SpaceXAI today released Grok 4.6, a large language model that it says can outperform Anthropic PBC’s Claude Fable 5 in some areas.

SpaceXAI was known as xAI until last month. The Elon Musk-founded artificial intelligence provider rebranded in connection with its acquisition by SpaceX Corp. In June, the combined company listed its shares on the Nasdaq via the biggest initial public offering on record.

Grok 4.6 is rolling out only a month after SpaceXAI released its previous flagship LLM. According to the company, one of the main improvements is that its engineers spent more time training the former model. The extended training run used an AI-generated dataset designed to improve Grok 4.6’s reasoning capabilities. SpaceXAI also provided the model with access to “high-quality engineering data.”

The initial training run was followed by two additional development steps. The first used a training method called supervised fine-tuning, or SFT, while the second used reinforcement learning.

An SFT training run refines an LLM’s output using a set of sample prompts and pre-packaged answers. Engineers mainly use the technique to ensure that LLM prompt responses are outputted in a user-friendly format. SpaceXAI used Grok 4.5, the predecessor of Grok 4.6, to optimize the latter model’s SFT training phase. The optimization workflow focused on improving its ability to tackle science and programming tasks.

SpaceXAI evaluated Grok 4.6 using the Artificial Analysis Intelligence Index. It’s a dataset that combines nine popular AI benchmarks spanning fields such as science, coding and financial services. Grok 4.6 scored 61, which put it on par with OpenAI Group PBC’s flagship GPT-5.6 Sol model and one point behind Claude Fable 5.

The company also compared the models across nine other benchmarks. Grok 4.5 managed to outperform Claude Fable 5 in three. One of the benchmarks, AA-Briefcase, evaluates LLMs’ ability to perform knowledge work projects that would take a human weeks to complete. The other two evaluations comprised tasks spanning more than a half-dozen industries.

Grok 4.5 is particularly adept at generating software prototypes based on high-level descriptions, according to SpaceXAI. Additionally, it’s better than its predecessor at creating visual assets such as interfaces. The model is also more likely to check its work for errors when working on long-horizon projects.

The standard version of Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens. SpaceXAI also offers a faster edition that costs twice as much. The LLM is available through Cursor, the vibe coding platform that the company bought for $60 billion in June, and an internally developed programming tool called Grok Build.

Photo: Unsplash

Denmark-based LuaBio ApS, operating as Again, today announced the acquisition of Genomatica Inc. to combine the company’s cutting-edge biomanufacturing capabilities with Geno’s artificial intelligence discovery and design platform for biomolecules.

Again co-locates modular bioreactors that use specially engineered microbes alongside industrial carbon dioxide emitters that feed waste gas directly into fermentation equipment rather than transport and sort the gas to distant facilities.

The company’s work involves using bacteria as living chemical factories. It uses a strain of CO2-fixing Moorella bacteria and AI-guided design to have them produce specialized enzymes that exchange CO2 and hydrogen for acetate, acetic acid and related derivatives.

These raw materials can be purified and refined into feedstock input for industries such as adhesives, solvents, plastics, textiles and cosmetics. As a result, any commercial-scale CO2 producer can be turned into a recycling point that lowers lifecycle emissions and can replace fossil-derived materials.

Geno produces and licenses biotechnology and simulation technologies for industrial process engineering for plant-derived chemicals and ingredients, replacing fossil fuel-based materials such as long-chain polycarbons. Its engineered microorganisms ferment plant sugars into materials used in consumer and industrial products.

Although the company uses a variety of microorganisms depending on the chemical it is trying to make, its primary workhorse is engineered Escherichia coli strains. This is because E. coli is one of the best-understood bacteria across bioengineering and the most deeply studied in academia, as it grows exceptionally well in massive, industrial-sized fermentation tanks, which is critical for making millions of pounds of eco-friendly chemicals.

Examples of product lines provided by Geno include renewable intermediates for spandex; multifunctional ingredients for cosmetics; bio-based alternatives for apparel, cosmetics, and carpets; nylon-based products; plant-sugar-derived products; and specialty chemical building blocks for fragrances.

Rather than owning every production plant, Geno also licenses its biomanufacturing processes to major manufacturers, including Novamont S.p.A. and Qore, a joint venture between Cargill Inc. and HELM AG.

The two companies did not disclose the terms of the transaction.

Again said that, at its core, the deal with Geno is data; the company is acquiring Geno’s entire AI computational biotech platform. This includes its entire patent portfolio, which includes data spanning experimental results, structural and scale-up insights and developmental outcomes.

More data will allow higher accuracy in discovery and enable faster progress, compressing the time required to produce new products when moving from idea to market.

Carbon dioxide is a powerful greenhouse gas. Together, Geno and Again will have an estimated combined capacity to divert up to 85 million tons of CO2 per year. As the two companies tackle carbon reduction on two different sides of biomanufacturing, their individual estimates reflect both direct carbon capture and indirect emission avoidance. Geno’s platform works for mass-scale industrial replacement and could prevent 85 million tons of carbon emissions annually; Again’s capture processes up to 1 ton of emissions per day. Again goes even further by preventing more pollutants from entering the atmosphere by bypassing the petrochemical routes required to make acetate, reducing the CO2 load further by up to 3 tons.

Now that the two companies are combined, they can accelerate these figures even further and pass the savings on to industrial outlets that license the technology for acetate and derivative stock production. Not only will this lower overall emissions and environmental stress — making regulators and people with lungs happier — but it will also recycle what was originally waste into more efficient revenue streams.

Image: Pixabay

River AI Inc., a startup that helps enterprises customize open-source artificial intelligence models, has raised $1.1 billion in early-stage funding.

The company stated in today’s announcement that it received the capital over two rounds, a seed and a Series A. General Catalyst and AMP PBC were the lead investors. They were joined by Nvidia Corp., AMD Ventures, Y Combinator and Temasek.

River AI is led by Chief Executive Officer Igor Babuschkin. He earlier co-founded xAI Corp. and worked at DeepMind as a researcher. Babuschkin helped develop the Alphabet Inc. unit’s AlphaCode system, the first coding AI that demonstrated competitive performance in a programming contest.

The startup’s inaugural product is a cloud service called the River API. It enables developers to tailor open-source large language models to their requirements by putting them through additional training. According to the company, the service supports LLMs with 35 billion to 1 trillion parameters.

River API customizes open-source models using a method called LoRA, or low-rank adaptation.It works by extending the model being customized with a small number of additional artificial neurons. Those extra neurons equip the LLM with capabilities that it doesn’t possess out of the box.

The primary selling point of LoRA is its cost efficiency. The standard way to extend an LLM’s capabilities is to retrain it from scratch, which can be highly resource-intensive. LoRA only requires software teams to train the small number of additional neurons they added to the model.

River AI says that the River API enables users to customize a model in 15 to 20 minutes. Additionally, it automates time-consuming prerequisites such as configuring the infrastructure on which training is carried out. The company says models customized using its service can be up to four times more cost-efficient than proprietary alternatives.

The River API is the first component of an expansive AI product suite River AI is currently developing. According to the company, the next addition will be a set of features designed to deliver “personalization and continual learning for agents.”

In a July blog post, Babuschkin wrote that the company’s long-term goal is to develop personal AI systems capable of adapting to user preferences. “It is yours, not rented, and you have real control over it,” he detailed. The executive also disclosed that the company’s engineering push is not focused solely on software.

A job posting indicates that River AI plans to develop a custom system-on-chip with an onboard machine learning accelerator. The company will produce the processor using “advanced foundry nodes.” River AI plans to offer a compiler that will automatically turn customer LLMs built using PyTorch, a popular AI framework, into a form that can run efficiently on its silicon. 

Image: Unsplash

Some of Wall Street’s biggest financial firms are partnering with Nvidia Corp. to pour a half-trillion dollars of funding into the artificial intelligence industry’s massive infrastructure buildout.

Nvidia said today it has struck deals with Apollo Global Management Inc., BlackRock Inc., Blackstone Inc., Brookfield Corp., Goldman Sachs Group and KKR & Co. Inc. For the first time, those investors are treating AI hardware and infrastructure as an asset class like stocks, bonds and commodities, the chipmaker added.

“In AI, compute is revenue,” said Nvidia Chief Executive Jensen Huang. “We are bringing the world’s leading long-term capital providers together to independently underwrite AI infrastructure.”

Nvidia said the funds from today’s deals will go toward both its own projects and those of its partners. Some of the projects backed by the fund will include the construction of new data centers needed to house, operate and maintain servers filled with hundreds of thousands of Nvidia’s graphics processing units, which are widely used to process AI workloads. The money will also be used to back new manufacturing facilities to produce those chips in order to meet growing customer demand.

Nvidia has become the single largest beneficiary of the AI boom. These days, basically every major AI firm and technology company uses its chips to power AI features, services and chatbots. It could even be argued that basically every large organization in the world has indirectly become a customer of Nvidia’s, for few companies these days don’t use some form of AI tools in their day-to-day business operations.

Some of Nvidia’s biggest direct customers include Google LLC, Microsoft Corp., Meta Platforms Inc., Amazon.com Inc., SpaceX Corp., OpenAI Group PBC and Anthropic PBC. Collectively, these companies have spent more than a trillion dollars on AI projects and infrastructure in the last three years, and they’re expected to invest even more in future. A huge chunk of that money has, and will continue to find its way into Nvidia’s bank accounts, which is why the chipmaker’s stock has increased fivefold over that three-year period.

By using institutional credit, insurance funds and private capital to underwrite new AI infrastructure projects, Nvidia is helping its customers secure the financing they need without drawing on their own balance sheets. “This is really the first time that technology chips have become an investable asset class,” Huang told CNBC in an interview. “These are revenue-generating assets now. They’re productive, long-lived, fungible and flexible.”

Traditionally, GPUs have always been seen as depreciating investments that quickly lose value upon delivery to customers. But Nvidia is challenging that assumption, arguing that compute capacity is a longer-term and bankable asset class that’s widely adopted and transferable across customers. It means lenders can reliably underwrite compute as a revenue-generating asset, though skeptics may question if GPUs really retain their value when newer generations of the chips emerge.

“What’s different about this industry and this way of doing computing is that the computer is now part of the infrastructure, like electricity, like the internet, and so you have to think about it like it’s infrastructure,” Huang argued.

KKR co-CEOs Joe Bae and Scott Nuttall have certainly bought into Huang’s argument. They said in a joint statement that AI is already so pervasive and important that compute has become a critical asset. “As we’ve scaled our approach to digital infrastructure, we’ve learned that delivery, not ambition, is the hard part,” they added.

However, not everyone agrees with Nvidia. Some investors have become wary of the chipmaker’s alleged “circular dealmaking” involving eyewatering amounts of money, and today’s announcement will likely stoke those fears.

Last month, for instance, Nvidia announced a $500 billion deal with the South Korean semiconductor giant SK hynix Inc. to secure a supply of memory chips. Shortly after announcing that deal, reports emerged claiming that the company was discussing a $250 billion deal with OpenAI to help finance the AI giant’s massive 10-gigawatt data center project in Ohio, which is expected to be one of the world’s largest “AI factories” once it’s completed in 2028. No agreement has been confirmed so far, but if it is, it would represent one of the company’s biggest deals with a customer, Bloomberg reported. The $250 billion would only cover the data center lease and debt, and talks are ongoing regarding a separate, $350 billion deal to finance chip purchases, the report added.

Critics say Nvidia is weaving a dangerous, tangled web of deals with a handful of companies that have overlapping interests, many of which have struck multibillion-dollar deals with each other. The concern is that these financial dependencies mean that if one deal falls apart, it could lead to a domino effect that engulfs not only the AI industry, but the entire global economy, given how much money is at stake.

Photo: Nvidia

Agentic artificial intelligence startup Echovane Inc. said today it has closed on a $1 million pre-seed funding round to accelerate the development of its AI-native market research platform.

Titan Capital and Neon Fund co-led the round, which will help the startup to build out its AI agent infrastructure and enhance its research capabilities.

Echovane was co-founded by Smriti Gupta (pictured, center), Vipul Nair (left) and Himadri Roy (right), who previously spent time on product development for companies that included Amazon.com Inc., Stripe Inc. and Razorpay Inc. Originally, they had a much more narrow goal of developing an AI agent that could interview humans to speed up qualitative market research projects.

However, by the time they had developed their first prototype, the co-founders realized that a simple automated interview moderator wouldn’t be enough to solve the bottlenecks in market research. As Chief Executive Gupta noted, the slow pace of interviews is just one reason why so many projects are delayed.

“We thought better AI interviews would unlock faster research,” she said. “But that was only one step in the research operation, for only one methodology of research. In reality, research is much more complex and can vary from interviews to unobtrusive observations and longitudinal studies. It has multi-level complexity, including finding niche participants, getting the research done on time with them and maintaining quality checks.”

According to Gupta, it was while she and her colleagues were building their original AI interview moderator that they discovered just how complex market research projects really are. They require recruiters, panel vendors, moderators, translators and data analysts, as well as someone to coordinate everything.

The complexities of this supply chain means that researchers often spend more time on logistics than actually interpreting what they discover and turning it into useful insights. What’s more, anytime a company has tried to speed up the research process and make it more efficient, the quality of the research often suffers because of poor screening, professional respondents and sometimes even fraud that compromises the quality of the study.

That’s why Echovane’s founders eventually decided that they need to take full ownership of the entire research process, from the initial brief to the decision-ready answers. Their solution is a team of AI agents designed to handle the market research process from end-to-end, starting with designing the studies, recruiting and verifying participants, executing multimodal fieldwork and synthesizing the qualitative and quantitative data those studies generate.

Once done, its AI agents don’t just conjure up a static report, but instead present their findings within an interactive dashboard and insight reels that make it easy for researchers to ask follow-up questions. Moreover, every insight can be traced back to its source, Gupta said.

Despite being bootstrapped until now, Echovane has already been adopted by global enterprises including Coca-Cola Co., Procter & Gamble Co., Haleon Plc, Kantar and Trustly AB. “Echovane turned a month of product research into just two days,” said Trustly Chief Product Officer Adam D’arcy. “They reached exactly the right participants and let us test, learn and iterate continuously across multiple audiences, countries and product variants, all within a single study that would be impossible to execute at speed and scale.”

It’s a promising start if nothing else, and Echovane now has ambitions to scale its global participant infrastructure to help more businesses reach specific, difficult-to-contact demographics anywhere in the world, without compromising on quality. “This funding allows us to deepen the agent infrastructure behind Echovane, expand our multimodal capabilities and strengthen the global participant network required to deliver increasingly complex studies with consistency,” said Chief Technology Officer Nair.

Photo: Echovane

Las Vegas was hotter than hell last week, but not as hot as the market for artificial intelligence-enabled security at Black Hat USA 2026. A bandwagon of million-dollar booths for overfunded agentic security startups arose like mirages from the desert. Fortunately, there was also real value to be found amidst the AI hallucinations.

AI supply chain challenges were the topics du jour for chief information security officers and researchers comparing each other’s level of preparedness for agentic AI disruptions such as the on-off introduction of Anthropic PBC Mythos-class escape exploits, and the recent OpenAI/Hugging Face attack, and even Meta Platforms Inc. saying it did one too, just to not get left out.

“I think this incident was a really good example of exploiting the nondeterministic capabilities we already know about AI, where there’s just a lot more code and traffic generated within a quicker time period, which can hide abuses,” said Patrick Duffy, head of product at Dropzone AI Inc., which announced its new AI Threat Hunter product at the show. “Human teams could not launch a thousand investigations at the same time to respond.”

One thing that has been thoroughly disproven since last year’s autonomous security operations center craze? AI automation is still not replacing professional security expertise. The ungovernable AI cat is already out of the bag, so we’ll need AI agents to work alongside us to stop it, because we’ll always have a shortage of skilled security professionals to follow up on the unknowns of an ever-expanding threat surface.

Here were some interesting new brain wrinkles I picked up at my first Black Hat:

Orchestrating the agentic SOC

Agents in the SOC, if governed, trained and employed correctly, can take a lot of work off the human security analyst’s plate, freeing up more time for critical investigations and strategic security architecture thinking, but there are many ways to get there.

Rather than keying off of alerts and incidents, Nebulock Inc. builds a behavioral world-model graph of an enterprise that overlays existing security information event management, information technology security management and alerting tools to conduct “hunt-first” agentic detections and investigations. Agent fleets can be triggered to respond autonomously to events such as patches or new CVEs, or a security analyst’s natural language request to follow up on a hypothesis informed by coverage gaps they are seeing in their security posture command center.

Huntress Labs Inc. provides a 24/7 managed platform used by hundreds of thousands of individual jacks-of-all-trades, service providers and mid-sized companies. Their in-context training, support experts and SOC workflow automation help customers that would have a difficult time covering every aspect of security.

“We try not to overuse AI buzzwords here,” said Aimee Simpson, director of product marketing at Huntress. “We do have Athena, an agentic investigator made of different subagents that are doing different tasks, compiling signals, investigations, writing incident reports, to take some of the workload from our SOC, but our customers care about results. If the agent is not confident in its findings, that still goes straight to our expert human analysts – and we’re definitely still hiring them.”

For a newer vendor on the market, Strike 48 brings a surprisingly broad agentic SecOps platform with multi-platform SOC, network operations center and DevOps connectivity. It has more than 300 MCPs under the hood, allowing its agents to resolve security issues alongside network and development engineering work. Larger companies can lean on their forward-deployed expert teams to build out specialized agents and further bespoke integrations.

Enabling agentic co-workers with less risk

Zero-trust policies with least privilege access controls, hardened containers, and microsegmentation have been around for years. These practices started gaining traction for securing containers and ephemeral workloads as cloud native architectures emerged. Now, the “IP-wandering” nature of nondeterministic AI agents makes setting boundaries – without becoming a blocker – more critical than ever.

“With microsegmentation, we’re narrowing down every asset in your ecosystem to put a protective bubble around it – workstations, servers, operational technology and cloud assets, and now we want to extend beyond infrastructure-as-code to cloud infrastructure itself, using native APIs built into the system of Azure, AWS or GCP,” said Chris Boehm, field chief technology officer at Zero Networks Inc., an identity and microsegmentation vendor that announced its enforcement of Open Worldwide Application Security Project Least Agency Principle for enterprise AI at the show.

“Agents can be your best worker, and your worst worker, and your adversary, all at the same time. You really need to take the perspective of agents to understand their intent, wherever they run,” said Jason Needham, CEO of Certiv Inc., an agent assurance firm that provides behavioral and intent monitoring to human agent owners, with judgments, policies and technical controls for agents, including local models.

Geordie AI Ltd. discovers agents wherever they exist across the enterprise, learns what they do and who is responsible for them, and helps security teams help their own organizations successfully adopt agents with less risk. Its Beam solution directly tests for exposure and operates agents, and acts as a control plane for managing work activities, with a light touch, so autonomy and adoption is not as inhibited as a typical security-oriented solution. 

“The trend for giving more autonomy to agents is only increasing this year,” said CTO Benji Weber. “You even have nontechnical sales and marketing teams adopting agents. Are these teams truly successful with agents, or are they just burning tokens?”

Enriching and optimizing SOC data estates for shared knowledge

AI SOC vendor Stairwell was there demonstrating its new Backstory threat intelligence data discovery and knowledge base that can pull multiple petabytes of contextual threat data in seconds into an all-hot live and historical data lake. It captures full-stack activity for multiple systems using parallel agents, without a time-consuming or costly search and ingest process.

“Rather than try to record every behavior, I want to find the thing that made the behavior, not just a bunch of logs,” said founder and CEO Mike Wiacek. “It changes the way teams think about security. If you were a bank that got robbed, would you want to replay the CCTV footage of the robber, or would you rather already have the guy in handcuffs?”

“We’ve seen this train coming for a while,” said Monzy Merza, founder and CEO and founder of Crogl Inc. The firm was at Black Hat promoting a free single-user download of its complete enterprise AI SOC platform to accompany its recent sovereign AI agent. “Sovereign data and private AI capabilities are becoming very important to customers who want to control their own data, and that’s why we have always been available on-prem as a customer-managed product.”

Endpoints are still the leakiest attack surface

On day three, I was going to record an interview, and my iPhone’s audio recording app said, “Cannot record audio during a phone call.” There was clearly no call going on, and the phone seemed kinda hot, so I rebooted it. Maybe some agent there at Black Hat was demonstrating a cool hack on my endpoint!

The most devastating attacks in play today are often insider threats on an employee’s desktop or phone, which are particularly difficult to recognize when the signature and intent of an agent’s actions on an endpoint are unknown.

Ent Security (Athena Formation Inc.) was there with a big presence just out of stealth, combining elements of user and entity behavior analytics, endpoint detection and response, application control, data and endpoint protection, user interface monitoring and remote access analytics, all of which together form an “intent-aware protection” layer for humans, AI and applications.

“You need to capture high-fidelity telemetry within the user’s workspace, everything from mouse clicks and opening files, screen shots, application focus changes and even remote workers and agents that might log in and take control of your desktop,” said Janani Nagarajan, VP of product marketing at Ent. “For instance, on day 1 of a customer install, they saw an employee invite a user in North Korea to a Zoom, and then hand over remote access to that user, and they were able to make an intervention.”

For mobile apps, which represent the majority of human user internet traffic, Appdome Inc. introduced new mobile test engineering capabilities and a remote management agent solution. It can make publisher-controlled live configuration and security updates to the application after build, packaging and deployment to an app store, in order to shrink the exploit window for live users if a new threat is discovered between major releases.

Improving DevSecOps collaboration

No, DevSecOps is not dead; it is just not as popular as it used to be. But security teams still need to collaborate with development and ops teams in order to realize the value the company expected from AI initiatives.

RevEng.AI (Binary AI Ltd.) built a machine code verification layer that explores production software, including binaries, for threats or risks. Rather than imitating static or dynamic application security testing scans, or building another large language model like OpenAI or Claude that understands English, it trained a large model that understands binary computer code at a root level.

“If AI is writing the majority of code today, In the future, why should AI write code that’s easy for human beings to understand? Why should it write Python, when it could go to compiled languages or even straight to binary at a 100-times or 1000-times quicker rate?” said CEO James-Patrick Evans. “But the problem is we’ve got to secure this code, and make sure that its output is actually correct.”

Software supply chain player RapidFort Inc. was demonstrating a new capability for monitoring an software bill of materials not just for approved software packages in builds, but within container deployments at runtime, in order to protect against drift in production environments. With agents writing infra code and introducing packages at an unprecedented rate, there are always new vectors to account for.

“We’ve seen developer credentials being compromised so what looks like a legitimate npm package gets pushed up to GitHub,” said Chief Marketing Officer Mike Wood. “People and agents think the update is OK to download, but it has malware. So we don’t make any of our images available to our customers unless they’ve been baked for two weeks and cooled off for two weeks, which guarantees that any issue will already have been discovered by the vendor community and thousands of developers and end users.”

I met with AISLE Inc., a deep-thinking startup focused on remediation in the AI-enabled software delivery lifecycle, that ended up exposing and publishing common vulnerabilities and extensions for several new zero-day vulnerabilties and topping research leaderboards, including 16 OpenSSL exploits and previously unknown vectors that Mythos-class models could generate. 

“These aren’t just rogue models, we’re talking about major AI companies that deliberately lowered the guardrails on their flagship models to see what they are capable of – when they weren’t supposed to be accessing the internet,” said Chief Operating Officer and CISO Jaya Baloo. “We want to verify them and prioritize remediation with a fix that doesn’t break anything new.”

When all else fails… kill the agent

Straiker Inc. was there with a futuristic cyberpunk presence for its multi-agent attack and defense solutions, and revealed a brand-new game with just one button, under glass: the Agentic Kill Switch. I couldn’t help but press the button, like I was the last man standing in a gas station with a leaking token pump.

‘We have tuned models that do detection, looking for remote code execution, indirect prompt injection, and fine-grained blocks, but sometimes, you actually need to just kill misbehaving agents with the push of a button,” said Parth Shah, Straiker’s head of product. “Ultimately, anyone that’s putting AI in their software needs to stay in the driver’s seat. They can’t have someone else in their supply chain that owns that kill switch.”

The Intellyx take

If there’s one thing that has been thoroughly disproved since last year’s agentic SOC craze, it’s that AI agents could replace human security professionals.

There will never be enough skilled SecOps team members with the insight to stay ahead of the rate of change AI is bringing to us, much less the chaotic behavior of fellow employees who trust AI with sensitive work.

Fortunately, at Black Hat it was impossible to ignore just how fast these vendors are moving forward with AI in their own products, and using AI to augment human awareness, in order to secure AI-driven applications that are appearing everywhere. The value AI and agents can offer human security teams is not a mirage.

The fact that new startups can build several enterprise-grade security solutions in less than a year is a testament to the incredible acceleration AI development can provide. But only working together as a community to improve the security awareness and expertise of builders and practitioners can save us from an uncertain future.

Jason English is a principal analyst and chief marketing officer at Intellyx. He wrote this article for SiliconANGLE. At the time of writing, Appdome, Dropzone AI and Straiker are current Intellyx customers, and Crogl and Zero Networks are former Intellyx customers. The event covered the analyst’s attendance cost, a standard industry practice. ©2026 Intellyx B.V.

Photo: Jason English

Artificial intelligence can be technologically transformative and still produce a capital bubble. Those two ideas are not in conflict.

The bubble bursting does not require AI to fail. It only requires deployable supply and capital commitments to grow faster than monetizable demand. When productive, revenue-producing AI capacity takes longer to materialize, pricing will normalize and financing will no longer bridge the gap. That’s when the capital cycle resets.

But here is the good news for investors: The AI supply chain remains constrained by high-bandwidth memory, advanced packaging, network fabric, power and site readiness. These bottlenecks not only slow deployment, they also delay price discovery (the point at which buyers have more choice); and they postpone the moment when the market discovers whether it has overbuilt.

Welcome to this week’s Breaking Analysis. We’ve titled it: “Forecasting the AI bubble: When scarcity turns to surplus.”

In this episode, we will build a framework for understanding what could delay the bubble popping, what could trigger it and which indicators will reveal when scarcity in supply becomes surplus that impacts the market.

We will also test that framework against the Oracle, OpenAI and Stargate buildout, which to us is the clearest current example of capital commitments racing ahead of deployment, utilization and cash flow.

Our premise today is the following:

The bubble will not burst because AI stops working. It pops if scarcity clears before utilization and cash flow catch up.

And the best place to begin is with the number that makes this cycle look almost unstoppable:

A projected $1.5T semiconductor market

Let’s begin with the sheer scale of the demand picture.

In 2024, David Floyer and I forecast that an expanded silicon ecosystem would approach $1 trillion by 2028. The scope of that model was broader than the WSTS semiconductor-product market shown here, so this is not a perfect apples-to-apples comparison. But directionally, the market is moving much faster than we anticipated.

Global semiconductor revenue approached $800 billion in 2025. The WSTS forecast shown here puts 2026 revenue at approximately $1.51 trillion – nearly double in one year.

So it’s clear that the underlying AI demand is substantial. Nvidia reported $75.2 billion of data-center revenue in its latest fiscal quarter, Broadcom reported $10.8 billion of AI semiconductor revenue, and AMD generated $5.8 billion in data-center revenue. Those are substantive proof points that AI-factory deployment and customer spending is unusually strong.

But the composition of the forecast is where it gets interesting.

More than half of the projected 2026 market comes from memory. So the revenue curve is being driven by two forces at once: enormous structural AI demand and extraordinary scarcity pricing.

That does not minimize the demand picture. It means the revenue line is rising faster than the physical unit and deployment curves.

And that is where the bubble analysis begins.

There’s little question that strong AI demand exists. The question is whether demand will continue absorbing capacity as HBM, packaging and other constraints ease – and as scarcity premiums begin to normalize.

That is why we need to understand two separate memory curves: physical bit growth, and the price and margin curve.

A decline in memory pricing alone would not mean the AI bubble has burst. The more concerning indicator would be prices falling while physical bit demand, deployment and productive monetization also begin to weaken.

To understand when that could happen, we have to understand why the supply constraints are not clearing simultaneously. The bottleneck moves across the system.

The AI bubble has a bottleneck clock

To forecast when scarcity could flip and become a surplus, we have to understand why the breaking point is likely to be delayed. AI infrastructure is not one market. It is a chain of interdependent gates as shown here.

A graphics processing unit allocation without sufficient high-bandwidth memory is not deployable capacity. HBM without advanced packaging does not become a working accelerator. Racks require network fabric to operate as a cluster. And a cluster still requires power, a ready site and capital before it becomes revenue-producing capacity.

The least available layer governs the output of the entire system.

But the key point is that the binding constraint moves. Early in the cycle, the dominant shortage was accelerator availability. It then shifted toward HBM and advanced packaging. As those constraints ease, the pressure moves outward toward networking, power, site readiness and financing. Solving one bottleneck often reveals the next one.

That does two things.

• First, it rations the amount of capacity that can actually reach the market.
• Second, it postpones the market-clearing test – the moment when we discover how much capacity customers will really consume and what they will pay, once supply is broadly available. That’s what we mean by “price discovery.”

David Floyer’s Volume-Value-Velocity framework gives us a useful absorption test. Additional volume is healthy when the technology creates enough value, and adoption moves quickly enough to absorb that capacity. But when system volume expands faster than useful demand and productive utilization, capacity turns into surplus.

So the forecasting question is not only: When will more GPUs or memory ship?

It is: When will the full system be capable of delivering more revenue-producing capacity than monetizable demand can absorb?

And the first potential release valves in that system are memory and advanced packaging.

Memory and packaging are the first release valves

The bottleneck clock gives us a reasonable place to look for the first signs that scarcity is turning into surplus, which leads us to memory and advanced packaging.

On the left, Micron’s fiscal third-quarter results show why the revenue reports can be misleading. Specifically, DRAM bit shipments increased by only a low-single-digit percentage sequentially. But the average selling price per bit increased in the low-60s percentage range.

Those are two very different signals. The first tells us how much additional physical memory Micron shipped. The second tells us how much more the market paid for each unit of that capacity. In other words, the surge in DRAM revenue was overwhelmingly driven by pricing and product mix – not by Micron shipping anything close to 60% more memory.

That does not mean AI demand is weak. It means scarcity is doing most of the work in the reported revenue line.

The bubble test is what happens as additional supply becomes available.

If memory pricing normalizes while physical bit demand and productive utilization continue to accelerate, that is potentially healthy. The market would be transitioning from a price-led cycle to a volume-led cycle.

But if average selling prices fall and bit growth also stalls, that is a more dangerous event. It suggests surplus is arriving faster than useful demand can absorb it.

Now let’s look at the right side.

Additional HBM supply does not automatically become deployable AI capacity. The memory has to be qualified, vertically stacked and integrated with the accelerator through increasingly complex advanced packaging – at acceptable yields.

A memory stack sitting in inventory is not productive capacity. It only becomes productive when it is packaged with the accelerator, installed in a working system and ultimately used for revenue-producing workloads.

So we need to watch four variables together:

• Bit shipments;
• Average selling prices;
• Package lead times; and
• Package yield.

When memory pricing and package lead times normalize together, the bottleneck will likely move downstream. But even then, assembled capacity does not earn a dollar until it reaches a powered site and gets put to productive use.

That brings us to the two clocks governing the AI buildout.

Capital can be committed years before capacity earns a dollar

Memory and packaging tell us how much accelerator capacity can be assembled.

This next slide asks the next question: When does that assembled capacity actually become economically productive?

AI factories operate on two very different clocks.

On the left is the short-cycle IT clock. That includes GPUs, custom accelerators and CPUs; HBM and server DRAM; networking and optics; and the servers and storage required to create a working computing system. These assets move comparatively quickly. They can be ordered, manufactured and delivered within quarters. And they are refreshed on a roughly three-to-six-year technology and economic cycle as new architectures improve performance, power efficiency and cost per useful output.

This is the clock most visible in semiconductor bookings, supplier revenue and near-term guidance.

But the right side operates on a much slower clock. The long-cycle site clock includes land and buildings, substations and grid interconnection, generation, cooling, water and the reusable physical infrastructure that may remain in service for decades. A new GPU can be designed and delivered in a few years with a cadence for follow-on products coming every 12 to 18 months. Meanwhile, a transmission line, substation and grid connection takes a decade to work through permitting and construction, with no follow-on sequence comparable to a GPU roadmap.

A data-center building can be physically complete while it is still waiting for transformers, switchgear, cooling infrastructure or usable power.

And that means the two clocks can move out of sync. Capital can be committed. Hardware can be allocated. Suppliers can report strong orders and revenue. But the capacity still has to be installed, accepted, energized and placed into productive use before it generates sustainable revenue and cash flow. By productive utilization, we mean more than whether the GPUs are turned on and the share of energized capacity running useful, revenue-producing workloads is throwing off sufficient volume, pricing and margin to recoup the capital invested.

That timing gap is where bubble risk accumulates in our analysis.

Many folks are hyper-focused on the risk that the industry is spending too much money. That’s not what concerns us. It’s more the risk that commitments, hardware deliveries and supplier revenue run ahead of the capacity that can be energized, productively utilized and converted into cash.

So the takeaway from the slide above is:

The short clock creates the appearance of rapid deployment. The long clock determines when that deployment actually becomes economic capacity.

And the clearest real-world example of this mismatch is the relationship among Oracle, OpenAI and Stargate.

Oracle, OpenAI and Stargate expose the commitment-to-deployment gap

Here we show a concrete example.

And it is important to understand that Oracle, OpenAI and Stargate are not three separate pools of demand that we should add together. They are one connected capital project, viewed from the buyer, supplier and physical-infrastructure perspectives.

OpenAI reportedly committed $300 billion over five years to purchase compute capacity from Oracle, with the contract beginning in 2027 as part of the broader Stargate buildout. That is an extraordinary demand commitment.

On Oracle’s side, that commitment is already showing up in capital spending, backlog and financing requirements. Oracle spent $55.7 billion in fiscal 2026 capex, up 162%, and has guided fiscal 2027 spending as high as $95 billion gross – or roughly $70 billion net as some customer monetization is showing up.

At the same time, Oracle’s remaining performance obligations reached $638 billion. But RPO is not deployed capacity, and it is not current revenue. Oracle expects only about 12% of that RPO to convert into revenue during the next 12 months. Much of the rest is scheduled across the following three to five years.

That is not, by itself, evidence that the contracts are weak. It is evidence that the spending and financing arrive substantially before the revenue.

The same gap appears from OpenAI’s side. OpenAI’s total forward compute commitments are estimated at roughly $600 billion to $665 billion through 2030 – many times its current annualized revenue – while the company is still reporting substantial cash burn.

So one side is committing to purchase enormous amounts of future capacity. The other side is spending and borrowing today to construct it. And physically, that capacity must still move through site construction, power, energization, customer acceptance and productive utilization before it produces any sustainable cash flow.

Oracle’s free cash flow illustrates the concern. It moved from approximately positive $26 billion in fiscal 2025 to negative $24 billion in fiscal 2026, while debt increased and Oracle’s credit rating moved to the edge of investment grade.

Again, this does not prove that Stargate fails. It shows where the risk accumulates.

Commitments prove intent. They do not yet prove deployment, productive utilization or return on capital.

The bubble risk lives in the time – and the financing – between those two conditions.

And this example tells us exactly what to watch: Does demand convert into accepted, highly utilized, cash-generating capacity before scarcity resolves and financing becomes less available?

That is the commitment-to-deployment gap we’re highlighting in this Breaking Analysis.

Two wildcards can change the bubble clock

Intel and China are relevant to this bubble framework for opposite reasons.

Intel can extend the capital cycle. China can accelerate the market-clearing test.

Let’s begin with Intel.

Our earlier and well-documented concerns with Intel were never that Intel would become irrelevant. It was that a leading-edge foundry without substantial external wafer volume could not achieve TSMC-like learning curves, utilization or merchant economics. And if former CEO Pat Gelsinger’s plan were allowed to continue without a restructuring of costs, the company would go bankrupt. And the probability of the Intel board allowing that was near zero.

With current CEO Lip-Bu Tan cutting more than 20,000 jobs, the U.S. government’s stake, Nvidia’s investment and those of others like SoftBank, that view now needs to be updated – but not reversed. Policy capital, strategic investors, product recovery and visible 18A learning have reduced Intel’s near-term failure risk and made the company a more important U.S. capacity option.

But the commercial proof point has not yet arrived.

The vast majority of Intel Foundry volume remains captive. The Nvidia relationship is strategically important, but it does not validate external wafer volume. The decisive milestone in our view remains a named external 14A anchor customer with committed wafer volume and credible ramp economics. Until that occurs, Intel is able to prolong the capacity buildout – and its associated capital intensity – before merchant economics are proven. But the risks remain until we see more proof.

China works on the other side of the clock.

Its progress is segmented, but mature-node manufacturing, NAND and commodity DRAM capacity are becoming meaningful. Open models and domestic platforms can also raise local utilization and accelerate ecosystem learning.

China still has major gaps in high-yield HBM/HBM4, EUV, advanced packaging and power efficiency.

But in our scenario, China does not need full frontier parity to affect the cycle. Enough memory and mature-node supply can pressure global pricing before those frontier gaps close. And rising domestic AI utilization can reduce the accessible market for Western suppliers.

So the bubble implication is asymmetric: Intel may delay the break by prolonging policy-backed capacity formation. China may pull the market-event forward by accelerating supply and price normalization.

Neither variable is simply bullish or bearish.

That is why the next slide resolves the outlook through scenarios rather than one smooth compound annual growth rate for the semiconductor value chain.

Three ways the AI capital cycle resolves

Intel and China do not map neatly into a single bullish or bearish outcome. That is why we believe the AI capital cycle should be expressed through three possible resolutions – not one single metric of market growth. Below we show our scenarios:

The first is a soft landing.

In this scenario, demand absorbs the capacity being built. Inference creates a second major volume curve (similar to what occurred with reasoning and agentic), productive monetization remains high and enterprise return on investment becomes increasingly measurable. In this scenario, memory prices normalize gradually rather than collapsing. Additional compute is consumed as it becomes available, and the primary constraint moves outward toward power, sites and energization.

The second scenario – which is our base case – is a delayed reckoning.

HBM, advanced packaging, network fabric and power remain constrained. The bottleneck continues to rotate rather than disappear. Scarcity premiums persist, capital commitments remain ahead of physical deployment and the market-clearing event is deferred.

This scenario will likely sustain the boom for longer (through 2027 or possibly longer). But it can also allow the gap between committed capital and cash-producing capacity to grow.

The third scenario is the bubble break.

Here, the scarcity event clears before cash flow catches up. In this scenario, memory and packaging supply expand. Lead times normalize. Memory ASPs and GPU rental prices fall. Backlog conversion slows, productive monetization weakens and customers become less willing to prepay for future capacity. At the same time, debt, equity and third-party financing become less willing to bridge the gap.

Importantly, AI does not have to fail for this scenario to occur. The bubble pops when deployable capacity begins growing faster than profitable utilization – and capital stops financing the difference.

The transition among these scenarios is observable. So the final question is not which outcome sounds most compelling.

It is: What indicators tell us that capacity is being absorbed – and what indicators warn that scarcity is turning into surplus?

Watch conversion, not announcements

Let’s close with the early-warning dashboard shown here.

The bubble question is not how many GPUs are announced, how many campuses are planned or how much backlog is reported.The question is whether those commitments move all the way through deployment, energization, productive utilization and cash flow.

The first sign is HBM bit growth versus average selling price.

The healthy outcome is that physical bit demand accelerates while pricing normalizes gradually. That would suggest the market is moving from scarcity-driven revenue to sustainable, volume-led growth.

The warning signal is that memory prices fall while physical bit growth also stalls. In that case, the industry loses both its scarcity premium and the underlying volume engine.

Second, watch advanced-package capacity, lead times and yield.

Additional supply is constructive when yields improve and the new capacity remains highly utilized – again meaning profitably monetized.

The warning is that lead times collapse and package capacity arrives into weaker demand. That would indicate the first major physical bottleneck is clearing faster than the market can absorb it.

Third, watch GPU rental pricing and productive cluster utilization.

Pricing can moderate without being bearish if utilization remains high. Cheaper compute paired with rising consumption could expand the market as the Jevons paradox kicks in.

The more dangerous combination is rental prices and utilization falling together. That would be the clearest signal that deployable compute supply is beginning to exceed profitable demand.

Fourth, compare energized megawatts with announced capacity.

A site or rack waiting for power is a deployment problem. A powered site that remains underused is a demand and monetization problem. The healthy outcome is that energized capacity converts rapidly into revenue-bearing workloads.

Finally, watch backlog, customer prepayments and free cash flow.

Contracts have to become accepted capacity, recognized revenue and operating cash flow. The warning signs are slower acceptance, weakening prepayments and rising dependence on debt or equity to sustain the buildout. There are signs that the debt and equity warning sign is flashing today, but capital appears to be unconstrained, for now.

Taken together, these five indicators separate our three scenarios.

• If new capacity is absorbed, the soft-landing case strengthens;
• If the signals remain mixed, rotating bottlenecks continue to delay the day of reckoning;
• But if pricing, productive utilization and financing weaken together, scarcity is turning into surplus.

So the action item is straightforward: Track the entire conversion chain – commitment, deployment, energization, productive utilization and cash flow.

The bubble breaks when scarcity resolves before productive utilization and cash flow catch up.

So, when does the bubble pop?

Here’s our best assessment of the timing and risk levels of supply resolving ahead of productive monetization. We would peg 2029 as the high-risk year as it’s likely that enough supply will be deployable to expose a pricing equilibrium – the bottlenecks will largely be resolved, China capacity will be significantly higher and Intel will be a viable second source to TSMC. Power delays could absolutely push this into the 2030s.

Our highest-risk window is 2028 through 2029, with 2029 the most likely year for a broad capital-cycle break. We expect the first cracks to appear earlier – but the bubble itself becomes visible only when enough memory, packaging and energized capacity reach the market to test whether productive demand can absorb them.

If we were forced to put a timeframe on when scarcity might resolve ahead of productive utilization, it’s 2029. But that doesn’t mean we won’t have a soft landing. Let’s hope if and when the bubble bursts, and it very likely will if history is an indicator, enterprises have broadly figured out the monetization equation and the pop will be less dramatic than other black-swan events.

Perhaps that’s wishful thinking….

Image: theCUBE Research

Disclaimer: All statements made regarding companies or securities are strictly beliefs, points of view and opinions held by SiliconANGLE Media, Enterprise Technology Research, other guests on theCUBE and guest writers. Such statements are not recommendations by these individuals to buy, sell or hold any security. The content presented does not constitute investment advice and should not be used as the basis for any investment decision. You and only you are responsible for your investment decisions.

OpenAI Group PBC today disclosed that one of its unreleased large language models may pose a significant cybersecurity risk.

The algorithm, which is known as Astra, was first detailed last week. OpenAI revealed in a Sunday blog post that the LLM had solved 10 long-running math problems. The company published the proofs and revealed that each one took about $2,000 worth of tokens to generate.

As part of its artificial intelligence safety efforts, OpenAI has published a 29-page document known as the Preparedness Framework⁠. One of the document’s sections contains a rating system for AI risks. The system ranks the cybersecurity risks posed by an LLM as “High” or “Critical” depending on its capabilities.

OpenAI’s flagship GPT-5.6 Sol model and a few earlier algorithms were given a High rating. According to the company, Astra is the first of its LLMs that may qualify for a Critical designation. Its engineers drew that conclusion based on a series of recent cybersecurity tests.

“These results, in addition to expert assessments, have led us to conclude last night that we cannot rule out critical cyber capabilities,” OpenAI stated in a blog post.

Under the Preparedness Framework⁠, a model poses a Critical cybersecurity risk if it can find zero-day exploits in “many hardened real-world critical systems.” The model qualifies if those exploits span multiple severity levels and are discovered without any human assistance.

OpenAI also designates an LLM as Critical if it can launch cyberattacks against hardened systems based on only a high-level hacking goal provided by a user. The company didn’t specify which of its two designation criteria were met by Astra. However, it did share details about how it’s tackling the risk.

The company is taking steps to ensure that Astra can’t access the public web. According to the company, its engineers will run the model in test environments that have restricted network and tool use permissions. OpenAI is pausing development activities that aren’t carried out in such sandboxes.

It’s also stepping up its efforts to prevent hackers from stealing Astra’s code. According to today’s blog post, the initiative will place particular emphasis on the encryption that protects the LLM’s weights. Those are the configuration settings that determine how a model processes data.

Astra powers a number of internal AI agents. The company has implemented observability mechanisms that monitor those agents for malicious activity. The mechanism spot suspicious behavior by analyzing agents’ chain of thought, a step-by-step summary of inference activities.

OpenAI will share some of the cybersecurity workflows it has developed with “third-party testing partners.” Those partners help the ChatGPT developer run the sandboxes in which it evaluates its LLMs’ capabilities. Additionally, OpenAI plans to loop in relevant government agencies and AI safety organizations. 

Image: OpenAI

In the rush to adopt AI, enterprise security architectures are diminishing, and the only appropriate response is a network security platform that incorporates networking and security into a single, consistently governed fabric.

That’s the vision Anthony James (pictured), executive vice president of marketing at Fortinet, pitched at Black Hat 2026. James, who first joined the company in 2004, said the vision that Ken Xie established more than two decades ago, consolidating fragmented security tools into a unified network security platform, is more relevant now than ever.

“If you look 20 years later after we started this whole idea of unifying threat management on a firewall, we’re now seeing that convergence of networking and security,” James said. “Platforms are becoming much more evolved, and CISOs are trying to cope with how do I make sure my infrastructure gives productivity to my end users, but also eliminates as much risk as possible.”

James spoke with Krista Case at Black Hat USA during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed how Fortinet Security Fabric differentiates from platformization, what the SASE Firewall vision means in practice and where enterprises are asking for the most help as AI moves from experimentation into production. (* Disclosure below.)

Network security platform strategy converges SASE, NGFW and AI governance into one fabric

Fortinet’s Security Fabric is built on a foundational principle that distinguishes it from what James termed bolted-on platform approaches. Every product Fortinet builds must integrate natively, share consistent data and simplify the operator’s experience from the beginning. The result is a single source of truth throughout the security stack instead of a collection of consoles with different data models that practitioners have to manually reconcile under pressure.

“Anything we build to help a CISO or a practitioner solve a security challenge, we build it with those foundational elements in mind,” James said. “Data is consistent, the experience is consistent, and it simplifies the user’s experience at the end of the day.”

The next step in that strategy is what Fortinet calls the SASE Firewall, a convergence of secure access service edge and next-generation firewall built on the same policy for both on-premises and cloud deployments. James said that cloud-first SASE vendors over-rotated, and the growth of sovereign AI and agentic workloads has confirmed what organizations still need in the network. As AI leaves experimentation behind and begins production, Fortinet is also building governance capabilities around internally deployed AI factories, controlling which agents and MCP servers are sanctioned, monitoring token usage and making sure machine-to-machine communications are legitimate.

“Now what we’re seeing is customers want to know what people are doing with their AI infrastructure internally,” James said. “Are they standing up elements of AI that’s not sanctioned, like agents and MCP servers and downloading code to their desktop? That’s a big no-no, and that could be sending data outside to train other models. So we’ve developed a lot of tools, and we have products that are in market and some that are coming that we’ll be excited to announce later this year.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of Black Hat USA:

(* Disclosure: Fortinet sponsored this segment of theCUBE. Neither Fortinet nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE

Palo Alto-based Naïve Inc., an artificial intelligence lab that’s developing autonomous agents capable of setting up and running entire businesses, said today it has closed on a $28.5 million Series A round of funding. Nexus Venture Partners led the round, which saw participation from Y Combinator, Zetta, Liquid 2 and a host of angel investors.

The startup is all about automation, noting that AI tools that can lighten people’s workloads have proven to be especially popular with people such as developers, enabling them to avoid repetitive and tedious tasks. Now it wants to advance this and use AI automation to perform basically all of the work that’s required to get an established business up and running. It’s an enticing concept, and so far Naïve has already attracted more than 30,000 developer customers to its platform.

Naïve goes further than traditional vibe coding. It has developed the infrastructure for AI agents to automate all of the steps required to establish a company and then run it on a day-to-day basis, so long as the customer provides the actual agents and the token budget. Essentially, it packages together the process of creating payments infrastructure, email accounts, phone numbers, cloud infrastructure, storage and business incorporation, making it all available through a simple application programming interface.

The company supplies a set of prompts that developers can use with third-party agentic tools such as Claude Code, Codex and Cursor. When prompted, these agents will connect to Naïve’s API to provision all of the infrastructure and other things needed to establish a company. For instance, it can tell agents how to orchestrate the creation of a U.S.-based limited liability company, providing details such as the state it’s based in, a business description, industry code and a proposed name. However, the owner of the business must still perform the Know Your Customer and Know Your Business processes themselves and make any payments that are required.

But just about everything else can be automated by AI agents. This includes tasks such as setting up email inboxes, phone numbers, databases, compute resources and links to payment processors like Stripe and accounting platforms like QuickBooks. Naïve has also built a comprehensive governance layer that can be used to restrict the capabilities of user’s AI agents, establish budgets and decide which actions require human approval before they’re taken by the agents. In addition, there are templates for things such as AI SEO, full-stack software-as-a-service applications, customer support, recruiting and accounting. It even has a mobile emulator that can be used by AI agents to navigate smartphone apps on virtual devices.

Naïve says it has enjoyed impressive revenue growth, with sales increasing more than tenfold over the last six months, achieving an annual revenue run-rate in the low double-digit millions. Customers have used its agentic infrastructure to create and run a range of autonomous businesses, including content channels on apps such as TikTok and YouTube, car rental firms that are fully managed by AI agents and more.

However, Naïve believes that while this toolkit is extremely useful to entrepreneurs, it has an even bigger opportunity in terms of reducing the costs associated with AI automation. That’s important, because while an almost fully-automated business might sound great to the person that owns it, the cost of running those agents can escalate dramatically as they make hundreds of calls to expensive models, pass around large volumes of context and eat up resources while sitting idle.

That’s why Naïve is also looking for ways to improve the efficiency of AI agents. Its innovations include a model router that instructs agents to always choose the most cost-effective AI model for each task and a memory system that allows them to store and reuse business context. It has also built a serverless runtime environment that allows agents to run in lightweight JavaScript environments, further reducing costs. In this way, it says it can help to slash the operational expenses of automated businesses.

Co-founder and Chief Executive Sean Dorje said this kind of optimization is vital because agent spending is likely to grow to trillions of dollars in future as more companies embrace automation. “Our vision at Naïve is to make each token do more, so autonomous companies can become a cost-efficient reality,” he said.

Ultimately, this could turn out to be a much more valuable business for Naïve than simply helping company founders automate the grunt work of setting up a new company. The platform will certainly be useful in getting a business up and running, but once that’s done, most entrepreneurs will hope to expand and grow that business for years to come. If Naïve can significantly reduce the cost of running the hordes of agents that automate their businesses, those customers will likely want to keep using its platform for as long as their companies are around.

Abishek Sharma of Nexus Venture Partners said autonomous software has already become an established fact, and the next step is to have fully autonomous companies. “Naïve gives millions of entrepreneurs and small businesses worldwide the turnkey infrastructure needed to build and run autonomous companies without needing to become AI experts,” he said.

Image: Naïve

Google LLC is reportedly in discussions with an artificial intelligence coding startup called Mechanize Inc. over a deal that would see it license some of its technology and hire a number of its most talented employees. Business Insider was the first to report the discussions, citing information from four people familiar with the conversations.

Mechanize was founded last year, saying that its overriding goal is to eventually create AI systems that can automate the work done by every single knowledge worker in the world. It’s currently focused on developing simulated virtual environments, evaluation benchmarks and specialized grading systems that can be used to test and train AI agents on various kinds of business tasks.

Earlier this year, the startup announced it had raised $9.1 million in funding from investors including former GitHub Chief Executive Nat Friedman, Stripe CEO Patrick Collison and the podcaster Dwarkesh Patel. Its CEO Tame Bisoglu is notable for having previously co-founded the AI research outfit Epoch Artificial Intelligence Inc. Business Insider said Google is currently discussing a sum of around $1.5 billion to strike a non-exclusive licensing agreement for Mechanize’s technology. As part of that deal, it would also hire some of Mechanize’s experts in model evaluation and development, the sources said.

It’s believed that Google is interested in using Mechanize’s technology to enhance the performance of its AI models, particularly when it comes to coding. That’s because Google’s coding agents are generally considered to be inferior to those of Anthropic PBC and OpenAI Group PBC, which have seen strong enterprise adoption with Claude Code and Codex, respectively.

However, Mechanize’s ambitions extend beyond coding. “Our current focus is software engineering, but our long-term goal is the full automation of valuable work across the economy,” it says on its rather threadbare website.

The discussions between Google and Mechanize highlight two key trends in the AI industry today. The first is that coding has emerged as one of the most lucrative applications of AI models, while the second is that big technology firms are willing to pay serious sums of money to secure the most promising AI developer talent.

Google has done this kind of thing before. Rather than buying AI startups outright, it structures the deals as hybrid transactions that enable it to secure their technology through licenses and poach some of their top talents. It’s a useful strategy as it gets around the antitrust scrutiny that comes with full acquisitions.

One of the most notable examples was Google’s acquihire of Windsurf’s most experienced AI engineers, and it also licensed the company’s tech. Former Windsurf CEO Varun Mohan now leads the development of Google’s agentic coding platform Antigravity. Two years ago, Google struck a similar deal with Character Technologies Inc. to re-hire its co-founder Noam Shazeer and obtain non-exclusive rights to its technology. Shazeer recently left Google again, however, joining OpenAI.

Image: SiliconANGLE/Dreamina

Cloudflare Inc. today announced the launch of Cloudflare OS, an open-source artificial intelligence agentic workspace available through the browser, filled with custom shared micro-applications for enterprise employees.

Most AI tools offered to the enterprise are generalized and understand a great deal about the world, but very little about the specific company they’re deployed for. Cloudflare OS was built from the ground up for its own employees, its own workforce, and its own teams, who use it daily to perform research, create documents and build apps for their day-to-day jobs.

“Cloudflare OS is how we run Cloudflare. For AI to truly transform an enterprise, it can’t live in a silo or behind a developer bottleneck. Every employee needs the ability to build, iterate, and automate safely,” said co-founder and Chief Executive Matthew Prince.

The platform is designed to capture a company’s working knowledge from day one and travel with employees by allowing them to create custom micro-apps that help them build their own workflows with their internal knowledge. They can then share that information and collaborate internally without worrying that the context will be lost.

Because the framework is open source and it runs on a company’s own Cloudflare account, organizations own anything they build on it. The company maintains and possesses the processes, context and internal system connections and will not get locked into a closed product.

According to Cloudflare, what they get is this: an AI agent workspace for everyone that is accessible via browser that doesn’t require developer expertise that allows employees to conduct research, produce documents tied to live data, run automated workflows, build tools and do busywork, all without waiting for the information technology or development teams.

This means they can quickly spin up smart applications they can build and share with everyone with their own isolated database, real-time capabilities, access controls and more, no developer required. As it is built on Cloudflare Access, it is zero trust by default, which verifies every user and every request before access is granted.

Another useful part of what Cloudflare built with Cloudflare OS is bring your own model. Through Cloudflare AI Gateway, organizations can use any AI model provider. They are not locked to any one vendor; this means administrators can control how tokens are used and see what’s being spent, broken down by person, team, and app. This allows them to set spending budgets, rate limits, or route tasks according to complexity. For example: routine tasks can be sent to small, efficient models and to frontier models when high reasoning is needed.

“We built this because nothing else did what we needed. Now any company can start from where it took us years to get,” added Prince.

Image: SiliconANGLE/Microsoft Designer

Shares of Arista Networks Inc. soared in extended trading today after the networking infrastructure firm easily beat Wall Street’s expectations in its second-quarter earnings report. The company’s stellar results provide an additional proof-point to claims that the demand for artificial intelligence infrastructure services is continuing to gain momentum. Arista, which is one of the main networking hardware suppliers for AI data centers, has emerged as one of the beneficiaries of that trend.

Arista reported earnings before certain costs such as stock compensation of $1.02 per share, easily beating the analyst forecast of 89 cents per share. Revenue for the quarter jumped 38% to $3.04 billion, ahead of the Street’s consensus forecast of $2.83 billion. Its stock gained more than 11% in late trading, adding to a gain of just over 3% during the regular session. Its stock is now up 45% in the year to date, and its market capitalization has now reached $233 billion.

Chairperson and Chief Executive Jayshree Ullal (pictured) said it’s clear that Arista’s 2.0 platform strategy is proving to be extremely compelling with the customers it serves. “Customers see networking as the central nervous system for infrastructure from the client to campus to data and AI centers,” she said in a statement.

Arista, which makes high-speed computer networking equipment for data center customers such as Microsoft Corp. and Amazon Web Services Inc., was also boosted by its strong guidance for the current quarter. The company said it’s shooting for earnings of between $1.06 and $1.08 per share on revenue of around $3.3 billion in the third quarter. That compares to the Street’s call for earnings of just 92 cents per share on sales of $2.95 billion.

The company posted similar results in its fiscal first-quarter earnings report, but its stock fell because executives warned that they’re struggling to secure supplies of components such as memory, silicon chips and optical cables. Chief Operating Officer Todd Nightingale provided an update on the situation in a conference call with analysts, saying that the company has been working hard to solve its supply problems for the last six months. Those efforts are beginning to pay off, he insisted.

“Our capacity has been increased in both manufacturing and distribution, and we’ve negotiated better component delivery terms,” he added.

Surprisingly, the company has also managed to blunt the impact of much higher component costs, which is an issue affecting virtually every hardware supplier today. Arista posted an adjusted operating margin of 49.9%, up from 47.8% in the first quarter and 48.8% in the same period one year ago.

Photo: SiliconANGLE

Nuclear energy startup Valar Atomics Inc. said today it has raised a bumper $1 billion Series B funding round to help it move from building prototype small nuclear reactors to manufacturing them at scale on a production line.

Today’s round was led by Sequoia Capital and saw participation from nine other venture capital firms, including Valor Equity Partners and Conviction. In addition, Valar said it has also secured a separate $200 million credit facility from Erebor, J.P. Morgan, Crescent Cove and Hercules Capital.

Valar Atomics founder and Chief Executive Isaiah Taylor said the investment will help the company to change what it does, rather than what it designs. It can now shift from “demonstrating the operability of an integrated reactor system to producing fleets of them en masse,” he told Bloomberg in an interview. The startup intends to build hundreds of small modular reactors, known as SMRs, which are basically miniaturized power plants, on production lines, similar to how cars are manufactured. Building them this way will be much faster and cheaper than the traditional way power plants are constructed, it says.

In June, Valar Atomics partnered with Nvidia Corp. to demonstrate the effectiveness of its small Ward-250 test reactor, which was installed in a data center and used to power Blackwell graphics processing unit running artificial intelligence workloads. “This marked the first and only time that an advanced reactor has directly powered AI infrastructure, and was the first time in history a startup generated nuclear power,” Taylor wrote in a blog post announcing today’s round.

Following that demonstration, Valar announced an expanded partnership with Nvidia to build a complete 30-megawatt nuclear-powered AI facility in Utah. The reactor’s design integrates a closed-loop cooling system that could reduce the facility’s water usage from around 2.6 million gallons per megawatt per year to just a few hundred gallons. Taylor told Bloomberg that the startup’s SMR development lifecycle is speeding up exponentially, as it implements the lessons learned in building them to improve its processes. “It took two years to complete the NOVA core, then it took seven months to take Ward 250 critical,” he said of the company’s first two SMRs.

“With each reactor built, the tick rate will become smaller until Valar is producing tens, hundreds and then thousands of reactors per year,” he added.

The idea is that Valar will eventually establish a giant factor that’s able to churn out SMRs that can then be delivered directly to data centers, industrial sites and other facilities that need a reliable energy source. The company believes its modular approach can dramatically reduce the costs in a nuclear energy industry that’s renowned for budget overruns and project delays. It’s an idea that’s also being pursued in China and in Russia, where state-backed corporations have already gotten the small power plants up and running. But to date, there are none in service in the U.S.

Taylor says the rise of AI and its massive power demands have added a sense of urgency to the company’s mission, which is why today’s round is such a critical milestone for his company. “AI is building very quickly, and we need a lot of power in every direction,” he insisted. “This is really about scale for us, this is the firepower that we need to go and build.”

It’s notable that Valar isn’t alone in trying to build fleets of SMRs that can power the AI factories of the future, In fact, it’s just one of several U.S. startups pursuing more cost-effective and scalable ways of starting self-sustaining nuclear chain reactions that can generate the continuous energy needed for AI. Rivals including X-energy Reactor Co., which raised $700 million in November before going public in April; Radiant Industries Inc., which raised $300 million in December; Aalo Atomics Inc., which pulled in $100 million in August 2025; and Bluecore Energy Inc., which last month closed on a $10 million seed funding round to pursue its idea of shrinking nuclear reactors and putting them on barges.

Holger Mueller of Constellation Research told SiliconANGLE that there is certainly a need for more affordable and abundant energy sources, because there are so many new data centers springing up to support AI workloads all over the U.S. However, he said while startups like Valar Atomic and the others have renewed interest in the potential of nuclear power, it’s not yet clear if they can provide the solution the industry needs.

“There’s no question that the demand for energy is there, and many people are interested in nuclear power because it’s affordable and can be made available in any location,” he said. “All you need to do is build the nuclear facility, and that’s what Valar is doing with it’s plan to mass produce small nuclear reactors. But there are still many unresolved long-term questions around nuclear energy in general, and it has yet to prove its small reactors can really scale. We’ll see how it fares soon enough.”

Image: Valar Atomics

Autonomous security solutions startup Horizon3.ai Inc. today announced it has raised $250 million in a Series E round on a valuation of more than $2 billion to build out its sales operation and expand into Australia and Singapore. Horizon3.ai closed its Series D in June 2025 at a valuation of $650 million.

Horizon3.ai’s NodeZero platform attacks its customers’ own networks. The tests run against live production systems, and Horizon3.ai says nothing breaks. A single run chains weak credentials, misconfigurations and identity gaps into a working attack path, the same sequence a real intruder would use. Fix guidance comes with the findings. NodeZero re-runs the test afterward to confirm the path is closed.

NodeZero can also seed an environment with honeypots while it tests. Horizon3.ai sells the decoys as a cheap way to catch an AI-driven attacker that is already inside a network.

Horizon3.ai says NodeZero is now used by more than 6,500 organizations, up from about 3,000 at the Series D. Customers include the U.S. National Security Agency, the Cybersecurity and Infrastructure Security Agency, four Fortune 10 companies, multinational banks and healthcare networks. Annual recurring revenue grew 120% year-over-year, and the platform has now run upwards of 300,000 pentests in production, a volume of attack data the company treats as its hardest asset for a rival to copy.

“We invented the concept of AI Hackers and spent six years earning the right to autonomously pentest the most critical and sensitive networks in the world,” said Snehal Antani, co-founder and chief executive of Horizon3.ai. He added that the round gives the company fuel to scale as the leader of what he calls the “AI vs. AI” era.

Antani started the company in 2019 with Anthony Pillitiere. Antani had been chief technology officer at Splunk Inc. Pillitiere came out of U.S. Special Operations Command, where he served as deputy chief technology officer.

Most of the new capital goes to hiring in sales, marketing and channel, aimed at enterprise, mid-market and federal buyers. Singapore and Australia are the first stops in an international push that also takes in more of Europe, the Middle East and Africa. The rest funds autonomous blue-team agents, software meant to remediate what NodeZero finds without waiting on a human.

Existing investors NightDragon and New Enterprise Associates Inc. co-led the round. New backers Acrew Capital Management, Blue Cloud Ventures, EDBI Pte. Ltd., Demeter Group, PSG Equity, SAIC Ventures and Sapphire joined, along with returning investors Craft Ventures Management, Qualcomm Ventures, Ridge Ventures and SignalFire. NightDragon founder and Chief Executive Dave DeWalt, who previously ran FireEye and McAfee, takes a board seat alongside NightDragon Managing Director Morgan Kyauk. DeWalt said Horizon3.ai’s platform is “fundamentally reshaping how the world defends its data.”

The Series E brings Horizon3.ai’s total funding to $428.5 million.

Image: Horizon3.ai

OpenAI Group PBC revealed Saturday that an internal version of Astra, the model family it calls its next major release, produced new results for 10 problems in mathematics and theoretical computer science that had been open for at least a decade, and it published machine-checkable proofs alongside the claim.

The company posted a 249-page manuscript collection, model-written reasoning walkthroughs and Lean 4 certificates for all 10 results. The certificates sit on GitHub under an Apache 2.0 license, and the repository reports a “sorry” count of zero, meaning no step in any of the formalized proofs has been left unproven.

The headline result is an explicit construction of a non-sofic group, a question left open since Mikhail Gromov introduced soficity in 1999. Astra also disproved Connes’s rigidity conjecture, constructing infinitely many non-isomorphic groups with property (T) that share the same von Neumann algebra, and it proved Ehrhart’s volume conjecture. Three problems from Paul Erdős’s catalog fell as well, including problem 183 on multicolor Ramsey numbers.

Stripped of the terminology, a group is the mathematical description of a set of symmetries, and a sofic group is one whose structure can be approximated by shuffling a finite deck of cards. Every group anyone had examined turned out to be sofic, and no one could prove that all of them are. Astra built the exception. Connes’s conjecture, posed in 1980, held that for one rigid class of groups, a related algebraic object acts as a unique fingerprint, pinning down the group it came from. Astra produced infinitely many distinct groups sharing a single fingerprint.

Erdős problem 183 is about Ramsey numbers. Color the links in a network with a fixed number of colors and past a certain size you cannot avoid a triangle whose three links match. The Ramsey number is the size at which that becomes true.

The remainder of the list runs across high-dimensional sphere packing, binary and spherical codes, arithmetic circuit complexity, quantum parallel repetition and the hardness of the closest vector problem, the last of which bears on lattice cryptography. Astra also produced counterexamples in extremal graph theory, resolving two more Erdős problems.

The Lean certificates are what give the announcement its weight. Lean’s kernel returns a binary verdict, either the proof compiles or it does not, which takes trust in the model out of the equation. What it does not take out is the need for a mathematician to confirm that each formal statement says what the open problem actually asks and to judge whether the result matters. None of the 10 has been through peer review.

OpenAI has been here before. The company’s then vice president of science, Kevin Weil, claimed in October 2025 that GPT-5 had solved 10 previously unsolved Erdős problems. Thomas Bloom, who maintains the erdosproblems.com database, called that “a dramatic misrepresentation.” The model had found papers in the literature that Bloom was personally unaware of. Weil deleted the post, and Google DeepMind Chief Executive Demis Hassabis called the episode embarrassing.

Bloom called the Astra results “big news” and rated them ahead of the Erdős unit distance counterexample an internal OpenAI model produced in May, a paper he helped verify.

Astra itself remains unreleased. OpenAI describes it as a model family built to run long tasks by coordinating multiple agents over extended periods, an extension of the test-time reasoning work associated with research scientist Noam Brown, who called the results “a major step for scientific reasoning” in a post on X. Human researchers turned the model’s output into publishable papers, though OpenAI said the mathematical arguments themselves came from Astra.

The compute bill was modest. OpenAI put the token cost for all 10 solutions at roughly $2,000 at GPT-5.6 Sol application programming interface rates.

Chief Executive Sam Altman demonstrated Astra to policymakers in Washington in recent days. The company has not given a release date, pricing or a decision on whether the model ships as GPT-6 or as another GPT-5 variant, and any launch will run through the federal AI safety review process that already staggered the GPT-5.6 rollout.

The timing is awkward for a mathematics community that has been pushing back. In June, the International Mathematical Union endorsed the Leiden Declaration, which warns that AI companies are “using published research without consent, bypassing peer review, and threatening the integrity of proof and attribution.”

Fernando Borretti, a software engineer who writes on technology, argued in a blog post responding to the release that the usual defenses of human mathematicians no longer hold and that the frontier of the field will recede past the point where anyone can follow it. “We will live in a demon-haunted world, full of marvelous devices whose operation we will not understand,” he wrote.

Image: SiliconANGLE/Ideogram

Three of Anthropic PBC’s large language models carried out successful cyberattacks during routine internal tests.

The company detailed the breaches on Thursday. A few days earlier, rival OpenAI Group PBC disclosed a similar incident. Two of the company’s LLMs escaped from an isolated sandbox that was being used to evaluate their cybersecurity capabilities. They subsequently hacked Hugging Face, a popular platform for hosting open-source AI projects.

OpenAI’s disclosure prompted Anthropic to check logs from its own model security evaluations. That review is what led to discovery of the cyberattacks disclosed on Thursday. According to Anthropic, its engineers identified three breaches carried out by three different Claude models.

All three cyberattacks occurred during so-called capture the flag evaluations. During such tests, Anthropic installs a Claude model in a sandbox that simulates the infrastructure of an external company. Claude is tasked with finding a way of stealing data from the simulated organization’s systems.

Anthropic developed the test environments in collaboration with Irregular, an AI security startup. Usually, the companies isolate their sandboxes from the web to reduce the risk of cyberattacks. A configuration error turned on internet access for the three AI model instances that carried out the cyberattacks.

The most severe breach involved Claude Opus 4.7, an LLM that Anthropic released in April. The simulated company that it was asked to hack shared a name with a real website. The model subsequently hacked the organization that operates the website by chaining together multiple vulnerabilities.

Opus 4.7 compromised a production database with several hundred rows of information. Additionally, it obtained access credentials for several applications and infrastructure assets.

The second cyberattack was carried out by Mythos 5, Anthropic’s most advanced commercially available model. The LLM wrote a malicious Python package, or code bundle, and uploaded it to a popular open-source project hosting platform. The file was downloaded by a cybersecurity company a few minutes later. The malicious package compromised the firm’s infrastructure and stole access credentials.

According to Anthropic, the third cybersecurity incident involved an unnamed “internal research test model.” It compromised an application using a set of simple hacking methods such as SQL injections. At one point, the model discovered that the application wasn’t a part of its security evaluation sandbox and stopped the cyberattack.

Anthropic is partnering with a nonprofit AI safety lab called METR to carry out a more detailed investigation of the breaches. Additionally, the company plans to improve how it develops and monitors its LLM evaluation sandboxes. 

Image: Anthropic

NXP Semiconductors NV is reportedly in talks to buy Ambarella Inc., a fellow provider of vehicle chips.

The Financial Times today cited sources as saying that other industry players could also make bids. The report didn’t divulge any other details, notably what sale price Ambarella is seeking. The company had a market capitalization of about $3.25 billion before the acquisition rumors boosted its shares by 16%. Any bids the company receives will likely value it higher. 

Netherlands-based NXP is among the world’s largest suppliers of car chips. Its processors can be found in infotainment devices, autonomous driving systems and a variety of other auto parts. NXP is also active in several additional segments including the data center market. Its chips are used for tasks such as distributing electricity to servers and managing the flow of coolant.

Santa Clara, California-based Ambarella competes in many of the same segments as NXP, most notably the auto sector. The company makes chips that enable carmakers to install artificial intelligence software in their vehicles.

Ambarella debuted its newest auto chip in January. The CV7 combines a four-core central processing unit with an image signal processor and a video encoder. The image signal processor enhances footage from the host vehicle’s sensors by removing errors, boosting brightness and performing other edits. The CV7’s video encoder, in turn, compresses footage into a memory-efficient format that can be more easily sent over the network.

The chip also features an artificial intelligence accelerator. According to Ambarella, the module is based on a custom architecture that can provide five times better power-efficiency than competing designs. The technology enables the CV7 to run transformer-based language models.

Carmakers can use the chip to power their vehicles’ ADAS, or advanced driver assistance systems. Such modules provided limited autonomy features such as automated parking and lane centering.

Ambarella offers the CV7 alongside a more advanced processor called the CV3-AD685. It can equip vehicles with L4 autonomy, or the ability to operate without driver input in most situations. The chip features a hardened processing module called a safety island that is optimized to run the host car’s most sensitive code.

Many autonomous vehicles enrich footage from their cameras with radar measurements. Radar sensors work by transmitting radio waves and measuring how they’re reflected by nearby objects. Each reflected radio wave becomes a dot in a point cloud, a three-dimensional map of the car’s environment. 

The CV3-AD685 ships with a set of radar algorithms called Oculii. According to Ambarella, the software adapts the radio waves generated by a car’s radars to the environment in which they operate. That reduces the number of antennas needed to collect radar data while boosting mapping resolution.

Ambarella’s technology could help NXP boost its presence in the autonomous vehicle market. Waymo LLC, Amazon.com Inc.’s Zoox unit and other industry players are currently in the process of growing their self-driving taxi fleets. Ambarella’s chips are also used in more than a half-dozen other segments including the industrial robot market.

Image: Unsplash

Artificial intelligence has quickly become a standard part of modern software development. Coding assistants, code completion tools and AI-powered integrated development environments are now widely available, yet many engineering organizations continue to struggle with the same fundamental challenge: developer productivity.

Approximately 65% of organizations report that engineering teams spend just 0–20% of their time on net-new innovation. The majority of developer capacity is still consumed by maintenance, migrations, reviews, operational toil and context switching. The problem is no longer access to AI tools but how organizations redesign AI-native software development around them.

In the latest episode of the AppDevANGLE podcast, Deepak Singh, vice president of developer agents and experiences at Amazon Web Services Inc., and Steve Tarcza, director of software development at Amazon, joined me to discuss why the next generation of software development is shifting from AI-assisted coding to AI-native engineering workflows.

AI productivity isn’t a tooling problem

One of the most notable observations from the discussion is that organizations using the same AI tools often achieve dramatically different outcomes.

“We’ve done studies both inside the company and externally,” Singh said. “Some teams are getting 15% to 30% increases in productivity. Others are getting three to 10 times — or even more. They’re using exactly the same tools.”

The difference isn’t the model or the IDE. According to Singh, the highest-performing teams rethink software development itself. Rather than inserting AI into existing workflows, they redesign planning, specifications, reviews and handoffs so AI agents become active participants throughout the development lifecycle.

That represents a meaningful shift in enterprise software engineering. AI is evolving from a coding assistant into a collaborative engineering system.

Context is becoming the new source code

As AI agents take on increasingly complex work, context is becoming one of the most valuable assets engineering organizations possess.

Foundation models understand programming languages, but they don’t understand an organization’s architecture, coding standards, operational practices or business priorities.

“What the AI doesn’t know is how you work,” Tarcza explained. “The teams that focus on getting that knowledge written down — whether it’s steering files, documentation or specifications — unlock the agents to take on much more work.”

This reflects a broader trend emerging across enterprise AI.

Organizations are beginning to realize that prompts alone are insufficient for production software development. AI agents require structured knowledge, engineering intent and reusable organizational context to consistently produce high-quality results.

That knowledge is increasingly becoming a strategic engineering asset.

Trust is the foundation of AI adoption

Another recurring theme throughout the conversation was trust.

While AI models continue to improve rapidly, organizations won’t allow autonomous agents to operate at scale unless engineers trust both the process and the output.

“Trust is the currency of AI adoption,” Tarcza said. “If you can’t trust the agents, nobody’s going to use them.”

AWS is approaching this challenge by emphasizing specification-driven development, structured engineering context and automated reasoning techniques that identify ambiguity before code generation begins.

The objective isn’t simply to generate software faster, but to generate software that developers are confident deploying into production.

That distinction becomes increasingly important as organizations begin allowing AI agents to execute longer-running development tasks with less human oversight.

AI-native software development expands beyond coding

Perhaps the most significant takeaway from the discussion is that AI is expanding well beyond writing code.

Within Amazon, engineering teams are already using AI agents to prioritize work, summarize Slack conversations, analyze tickets, generate specifications and automate portions of daily engineering operations.

These systems function less like coding assistants and more like engineering teammates.

Tarcza shared one example in which an Amazon retail feature called “Add to Order” was delivered two months earlier than originally projected after the team shifted to spec-driven development, placing AI at the center of planning, execution and implementation rather than simply using it for code generation.

This evolution suggests the future of software engineering may be defined less by how quickly developers write code and more by how effectively humans and AI agents collaborate throughout the software delivery lifecycle.

The bottom line

The first wave of generative AI focused on accelerating individual developer tasks. The next wave is transforming software engineering itself.

Organizations that simply layer AI onto existing workflows will likely continue to realize incremental gains. Those willing to redesign engineering around specifications, trusted context, autonomous agents and AI-native processes may unlock far greater improvements in productivity and innovation.

The competitive advantage is shifting from adopting AI tools to building organizations that know how to work alongside them.

Here’s the complete conversation with Deepak Singh and Steve Tarcza, part of the AppDevANGLE podcast series: 

Image: SiliconANGLE/ChatGPT

Israeli network chip startup Xsight Labs Ltd. said today it has closed on a $300 million funding round that brings its valuation to $2.8 billion. It’s planning to use the funds to help its programmable Ethernet switches and 800-gigabit data processing units push deeper into cloud and artificial intelligence networks.

Today’s round was led by Fidelity Management & Research Co., and saw participation from Aliya Capital Partners, Atreides Management, Artisan Partners, Battery Ventures, Diagonal Capital, Intel Capital, Key1 Capital, Maverick Capital, Sienna, T. Rowe Price, Union Group and Valor Equity Partners..

Investors are making a fairly straightforward bet on Xsight. They believe AI data center networking will balloon into a $150 billion market by 2028, and Xsight’s chips, which prioritize power efficiency and programmability, are positioned to capture a meaningful slice of that growth

Xsight, founded in 2017 and based in Tel Aviv, develops two key components that are used in data center networks – the E1 data processing unit or DPU and the X2 Ethernet switch, which are both designed specifically for AI connectivity. The company’s calling card is power efficiency, with its chips consuming under 200 watts for a 12.8 terabit-per-second switch. It also emphasizes customer-driven software extensions, making its chips programmable rather than fixed-function.

The startup is currently developing new generations of both products, which will introduce support for the latest standards backed by the Ultra Ethernet Consortium. To develop those products, it’s planning to expand its engineering teams across its facilities in Israel, the U.S., Europe and Asia, and increase its manufacturing and supply chain capacity so it can take orders from “tier-1 customers.”

The startup’s E1 chip pairs two 400G ports with 64 Arm-based Neoverse N2 cores built on a 5-nanometer process and 32 megabytes of shared cache. It also features PCIe Gen5 connectivity and encryption engines that run at line rate. Where it differs from other DPUs is that those 64 cores sit directly on the data pathway, so that routing., telemetry and packet inspection occur on every data packet instead of being handed off to a smaller exception-handling engine, enhancing overall network bandwidth with four-times the performance per watt of previous-generation designs. The chip has also achieved SONiC-DASH Hero 800G validation, sustaining more than 14 million connections per second with no dropped packets in that benchmark.

As for the X2 switch, this provides 12.8 terabytes of full-duplex bandwidth at under 200-watts, plus first-bit-to-first-bit latency of less than 700 nanoseconds, with 128 lanes of 100G PAM4 SerDes. It features a reprogrammable data plane that allows it to be reconfigured for different network applications on the fly. According to Xsight, it draws 40% less power than other kinds of 12.8 Tbps switches.

Xsight says the E1’s performance-per-watt is its main advantage. In AI data center racks, where accelerator chips consume the bulk of the power budget and cooling capacity, the watts spent on network fabric are all considered overheads.

Xsight’s most notable customer so far is SpaceX Corp. The company said last year that the X2 has been chosen as the high-speed networking core of Elon Musk’s Starlink V3 satellites, which will be capable of moving more than 1 Tbps of fronthaul traffic and around 160 gigabytes-per-second of uplink capacity. The switch was chosen as much for its performance as its ability to operate in radioactive environments and thermal extremes, the company said.

Xsight’s E1 and X2 have both also been deployed by multiple data center operators globally within edge and network infrastructure, and its products are now being evaluated by a number of Tier-1 hyperscalers. “This valuation is a testament to the team’s relentless execution and reflects the market need for a high-performance alternative to closed, legacy architectures,” said co-founder and Chief Executive Yossi Meyouhas.

The startup is also hoping to increase its presence in AI data centers, as Ethernet is believed to be closing the gap with the proprietary interconnects that link most AI chip clusters today. Last year, the Ultra Ethernet Consortium published its 1.0 specification, which defined a remote-memory-access transport with multipath routing and congestion control that’s designed for AI training and inference traffic instead of general-purpose computing workloads. Xsight says the X2 is one of the first switches to meet the new standard.

Incumbents such as Cisco Systems Inc., Broadcom Inc and Marvell Technology Inc. currently dominate large portions of the data center networking market. The question is whether Xsight’s promise of lower power consumption and competitive cost per token processing will hold up against those player’s deeper customer relationships and broader product portfolios.

Image: Xsight Labs

Object storage software company MiniIO Inc. says it has cracked the persistent memory problem for artificial intelligence agents with the launch of a new offering called AIStor Memory.

Whereas conventional chatbots like Claude and ChatGPT can generally get away with much more limited context because they usually only need to respond to a few prompts, AI agents often carry out much longer, multistep tasks that can span multiple sessions. This means that having persistent memory is a much more pressing concern.

AI agents perform tasks such as creating documents, drafting analyses and answering complex questions that could previously only be done by humans. The resulting outputs, according to MiniIO, are basically organizational knowledge – a record of what a business knows – and believes that controlling these are critical to scaling AI agents.

AIStor Memory is MiniIO’s attempt to enable this. In a blog post, the company explained that it treats agent memory as a native data type that lives alongside standard objects and tables, allowing AI agents to retain context across multiple sessions more easily. This means they can quickly pick up where they left off on unfinished tasks and operate on enterprise data with existing governance controls. According to MiniIO, the resulting knowledge the agents generate will remain on infrastructure controlled by customers themselves.

Co-founder and co-Chief Executive AB Periasamy said all of the knowledge generated by AI agents should live on enterprise-controlled infrastructure. “AIStor Memory brings long-term memory, persistent workspaces and secrets together on a single enterprise-controlled foundation,” he explained. “A single agent’s memory becomes the shared substrate for the entire organization.”

At present, enterprises face an uphill struggle to give agents a persistent memory. They’re required to combine their object storage resources with metadata databases, vector stores, secrets managers, governance tools and synchronization pipelines, and that is a major drain on information technology team’s resources. With AIStor Memory, the company is offering an integrated alternative that can be mounted onto existing sandboxes, which also plays nicely with existing AI infrastructure, tools and frameworks without any modifications required.

The new offering is designed to support long-running multistep agentic workflows, where the work performed by AI agents needs to be able to survive interruptions and where sensitive data must remain fully governed. The company said AIStor Memory can enable specific agentic use cases such as software engineering agents working on large codebases, deep research and analysis that spans days, human-in-the-loop workflows that must be paused and resumed later, and enterprise AI systems that rely on regulated data.

MiniIO explained that AIStor Memory is designed to sit beneath the KV-cache layer served by its own MemKV service, where it can preserve the complete state of an agent’s work, rather than only the context used during the inference process. This enables it to support “infinite context,” with persistent memory scaling based on the available capacity rather than a model’s context window, without the data having to be summarized, truncated or otherwise minimized.

AIStory Memory is tightly integrated with existing infrastructure, which means there’s no need to set up a separate database, vector store, metadata tier or synchronization pipeline, and it employs techniques such as erasure coding, bitrot protection, encryption, compression and fault-tolerance to protect against rack, drive and data center failures.

Because it lives entirely on the customer’s infrastructure and is protected by customer-owned keys, organizations won’t have any concerns about memory leakage, the company said.

Image: MiniIO

Pangram Labs Inc., an artificial intelligence research lab that develops AI detection software, today announced it raised $9 million, led by Menlo Ventures, to improve the accuracy of its core text detection and expand into other media, starting with images.

Haystack, ScOp Venture Capital, Script Capital and Cadenza also participated in the investment round. The funding brings the total raised by the company to almost $13 million after raising $2.7 million in June 2025.

Pangram is best known for its AI detection platform that the company claims is capable of industry-leading 1 in 10,000 false positives.

“Fully AI-generated content is everywhere now, and much of it is undisclosed,” said cofounder and chief executive Max Spero. “Pangram exists to make authorship legible for publishers, teachers, journalists and anyone who relies on the truthful and authentic communication of information.”

The company uses what it calls a “classifier model,” a type of neural network that can estimate whether a portion of text is AI-generated or human-written. According to Pangram’s page on how the product works,  says that the system functions by attempting to determine what a passage “sounds like” an LLM or a human.

Training the classifier involved pulling known-human text drawn from before 2021 and pairing it with AI-generated text so that the classifier could readily distinguish the two. The company acknowledged that this works well now, but language style and use drift over time, and the training model will have to adjust with it.

The company also announced Pangram 4, the company’s most powerful AI detector to date, and introduced Pangram Image detection in research preview.

The company said in internal benchmarks, Pangram 4 has a false positive rate of 0.0041%, or around once for every 24,000 documents. The company added that the new model also greatly reduces false negatives, when it fails to detect AI, and it is robust against humanizers, which attempt to make AI text look human-written.

The company’s Image model for detecting AI-generated images can catch images created by image providers including OpenAI Group PBC’s GPT Image, Google LLC’s Gemini Nano Banana, Midjourney Inc., FLUX, and Grok Imagine, and some AI video providers, including Kling AI Pte. Ltd., Seedance, Google’s Veo and Wan.

The company said that a new breed of AI image detectors is needed because deepfakes and AI-generated images passed as truth are becoming more prevalent. Although other companies have begun to embed invisible watermarks and other markers in their content, for example, Google’s SynthID, these only work on frontier models that embed them.

Even as AI images proliferate, humans are getting worse at detecting them. According to a report from Let’s Enhance, a blog focused on AI creative tools, overall detection rates hover around 63.7%, but for high-capability image generators such as FLUX, rates drop to near 29%. Research showed that distinguishing AI from natural is falling to close to 50% on average, essentially a coin toss.

AI detectors and the reliability gap

Pangram’s detection accuracy and false positive rate claims come from what appears to be primarily internal benchmarks, a technical white paper and a small number of favorable third-party studies. The company wants to set itself apart from other AI detectors because accuracy is meaningful.

A key point to examine for AI detectors is that they are by and large unreliable. MIT Sloan Teaching and Learning Technologies pointed out that the technology is “far from foolproof” and often features high error rates that can lead to false accusations. The Mozilla Foundation found that detector tools are not as reliable as they claim and that AI detectors can be biased against non-native English speakers.

Reliability itself is highly context-dependent, with false positives being the main concern.

While some newer AI detection vendors, including Pangram, claim major improvements, the broader literature on the subject still builds on a foundation that AI detection is a probabilistic signal rather than reliable for actual writing.

The result is that numerous educational facilities have either discontinued or banned the use of AI detectors or provided guidance to professors and teachers that it should be used as a data point and not a verdict. The University of Waterloo discontinued using Turnitin LLC, one of the leading detectors, in September; MIT’s position is that AI detectors just don’t work — educators and professional work should be built on adaptive policies and expectations instead.

Image: SiliconANGLE/Microsoft Designer

Meta Platforms Inc. today announced plans to build an artificial intelligence data center campus in El Paso, Texas.

The project is a collaboration with BlackRock Inc., the world’s largest asset manager. The two companies plan to invest $14 billion in the campus.

Meta and BlackRock expect to finalize their joint venture agreement in the coming days. Upon signing, the Facebook parent will contribute $2.3 billion worth of physical assets including on-site facilities that are currently under construction. BlackRock, in turn, will invest $4.9 billion. The asset management also intends to make a $1 billion payment to Meta.

BlackRock is set to receive a 80% stake in the campus while the social networking giant will own the remaining 20%. Meta will be the site’s sole user when it comes online in 2028. According to the Facebook parent, its lease has a four-year initial term that can be optionally renewed four times.

Meta stated that the initiative’s $14 billion budget covers “total development costs for the buildings” along with power, cooling and networking expenses. Notably, the company didn’t mention the cost of the chips that the campus will host, which suggests the total project price could be higher. Earlier this month, Meta announced plans to spend more than $50 billion on a data center campus in Louisiana. Bloomberg reported that the price tag will be $250 billion when taking into account chips and certain related items.

The Louisiana project has a similar financial structure as Meta’s El Paso development. Last year, the company sold an 80% stake in the former campus to Blue Owl Capital. In return, the investment firm made a $7 billion cash contribution to the project.

Meta expects its El Paso campus to provide 1 gigawatt of AI-optimized computing capacity. The site joins a string of 1-gigawatt data center projects that the company has announced over the past two years. Those campuses are located in Ohio, Indiana and Canada.

Meta expects its three other 1-gigawatt sites to cost around $10 billion. The fact that the El Paso site has a $14 billion budget hints the company may be planning to use a different, more advanced data center design. It’s also possible that Meta will use the campus to power the cloud infrastructure business it’s reportedly building.

The Facebook parent ended  2055 with $72 billion in capital expenses. Meta expects that number to range between $125 billion and $145 billion this year. The cloud business that Meta is believed to be building could ease its efforts to realize a return on investment. Meta already sells access to Muse Spark 1.1, its newest AI model, via an application programming interface that enables other companies to integrate it into their applications.

Photo: Meta

Investors are throwing money at quantum computers again, with the Swiss startup ZuriQ AG becoming the latest beneficiary of their largesse after closing on a $25.5 million seed funding round. The investment was led by Quantonation and saw the participation of Forward.one, Extantia, Firgun Ventures and all of the startup’s existing backers, building on a previously undisclosed $4.2 million pre-seed round that closed when it spun out of ETH Zürich last year.

ZuriQ is the latest in a long line of startups trying to solve the problem of scaling quantum computers, which continues to befuddle some of the world’s brightest minds. The problem is that the fundamental building blocks of quantum computers, known as “qubits,” are extremely sensitive to environmental noise. Qubits are so sensitive, in fact, that even something as faint as the vibration caused by a pin dropping on the floor could be enough to disrupt their state, leading to errors in quantum calculations.

Trapped-ion systems have long been positioned as a potential solution to the problem of qubit instability, yet the companies that have attempted to build them have struggled to achieve the scale required to build systems large enough to perform meaningful calculations. According to ZuriQ’s co-founder and Chief Executive Dr. Pavel Hrmo, the reason is that they’re still reliant on a legacy architecture that was first proposed over two decades ago. He explained that these legacy designs are based on ions held in one-dimensional, single-file chains stitched into large grids using complex junctions. But this architecture severely restricts their ability to scale. As such, trapped-ion systems have only been able to grow one qubit at a time, preventing anyone from creating the thousands of qubits needed to build a commercially viable quantum computer.

“The trapped-ion companies that started the race began with one-dimensional designs that were useful stepping-stones, but the real challenge is whether they can successfully pivot to two dimensions as they attempt to scale,” Hrmo said. “We spent longer in the lab, and that time allowed us to identify an alternative route that is inherently easier to scale. Our architecture is two-dimensional from the ground up, so the number of qubits we can place on a chip will grow far more readily than in systems built on a legacy blueprint.”

ZuriQ’s processors are instead based on two-dimensional designs. Instead of using the rapidly oscillating fields found in conventional systems, it uses Penning micro-traps combined with a static magnetic field, allowing its trapped ions to move freely in any direction, without the constraints imposed by the complex junctions essential to one-dimensional architectures. Because ZuriQ’s architecture is natively 2D, it means it can grow its qubit count across the entire surface area of a chip, as opposed to only growing in a straight line.

Professor Jonathan Home of ETH Zürich, a scientific adviser to ZuriQ, said the secret sauce behind the startup’s approach is geometry. “Hold ions in a line and the count grows one at a time; hold them in two dimensions, and it grows with the area of the chip — on a standard chip, that is the difference between tens of ions and many thousands,” he said. “Just as important, the ions can be moved freely in three dimensions, so they can be connected together far more flexibly, and that connectivity is what ultimately makes a quantum computer more capable.”

Given that scale is the primary challenge ZuriQ wants to solve, it’s no surprise that its overriding focus now is to show that it can actually scale its 2D trapped-ion qubit architecture. So far, the company has already developed a working prototype to demonstrate its technology. It built a three-by-three array of nine individually controlled ions, fabricating it in collaboration with a company called Infineon AG using established chipmaking processes to show that it’s possible.

Going forward, the startup now wants to show it can dramatically increase the number of ions on a chip and develop a processor that can host hundreds and later even thousands of qubits. To that end, the funds from today’s round will be directed towards advancing its research, scaling its chip fabrication operators and hiring more experts from rivals in the quantum computing industry.

Quantonation’s founding partner Christophe Jurczak said that many people are of the opinion that the winners of the quantum computing race are already known, when the truth is that it’s far from settled. “ZuriQ is demonstrating that there remain significant and transformational physics breakthroughs still to be made in quantum architectures,” he insisted.

Images: ZuriQ

Nvidia Corp. today announced the launch of the Open Secure AI Alliance, a new organization founded by technology, cloud computing and cybersecurity leaders to build and share open artificial intelligence tools.

As the capability and strength of AI tools grows and shapes the technology industry, it is becoming instrumental for safety, security and trust to remediate and secure vulnerabilities within this emerging software. The company said just as open source created a shared foundation for software, AI security now faces an equally important critical mass where too much of the infrastructure is closed, opaque and difficult to tap into by the larger community.

A who’s who of leaders from cloud computing, cybersecurity, enterprise software, open-source foundations and AI research joined the inaugural launch of OSAA. In addition to Nvidia, founding partners included Adobe Inc., Cisco Systems Inc., Cloudera Inc., Cloudflare Inc., Databricks Inc., Hugging Face Inc., OpenClaw, IBM Corp., Red Hat, Salesforce Inc., Snowflake Inc. and Thinking Machines Lab Inc., among almost two dozen more.

The mission of the OSAA is to ensure that defenders have access to open, frontier tools that they can trust and control.

This alliance forms even as the industry struggles to handle the emerging use of AI as both cybersecurity tools for defense and attack. AI agents, autonomous software capable of operating on its own with little or no human oversight, are becoming a mainstay of cybersecurity operations. This happens even as the enterprise continues to infuse AI into every layer of the technology stack, from building code and debugging to maintenance, operations and customer-facing interfaces and websites.

Recently, Hugging Face, the open-source AI model and tool repository, suffered a breach from a then-unknown attacker and attempted to use current centralized closed-source frontier AI tools to defend itself – however, safeguards on these models blocked it from doing the necessary analysis. As a result, the company pivoted to the open-source Z.ai GLM 5.2 to develop a rapid response and repel the attack.

A week later, OpenAI Group PBC revealed that one of its own frontier models running agentic capabilities had escaped containment and hacked into Hugging Face while under testing and evaluation. This made headlines primarily because it is the first example of a frontier model reaching out of its own sandbox and executing a cyberattack on its own.

This is not the first time AI models, including frontier models, have been used by cyberattackers to enhance and execute attacks. Autonomous software driven by powerful AI models can adapt and react faster than humans, orchestrate and control wide varieties of tools using large datasets of vulnerabilities and exploits to determine critical paths to attack networks. Already, a growing number of cyberattacks are AI-assisted or driven, with 87% of organizations reporting experiencing AI-related attacks in the past year, according to AllAboutAI.

AI tools are not only becoming a mainstay of defense; the software itself is becoming a new vulnerable surface that must be secured against attackers. As the industry evolves, Nvidia and the new alliance argue that AI safety will require using the full agent stack – identity, permissions, harnesses, guardrails, logs and evaluation – to understand and respond to emerging threats. Open harnesses, tools and other elements will matter just as much as open models when equipping defenders with the ability to customize and control their capacity to inspect, test, improve and share knowledge.

Alliance membership brings open tools to the defense

Nvidia is contributing open models, model weights, and new agent harness research to the OSAA with the intent of seeding the future development of cybersecurity tools and techniques.

This includes the new open-source Nvidia Labs Object-Oriented Agent project, now available on GitHub, which enables the development of advanced AI safety capabilities for agentic control systems. Research from this project is building frameworks to enable harnesses to better integrate with models and control agent behavior for testing, auditing, tracing and responding to cyberattacks.

Other alliance members are also offering open tools, including HPE contributing the zero-trust identity framework SPIFFE/SPIRE; Hugging Face offering Safetensors, a safe format for storing and sharing model weights; and Microsoft contributing MDASH, a multi-model agnostic scanning harness designed to orchestrate AI agents to discover, debate and dissect exploitable bugs.

The new alliance of industry leaders and heavyweights will also work to educate and inform policymakers and regulators about the state of AI security. In the announcement, Nvidia noted that it will be crucial to understand the necessity of open models, harnesses and security.

This comes in the wake of the United States Treasury threatening to sanction open-weight Chinese AI model makers and pressure to curtail or ban the use of open-source tools. Nvidia, Microsoft, IBM, Meta Platforms Inc. and more than a dozen other tech firms penned an open letter last week calling upon policymakers to take care in regulating open-source models and avoid banning them outright. The letter argued that such a move could have a significantly negative effect on the industry.

For example, open-weight models and open tools promote community ecosystems for sharing, creating and evolving technology – a critical element of innovation that could be stifled if open-source model and tool access was curtailed. Open-weight models also allow companies to readily customize and run AI on their own infrastructure, allowing them greater control over sensitive data and computation.

Image: SiliconANGLE/Microsoft Designer

As AI expands into robotics and other real-world applications, the focus is shifting toward integrated platforms that simplify development and deployment. Open ecosystems and standardized architectures are emerging as key priorities for the future of physical AI computing.

Agentic AI is creating the same requirements and challenges across markets, all of which demand a significant amount of compute. Whether it’s a humanoid robot or another type of autonomous robot, the need extends beyond GPU computing alone, according to Kirk Saban (pictured), corporate vice president of product marketing and management at Advanced Micro Devices Inc.

“You need GPU compute, but you need CPU compute as well for all of the autonomous functions,” Saban said. “It’s really the ultimate application of agentic AI, the ultimate embodiment, is a humanoid robot.”

Saban spoke with theCUBE’s Dave Vellante and John Furrier at the AMD Advancing AI event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed physical AI computing and how AMD is using an open, developer-centric platform to accelerate robotics innovation. (* Disclosure below.)

Physical AI computing in focus

AMD’s embedded business serves a broad range of industries, including healthcare, agriculture, automotive and industrial IoT. That broad reach creates opportunities for across a wide variety of applications.

“Part of the flexibility and really the value prop here is that with the FPGA technology that we have for sensor aggregation, it doesn’t matter what market you’re going after,” Saban said. “We can support that, we can scale it.”

The same hardware can be reprogrammed and repurposed for multiple applications, providing flexibility for sensor aggregation. Rather than relying on a custom chip, developers can tailor the platform to support different sensors, cameras and MIPI interfaces, according to Saban.

“That is really part of our entire value prop,” Saban said. “We can service the entire robot spectrum, spine, brain, joints, we now have the complete solution.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the AMD Advancing AI event:

(* Disclosure: TheCUBE is a paid media partner for the AMD Advancing AI event. Neither AMD, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE

HPC AI infrastructure is no longer a parallel track to traditional computing. They are merging into a unified computing foundation that the supercomputing world has been building toward for years.

The same infrastructure once reserved for national laboratory modeling and simulation has become the foundation for enterprise AI factories, and agents are already running scientific simulations autonomously, said Trish Damkroger (pictured), senior vice president and general manager for HPC and AI infrastructure solutions at Hewlett Packard Enterprise Co.

“Those things that we call supercomputers are now the exact same infrastructure that you need for modeling and simulation that you need for AI,” Damkroger said. “I’m finally seeing a lot of this convergence, where you have AI steering your HPC workflows, or agents running your simulations.”

Damkroger spoke with theCUBE’s Dave Vellante at the AMD Advancing AI event, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed the deepening AMD and HPE partnership, liquid cooling as an enterprise imperative and why sovereign AI is expanding beyond national laboratories. (* Disclosure below.)

HPC AI infrastructure converges around liquid cooling, sovereignty and AMD partnership

HPE’s relationship with AMD encompasses the most powerful HPC AI infrastructure deployments in the world. The Frontier system at Oak Ridge National Laboratory, the first to reach hyperscale, is an AMD and HPE platform, and HPE is under contract to deliver Oak Ridge’s next Discovery system, also AMD-based.  Lisa Su, president and CEO of AMD, announced both the EPYC 6 processor and AMD Instinct MI430X accelerator at the event, reinforcing an AMD and HPE partnership that Damkroger said extends from national labs to regulated commercial enterprises.

“We have an amazing relationship with AMD,” Damkroger said. “The number two system on the TOP500 is also an AMD system, El Capitan at Lawrence Livermore, where I came from.”

As silicon power requirements continue to soar, liquid cooling is becoming essential. HPE’s Cray heritage gives it a deep foundation in that space, and its GX5000 platform, part of the Discovery system, is designed to handle high thermal design power parts while meeting European requirements for warm-water cooling at up to 45 degrees Celsius. Damkroger said that the same learning curve that shaped HPC AI infrastructure is now a complication for mainstream enterprise data centers and organizations that drag their feet will face mounting cost and density disadvantages as AI silicon continues to scale.

“You’d have one server in a rack if you’re going to air,” Damkroger said. “It’s going to get to the point where it’s going to be ridiculous.”

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the AMD Advancing AI event:

(* Disclosure: TheCUBE is a paid media partner for the AMD Advancing AI event. Neither AMD, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE

Meta Platforms Inc. today made its latest large language model available to consumers through its Meta AI chatbot.

The update is rolling out alongside several enhancements to Facebook. Meta’s flagship social network is receiving a new video-based account verification tool. Additionally, users can now access the Facebook Marketplace e-commerce platform through a mobile app.

Meta AI is a free AI service with a similar feature set to ChatGPT. Today’s update switched the chatbot to Muse Spark 1.1, an LLM that Meta introduced earlier this month. The model supports prompts with up to 1 million tokens and can create groups of AI agents to tackle complex work.

According to Meta, the new version of Meta AI is better at tasks that require it to retrieve data from multiple sources. If the chatbot is asked to plan a trip, it can research destinations and check the user’s calendar for scheduling information. A consumer working on a kitchen renovation can ask Meta AI to find online furniture stores.

The new version of chatbot also lends itself to more complex research tasks. It can spend up to several minutes collecting information about a user-specified topic. Meta AI aggregates data from research papers, social networks and other sources. Users can optionally have it visualize the retrieved data.

Rounding out the list of AI enhancements is a new notification capability. Meta AI can inform users about events of interest and generate recurring alerts such as daily weather updates.

On launch, Muse Spark 1.1 is available to users in a limited number of markets via the web and mobile versions of the chatbot. Meta plans to bring the model to the version of Meta AI embedded in WhatsApp within a few weeks. 

The company is rolling out the chatbot upgrade alongside two new Facebook features called Facebook Verified and Seller.

The former tool enables users of the social network to verify their accounts by uploading a video selfie. Consumers who complete the process receive a checkmark badge that appears on their profiles. According to Meta, Facebook Verified is available at no charge to users who are over 18 and don’t breach its terms of service.

The other highlight of today’s Facebook update is Seller, a mobile app for users who sell items on Facebook Marketplace. It includes an AI tool that speeds up product listing creation. The feature generates marketing copy, displays pricing suggestions and organizes each newly created listing under the relevant category.

Like the new version of Meta AI, Facebook Verified and Seller will roll out gradually. Meta plans to pilot the features in a limited number of markets before making them available worldwide. 

Image: Meta

Database giant Oracle Corp. has signed a new, multibillion-dollar deal with the Pentagon that will see it provide its software to branches of the U.S. military, Coast Guard and intelligence community over the next 10 years.

The deal, valued at almost $7 billion, includes a five-year base period for both perpetual and subscription-based software licenses, maintenance and consulting, Pentagon officials said.

Kirsten Davies, who serves as the Department of Defense’s chief information officer, said that her agency managed to negotiate a contract that will result in savings of at least $441 million for taxpayers by “fundamentally improving how we procure on-premises Oracle capabilities.”

Investors reacted positively to the deal. Oracle’s stock, which has come under pressure this year amid fears that artificial intelligence will disrupt traditional software vendors, gained more than 3% on news of the contract.

The deal stems from a push by the Trump administration to rein in DOD spending on software services and licenses. Since taking office last year, Defense Secretary Pete Hegseth has detailed multiple efforts to cut the agency’s technology spending by eliminating duplicate contracts and non-essential software agreements for the Pentagon’s information technology systems.

A 2025 memo called for the DOD to create a plan of action about how it will negotiate more favorable prices on cloud and on-premises software, so that it doesn’t pay any more than other American enterprises. However, while some may consider the savings made by the Pentagon to be a win, it’s a figure that pales in comparison with the estimated $37.5 billion the U.S. has spent on the Iran war since February.

Oracle has had a long history of serving the U.S. government. In fact, the Central Intelligence Agency was its first ever customer. Back in 1977, co-founders Larry Ellison (pictured), Bob Miner and Ed Oates were awarded a $50,000 contract to build the CIA a relational database management system under the codename “Oracle,” and that project ultimately inspired the name of the company they were to establish.

Ellison, who is now Oracle’s chief technology officer and executive chairman, is a longtime supporter of U.S. President Donald Trump, and reportedly contributed $45 million towards his 2024 presidential campaign via a donation to a nonprofit group. He was notably one of the first guests to visit the White House following the start of Trump’s second term, using the occasion to announce his company’s involvement in $500 billion Stargate initiative to build AI data centers in the U.S. Trump was also a supporter of Oracle’s push to secure a 15% stake in TikTok’s U.S. business.

Despite the strong government backing Oracle has received, its stock has taken a battering this year due to the rise of AI software. In addition, Oracle is also taking on tens of billions of dollars in debt as it races to become an AI infrastructure player by building out its network of global data centers.

In its most recent earnings report last month, Oracle revealed that its software revenue declined 2% from the previous year. However, its cloud revenue remains a bright spot, growing 47% thanks to strong demand for compute resources from its AI customers.

Photo: Oracle

Networking and security company Cato Networks Ltd. today unveiled a partnership with CrowdStrike Holdings Inc. to integrate their platforms and give security teams a single view of network and endpoint data.

The integration ties the Cato SASE Platform to the CrowdStrike Falcon platform. Analysts get endpoint alerts and network activity side by side, which the companies say should shorten investigations and speed response.

Network and endpoint monitoring usually sit in separate tools. An analyst chasing an alert in one has to go dig up the matching activity in the other, which is slow and leaves room to miss things. Cato and CrowdStrike say wiring the two feeds together removes that manual step and gives detections more context to draw on.

The integrations cover three areas. Cato XOps and CrowdStrike Falcon Discover correlate endpoint detections with network telemetry to sharpen threat visibility and accelerate investigations. Cato Asset Security and Falcon Discover combine device intelligence with endpoint context for a fuller view of managed assets. The Cato SASE Platform also streams network telemetry into Falcon Next-Gen SIEM, letting analysts hunt threats and build detections from a broader set of signals.

“Security outcomes, especially in the AI era, can only improve when fueled by rich and connected data,” said Karl Soderlund, global channel chief at Cato Networks. “This partnership further enhances Cato’s ability to deliver complete visibility, full context and agentic control to make security operations teams more effective in defending against the next wave of AI threats.”

Chris Stewart, vice president of global cloud and technology alliance partners at CrowdStrike, said integrating the two platforms lets customers correlate network and endpoint telemetry, streamline investigations and respond to threats faster.

The integrations also reach customers through distribution. Francisco Criado, senior vice president of security, cloud and AI at TD SYNNEX Corp., said the collaboration gives channel partners a more integrated approach to security that simplifies deployment and scales more easily.

The technical integrations are generally available to customers worldwide through the CrowdStrike Marketplace.

Cato Networks is a venture capital-backed company that was valued at more than $4.8 billion in its most recent funding round of $359 million in June 2025.

Image: Cato Networks

Enterprise intelligence depends on data that is available, accessible and connected.

Graph databases prioritize relationships, connecting data with intelligent context that models and agents can better understand. The Neo4j Graph Platform uses that technology to help developers create applications based on graph-powered systems, with use cases ranging from fraud detection to financial services. Those capabilities will be the primary focus at Neo4j’s GraphTalk event, taking place in San Francisco on July 23.

“GraphTalk underscores one of the most important trends emerging in enterprise AI,” said Paul Nashawaty, principal analyst at theCUBE Research. “Context is becoming the new competitive advantage. Graph technologies are increasingly serving as the connective tissue for AI agents, GraphRAG architectures and decision intelligence platforms, enabling organizations to move beyond generating answers toward delivering trusted, explainable outcomes.”

Neo4j’s GraphTalk event brings together leaders in enterprise AI and data context. On July 29, theCUBE, SiliconANGLE’s livestreaming studio, will present post-event coverage featuring expert insights on how Neo4j and graph technology are shaping autonomous systems. (* Disclosure below.)

Neo4j challenges Palantir on AI intelligence

TheCUBE’s research has found that while AI adoption is accelerating, organizations continue to struggle with data fragmentation and knowledge discovery across the software ecosystem, according to Nashawaty. Neo4j’s tools bring those disparate islands of data together, strengthening the infrastructure of autonomous systems.

Most recently, Neo4j acquired GraphAware, an intelligence analysis software company for government agencies, in a bid to compete with Palantir. The resulting AI-powered graph solutions are intended to be an alternative to Palantir Gotham, paving the way for a graph intelligence platform that can handle critical workloads from government and enterprise clients.

“For over a decade, governments have utilized intelligence analysis software from Palantir,” said Emil Eifrem, founder and chief executive officer of Neo4j. “Now they will have a legitimate choice with GraphAware Hume, powered by the Neo4j Graph Intelligence Platform. We aren’t just launching a new intelligence analysis alternative to Palantir Gotham. We’re acquiring a proven solution that’s built on open-standards technology and already trusted by government agencies throughout the world.”

Government agencies require a reliable data foundation for their AI systems and specialized expertise for use cases such as law enforcement or cyber defense. GraphAware Hume is designed for these types of mission-critical environments.

“The future of intelligence analysis is undeniably graph-powered and AI-assisted,” said Michal Bachman, founder and CEO of GraphAware. “Having partnered with Neo4j for over 10 years, integrating GraphAware Hume into the Neo4j ecosystem is an extremely exciting next step. We have long relied on Neo4j’s platform to put the power of connected data into the hands of analysts. By joining forces, we will drive tighter integration and faster innovation, delivering game-changing AI capabilities to our existing joint customers and new organizations worldwide.”

TheCUBE event livestream

Don’t miss theCUBE’s post-event coverage of GraphTalk on July 29. Plus, you can watch theCUBE’s exclusive content on demand after the livestream.

How to watch theCUBE interviews

We offer you various ways to watch theCUBE’s post-event coverage of GraphTalk, including theCUBE’s dedicated website and YouTube channel. You can also get all the coverage from this year’s series on SiliconANGLE.

TheCUBE podcasts

SiliconANGLE’s “theCUBE Pod” is available on Apple Podcasts, Spotify and YouTube, which you can enjoy while on the go. During each podcast, SiliconANGLE’s John Furrier and Dave Vellante unpack the biggest trends in enterprise tech — from AI and cloud to regulation and workplace culture — with exclusive context and analysis.

SiliconANGLE also produces our weekly “Breaking Analysis” program, where Dave Vellante examines the top stories in enterprise tech, combining insights from theCUBE with spending data from Enterprise Technology Research, available on Apple Podcasts, Spotify and YouTube.

Guests

During theCUBE’s post-event coverage of GraphTalk, industry experts discuss how graph technology is advancing enterprise AI, strengthening data context and supporting autonomous systems. The conversations feature leaders from Adobe, Boeing, Intuit, GoFundMe and other organizations.

(* Disclosure: TheCUBE is a paid media partner for the Neo4j GraphTalk event. Neither Neo4j, the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Image: SiliconANGLE

French lawmakers today voted in favor of banning social media use for children under the age of 15, making it the first nation in the European Union to introduce an age limit in line with legislation that is slowly taking hold globally.

President Emmanuel Macron, currently serving his last term, welcomed the ban, calling it a “major step forward.” He added, “France is leading the way in Europe when it comes to protecting our children and teenagers. The Constitutional Council must now rule on it, and then it will be time to take action to make this measure a reality and protect our children online.”

Following Senate approval, the country’s National Assembly members passed the bill by 279 votes to 81.

After years of widespread criticism concerning how social media is used by youngsters and how it affects their mental health, Australia made history when it became the first nation to wrestle under-16s off social media in 2025 after landmark legislation was passed in 2024. The United Kingdom has also announced a ban that will take effect next year, while Denmark, Greece and Spain are expected to follow. Separately, the EU is considering a wider ban across its 27 member states.

In France, Macron instructed the French government to expedite the bill’s passage through parliament so the legislation would be in place before the new school year begins in September. Digital Minister Anne Le Henanff told the press that the ban shouldn’t be a problem because “age-verification tools already exist,” adding, “If someone is under 15, the account will be closed.”

France’s public health watchdog has been very critical of apps such as Instagram, TikTok and Snapchat. It said last year that they have been proved to be harmful to children and addictive, one of the reasons for a decline in mental health in the country among the young, currently being described as a “pandemic.”

“The brains of our children and our teenagers are not for sale,” President Macron said earlier this year, taking aim at big tech firms. “The emotions of our children and our teenagers are not for sale or to be manipulated, neither by American platforms nor by Chinese algorithms.”

As bans spread worldwide, there has been criticism from speech advocates who maintain there are better solutions than blanket bans and that some applications can be beneficial to children. French senators agreed, voting for a two-tier system wherein some platforms would be banned and some allowed with parental consent, but it is the lower house’s blanket ban that will be implemented.

Photo: Unsplash

A federal judge today approved a $1.5 billion settlement awarded to authors and publishers whose works were used by Anthropic PBC to train the AI chatbot, Claude, in what is the largest copyright class action settlement in history.

Judge William Alsup issued a preliminary approval of the settlement last year, though he has since retired, with U.S. District Judge Araceli Martínez-Olguín today calling the settlement fair and adequate.

“The $1.5 billion settlement provides substantial benefits to the class in light of the novel claims asserted,” she wrote. “Success at trial was not assured, and a loss would have left the class with no recourse.”

The payout means authors and publishers will receive $3,000 for each of the roughly 500,000 works that Anthropic copied from pirate libraries while developing its AI chatbot, Claude. The plaintiffs alleged that Anthropic trained its models on hundreds of thousands of copyrighted books obtained from illegal piracy websites such as Library Genesis and Pirate Library Mirror.

“Anthropic has attempted to steal the fire of Prometheus,” wrote the plaintiffs. “It is no exaggeration to say that Anthropic’s model seeks to profit from strip-mining the human expression and ingenuity behind each one of those works.”

The company is now required to destroy all the pirated material.

The central issue that has spawned a plethora of lawsuits — whether it is legal to train AI models on copyrighted material — has not yet been definitively resolved. Judge William Alsup sided with Anthropic, ruling that training its AI on copyrighted books constituted fair use. However, he also held that obtaining those books from pirate websites fell outside the protection of fair use and therefore infringed copyright.

Countless lawsuits are still awaiting resolution, with companies including Google LLC, Meta Platforms Inc., Midjourney Inc., Perplexity AI Inc. and OpenAI Group PBC all hoping for a favorable outcome.

“We reached this settlement in 2025, after the court’s landmark ruling that training AI on books is fair ​use under copyright law — which remains the law today,” Anthropic Deputy General Counsel Aparna Sridhar said in a statement after today’s judgment. “We are pleased that more than 91% of authors and publishers covered by the settlement have claimed their share of the payment, and we’re looking forward ​to bringing this matter to a close.”

Photo: Unsplash

Artificial intelligence companies are racing to control the infrastructure, data and software layers that will power enterprise intelligence.

Nvidia Corp. remains far ahead in accelerated computing, but Advanced Micro Devices Inc., Broadcom Inc. and other challengers are positioning themselves for a market in which demand may support multiple winners. At the same time, IBM Corp. is under pressure to prove that its software, data and hybrid cloud assets can secure a place in the emerging AI stack.

The next competitive advantage will come from turning proprietary data and domain expertise into an internal system capable of reasoning and executing work, according to John Furrier (pictured, left), executive analyst at theCUBE Research.

“The moat of a business in the future is the brain of the company, the data that they have,” Furrier said. “Can it be organized in a way to reason — system of intelligence, systems of execution, systems of agency? Can they reason?”

On the latest episode of theCUBE Pod, Furrier and Dave Vellante (right), chief analyst at theCUBE Research, discussed IBM’s execution challenges, Nvidia’s infrastructure lead and the growing opportunities for AMD and Broadcom. They also examined open versus closed AI models, token economics and the race to build the enterprise system of intelligence.

IBM faces an AI execution test

IBM’s sharp stock decline prompted speculation that enterprise AI adoption is progressing more slowly than expected. IBM’s performance says more about the company’s portfolio and execution than the health of the broader AI market, according to Furrier and Vellante.

Much of today’s infrastructure spending is flowing toward hyperscalers, neocloud providers and chipmakers. IBM’s mainframe and infrastructure businesses are not as closely aligned with that buildout, while its newer AI offerings have yet to grow enough to offset pressure on older products.

“Their infrastructure business is not aligned with the wave in the industry. They’ve got mainframes,” Vellante said. “What always happens in these waves is the new is not big enough to offset the decline in the old, and I think that’s what’s happening with IBM.”

IBM has many of the assets needed to participate in enterprise AI, including Red Hat, watsonx, governance software and a broad data portfolio. The missing piece is a unified system of intelligence that connects data, context, reasoning and applications. Databricks Inc., Snowflake Inc. and the hyperscalers are already competing for that position. IBM has the components, but it has not assembled them into a platform that commands the same attention, according to Vellante.

“They have all the ingredients, but they’re not putting them together that way,” he said. “Their software portfolio could be. They just need a little bit better focus, in my opinion.”

Furrier dismissed calls to break up IBM, describing its problem primarily as one of execution. Enterprises may also be delaying software decisions while they rebuild data pipelines and prepare infrastructure for agentic workloads.

“I think there is nothing wrong with IBM other than execution,” he said. “I think they could have been more data specific in that system of intelligence. They had all the piece parts. That’s just focus, right?”

Nvidia’s lead leaves room for challengers

Nvidia remains the dominant force in accelerated computing, supported by its graphics processing units, networking, software and rack-scale architecture. Vellante expects the company to retain between 75% and 80% of the AI accelerated computing market. That dominance does not eliminate opportunities for AMD and Broadcom. Demand for AI infrastructure is growing quickly enough that competitors can build large businesses without displacing Nvidia.

AMD has expanded beyond central processing units through graphics accelerators, networking and systems capabilities. Its acquisitions of Xilinx Inc., Pensando Systems Inc. and ZT Systems have helped transform the company into a broader infrastructure provider.

“[AMD CEO Lisa Su] is basically compressing 20 years of ecosystem development by Nvidia, and she’s compressing that into five years of capital allocation,” Vellante said. “It’s actually remarkable what she’s done when you think about that, because she recognizes she’s got to move at the speed of Nvidia.”

Broadcom is pursuing a different opportunity through custom silicon and networking. Furrier expects more alternatives to emerge as enterprises seek smaller, less expensive inference systems that can operate outside hyperscale data centers.

“Everything that AMD and Nvidia makes will sell; Nvidia specifically because they are the leader,” Furrier said. “But the enterprise, they have to start thinking about the budget for the AI infrastructure because they don’t have the big bucks.”

Companies race to build enterprise intelligence

The greatest value may ultimately sit above the infrastructure layer. Companies will need to connect proprietary information, domain knowledge, models and applications into systems capable of reasoning and acting.

That enterprise brain will require relational, vector and graph databases, governed data pipelines and specialized models working alongside general-purpose models. Furrier questioned why companies would outsource all of that intelligence to a small number of model providers.

“If you’re a company and you have to build the next 20-year competitive advantage, you’ve got to build the company brain, the intellect, intelligence, cognition for the company,” he said.

OpenAI, Anthropic PBC and other frontier model companies will need to expand deeper into software and build ecosystems around specialized models, according to Furrier and Vellante. Open-weight models will also increase competition, though they may not match frontier systems in both capability and efficiency.

Enterprises are meanwhile becoming more disciplined about AI costs. The focus is shifting from consuming the largest number of tokens to measuring how efficiently AI produces a useful result. Companies may begin using token-to-value ratios to connect model consumption with measurable outcomes, according to Furrier. That shift could create an AI version of financial operations, with organizations monitoring spending, performance and return in real time.

“Right now, in real time, you can actually peg the value to the outcome and be like, OK, you built an app, everyone’s using it. That’s good. Or you built an app and no one’s using it,” Furrier said.

Human workers will remain involved, but their roles will change. Rather than completing every task themselves, people will manage context, validate outputs and supervise AI systems. The winners will be the companies that combine infrastructure, intelligence and human judgment into a reliable operating model, according to Furrier and Vellante.

The conversation also previewed theCUBE’s upcoming coverage of AMD’s Advancing AI and Neo4j’s GraphTalk events. Stay tuned for more reporting and interviews as theCUBE continues tracking the infrastructure, data and software layers shaping enterprise AI.

Here’s the full episode of this week’s theCUBE Pod:

Image: SiliconANGLE

Alibaba Group Holding Ltd. today previewed the most powerful artificial intelligence model in its Qwen family, Qwen3.8, and claimed the system trails only Anthropic PBC’s Claude Fable 5 among the world’s frontier models.

The company revealed the model, offered as a preview build called Qwen3.8-Max-Preview, today at the World Artificial Intelligence Conference in Shanghai. It accommodates 2.4 trillion parameters, making it the first Qwen model above 1 trillion parameters to process images, video and documents alongside text.

Alibaba’s Qwen team said in a post on X that the model ranks “second only to Fable 5,” a reference to Anthropic’s recent release. It offered no benchmark scores to support the claim and no independent evaluation has verified the ranking.

That absence stands out. Alibaba’s previous flagship, Qwen3.7-Max, shipped in May with a full set of published results, including a score of 56.6 on the Artificial Analysis Intelligence Index. The Qwen3.8 preview arrived with no model card, no activated-parameter count and no benchmark data at all.

Alibaba published no task-level comparison with Qwen3.7-Max and described the new model only as “continuously evolving.” Open weights are promised “soon,” though the company has not set a date or published license terms.

The preview is available now through Alibaba’s Token Plan subscription and its Qoder and QoderWork developer platforms, priced at 10% of the standard rate during the trial period.

The timing places Qwen3.8 directly against a domestic rival. Chinese AI startup Beijing Moonshot AI Technology Co. Ltd. released Kimi K3, a 2.8 trillion-parameter model, three days earlier. Both launches push Chinese developers into the multi-trillion-parameter tier, a scale until recently associated mainly with the largest U.S. labs

Alibaba has made Qwen central to its effort to position itself as China’s default AI supplier. Over the past year the company has released a steady run of open-weight models while building out cloud infrastructure and custom chips to run them. The open-weight approach has helped Qwen assemble one of the largest developer followings of any Chinese model family, a base it now hopes to carry into the trillion-parameter class.

For enterprise buyers weighing the announcement, the comparison with Fable 5 is Alibaba’s alone. No independent leaderboard has scored Qwen3.8 yet, and the last Qwen model that was ranked, Qwen3.7-Max, sits well down LMArena’s list, where Fable 5 is No. 1. Benchmarks and the promised open weights would settle it, but Alibaba has released neither.

Image: Alibaba/X

Advanced Micro Devices Inc.’s first reinvention rebuilt the company. It was frankly about survival. Its next reinvention must redefine the company. AMD’s resurgence over the past decade came from doing what many thought was improbable – rebuilding its processor franchise, taking meaningful share from Intel Corp., and restoring credibility through disciplined execution.

In our view, AMD’s next chapter is fundamentally different.

The company is no longer trying to defeat Intel in a mature x86 market. Instead, it’s positioning itself as the indispensable second platform in a rapidly expanding artificial intelligence infrastructure market – one where Nvidia Corp. is likely to remain the dominant player for the foreseeable future.

That requires an entirely different playbook.

AMD must continue to build world-class silicon. But it needs to do more and bet its future on innovation, openness, heterogeneous computing, rack-scale systems, software, networking, strategic acquisitions and relentless execution.

Our assessment is AMD is not trying to repeat the beat Intel playbook. It doesn’t need to.

Thesis: AMD won its first turnaround by building a better central processing unit. It will win its next chapter — if it succeeds — by building a better AI platform built around EPYC CPUs, Instinct accelerators, ROCm software, rack-scale systems and an open ecosystem. We believe Chief Executive Lisa Su recognizes that the “Beat Intel” playbook won’t work against Nvidia. Instead, AMD is attempting to compress nearly two decades of ecosystem development into just a few years through disciplined capital allocation, strategic acquisitions and relentless execution.

In this Breaking Analysis, we’ll examine how AMD engineered one of the most impressive turnarounds in semiconductor history, why that formula worked against Intel, why it won’t work the same way against Nvidia, and whether Lisa Su’s strategy can transform AMD from an x86 comeback story into one of the defining AI infrastructure platforms of the next decade.

Core premise

AMD’s first reinvention was a comeback story. It rebuilt the company by taking share from Intel in a mature x86 CPU market. But this next reinvention is fundamentally different. The objective isn’t to beat Nvidia. In our view, that’s the wrong way to view the dynamics of the AI infrastructure business.

Nvidia has built perhaps the strongest AI infrastructure platform in the industry and we believe it’s likely to remain the dominant AI infrastructure supplier for the foreseeable future. Instead, AMD’s challenge is to define what success looks like in a market where Nvidia remains number one.

That’s a very different strategic problem. The graphic below describes in detail the thinking behind this premise.

Rather than trying to displace the incumbent, AMD is positioning itself to become the indispensable second platform for AI infrastructure – built around EPYC CPUs, Instinct accelerators, ROCm software, rack-scale systems such as Helios, and a strategy centered on openness, heterogeneous computing and disciplined execution.

To emphasize the key point of our analysis – The goal for AMD isn’t to be No. 1. The goal is to become an indispensable alternative in a market where supply is far outstripping demand for the foreseeable future — in the largest market in the history of tech.

To understand AMD’s next reinvention, it’s instructive to understand its first transformation.

AMD’s turnaround was the result a convergence of three strategic decisions, detailed below.

• First, the company finally let go of what had become a structural disadvantage. Founder Jerry Sanders famously said, “Real men have fabs.” Clinging to that integrated model nearly killed the company. But by spinning out GlobalFoundries in 2009, AMD embraced the fabless model, allowing it to focus on design while ultimately leveraging TSMC’s manufacturing leadership.

• Second, AMD rebuilt its technical foundation. Jim Keller, who by the way is now the CEO of Tenstorrent — which was reportedly in talks with Qualcomm to be acquired for $10 billion — returned to help architect Zen, with Mike Clark leading the CPU micro-architecture effort. Zen was more than just another processor. It completely reset AMD’s roadmap and restored the company’s engineering credibility.

• Third, Lisa Su transformed a sound strategy into disciplined execution. After becoming CEO in 2014, at the young age of 44, she consistently delivered on the roadmap, launching Ryzen into the PC market and EPYC into the data center. More importantly, she rebuilt credibility with customers, partners and Wall Street by doing what she said she would do.

The key point is AMD’s first reinvention wasn’t just about a CPU comeback. It rebuilt the company through architecture innovation, focus and execution – and laid the foundation for everything the company is trying to accomplish in AI today.

Why AMD’s attack on Intel was so successful

AMD’s first reinvention succeeded because it attacked a very specific problem. Intel’s manufacturing leadership began to falter just as AMD’s architectural execution improved dramatically.

The Zen micro-architecture reset the company’s CPU roadmap. The choice of chiplets allowed AMD to give its customers choice, economic advantage and engineering flexibility. This was a big deal as it lured customers away from a faltering Intel.

Then EPYC execution came with remarkable consistency, delivering a predictable cadence of new products into the data center from the start of Naples in 2017 to Milan and Genoa during the pandemic into Venice (not shown on the slide above), which will be highlighted at AMD’s Advancing AI event this coming week. It’s the latest in a long line of world-class processors.

Say:Do – Perhaps most importantly, Lisa Su established a reputation for execution. Quarter after quarter, generation after generation, AMD did what it said it was going to do. That restored confidence with customers, partners and investors.

The result was a remarkable comeback.

As the slide above shows, AMD has clawed back meaningful share in both PCs and the data center – in particular roughly 55% revenue share in the x86 data center market, while maintaining one of the industry’s most disciplined product roadmaps.

But as we’ll discuss momentarily, every one of these advantages played out within a relatively mature x86 market. Architecture. Execution. Process. Price-performance. These were the levers AMD pulled.

The next chapter will be different, because the market itself has fundamentally changed.

How Wright’s Law and the volume conundrum helped AMD beat Intel

Before we dig into the fundamental changes in the market, we want to review in more detail, the core challenges Intel faced, which AMD exploited. Intel’s troubles started to show up early last decade, well before most observers realized. The chart below shows x86 volumes peaking on the orange line (as PC volumes peaked) 365 million units (circa 2011). We saw a “dead cat bounce” during the pandemic, but the long slow decline of x86 continues today. It remains a multi-hundred-million-unit market; however, the blue line represents Arm unit volumes. Notice this is a double Y-axis chart and the Arm units are in billions while x86 is in the millions. So Arm wafer volumes are 10 times those of x86, which confers significant cost advantages to TSMC. So Intel was fighting a two-front war – with AMD eating share in the core x86 market, and Intel’s foundry at a significant cost disadvantage relative to TSMC.

The important point isn’t simply that PC volumes peaked around 2011. It’s that once x86 stopped being the volume engine of the semiconductor industry, the economics changed. Wright’s Law tells us that manufacturing costs decline by a constant as cumulative production doubles. As Arm became the volume architecture through smartphones and embedded devices – and fabless companies increasingly relied on external foundries – the center of gravity shifted. AMD’s first comeback occurred just as the old x86 economics were beginning to plateau. That is important to understand because you shouldn’t think of the AI era as another CPU cycle – it’s an entirely new volume curve.

Now in some ways, relative to Intel, AMD has different challenges around x86 but also faces similar headwinds. AMD doesn’t have the foundry commitment (and the drag on its P&L) that Intel has; but the market is shrinking for both companies. It will become increasingly difficult for AMD to gain share at a rate similar as it has in the past, especially as Intel gets its financial act together under CEO Lip-Bu Tan.

A parallel play AMD can run in our view is to be a bridge from x86 to the AI factory era. Because it has a strong position in x86 data center and is ahead of Intel in AI, it is in a position to effect that transition. Of course, a wildcard is the deal that Intel has with Nvidia as part of its $5 billion investment in Intel. Specifically, we’re referring to the integrated dual-chip architecture optimized for AI data centers.

The key point is, in the fullness of time, we predict that much of today’s software stack functionality, built around general-purpose x86 systems, will be re-architected around modern AI infrastructures. If and when this evolves, independent software vendors will be forced, for economic and functionality reasons, to port their software to AI systems – CUDA, ROCm and the like — and this is an opportunity for AMD to make money despite the x86 unit volume decline.

New rules for AI infrastructure

This graphic below is perhaps the most important in today’s Breaking Analysis.

AMD’s first turnaround worked because it attacked a mature x86 market where the competitive variables were well understood – process technology, CPU architecture, core counts, power efficiency and price-performance. But the rules have changed. The market AMD is entering today bears almost no resemblance to the market where it defeated Intel. The battleground is no longer the CPU. CPU plays an important role. But it’s a role in the larger AI factory.

With AI, winning in semis has shifted from designing the fastest and best price/performance processor to delivering the best integrated system. That means integrating silicon, software, networking and a developer ecosystem into a total system. Ultimately, this drives the economics of AI itself.

In other words, the basis of competition has shifted from individual components to complete platforms. That’s where Nvidia’s advantage becomes much more difficult to overcome. Nvidia’s lead isn’t just silicon. It’s software/CUDA, networking (Mellanox/NV-Link/Spectrum-X), rack-scale engineering, system integration and an ecosystem that has been growing for nearly two decades.

That’s why we believe AMD can’t simply repeat the Intel playbook.

The game itself has changed.

The CPU became the center of computing. The GPU became the center of AI. The AI factory is becoming the center of enterprise infrastructure. We’ve argued for some time that the AI factory becomes the physical foundation of what we call the System of Intelligence. That’s further up the stack than today’s discussion allows for, but that’s ultimately where the customer value resides. 

New rules, new playbook

So if AMD can’t repeat the Intel playbook, what exactly is the new model? In our view, it can be summarized in three layers as shown below:

• Build the core;
• Buy the gaps;
• Seed the ecosystem.

First, AMD is continuing to invest organically in what differentiates the company – EPYC CPUs, Instinct accelerators, ROCm software, chiplet innovation and its product roadmap. Think of these as the crown jewels.

Second, where time-to-market matters more than building everything internally, AMD has been remarkably disciplined with acquisitions. To wit:

• Xilinx brought adaptive computing via field-programmable gate arrays or FPGAs. Lisa Su at last year’s investor day called out $60 billion in acquisitions. Some $49 billion of that was Xilinx.
• Pensando added DPUs and networking expertise and is a critical part of the portfolio, which you’ll see at Advancing AI this coming week.
• ZT Systems gets AMD into rack-scale system integration ore quickly.

Rather than acquiring adjacent businesses, AMD has largely acquired bottlenecks – pieces that would have taken years to develop organically.

So you have EPYC – here comes Venice, Instinct, ROCm, Xilinx, Pensando, ZT Systems….

We think that’s exactly what’s happening here. AMD isn’t acquiring revenue. It’s acquiring where it has bottlenecks. And it’s investing where ecosystem flywheels can be created.

And that’s why there’s a big push by AMD into open standards.

Which brings us to the third layer – investments in the ecosystem. That means software, developer tools, open standards to facilitate partnerships with original equipment manufacturers, hyperscalers and Neoclouds; and encouraging broader adoption of ROCm and an open AI software stack.

Taken together, this is much more than a silicon chip roadmap. It’s a platform strategy. That’s an important distinction because we know already that AMD can build great CPUs and let’s agree they’ll be build excellent graphics processing units too. The new game and the real test is whether it can make all of these pieces behave like a single, deployable platform.

In many ways, Helios becomes that integration test. If AMD can integrate CPUs, GPUs, software, networking and rack-scale systems into a coherent platform, then it has a credible path to becoming the industry’s indispensable second AI platform.

If it can’t… then it risks remaining a supplier of excellent components in a market increasingly defined by complete systems.

If you think about what Lisa Su is doing, AMD is trying to compress 20 years of ecosystem development into five years of capital allocation. This is her strategy to move at the speed of Nvidia and not get left behind. In our view, the playbook of beating a wounded Intel has changed based on the actions Lisa Su is taking. It’s clear AMD is moving rapidly in a new direction.

How AMD stacks up to the Nvidia gold standard

At the risk of oversimplifying things, the slide below summarizes where we believe AMD stands today relative to Nvidia.

The first takeaway is the obvious: Nvidia leads the integrated AI platform today.

Its advantage extends well beyond GPUs. CUDA, developer mindshare, networking through Mellanox, NVLink and Spectrum-X, rack-scale systems,and nearly two decades of ecosystem development create a formidable moat that AMD will not erase without a major stumble from Nvidia. This we feel is unlikely.

But that doesn’t mean AMD can’t do well. Its strengths are different as is its value proposition.

AMD remains a leader in x86 server CPUs with EPYC. It has built a compelling portfolio through Instinct, Xilinx, Pensando and Helios. And perhaps most importantly, it offers customers something the market increasingly values – optionality.

Our assessment is that while Nvidia maintains clear leadership in software, networking, integrated systems and overall platform maturity, AMD’s opportunity lies elsewhere.

• First, becoming the industry’s most credible second platform;
• Second, leveraging its existing enterprise CPU relationships;
• Third, competing aggressively in inference, where cost, availability, power efficiency and workload economics may matter more than absolute peak performance;
• Finally, AMD’s commitment to open standards gives it a messaging angle that will resonate with customers in a market where many enterprises are becoming increasingly concerned about dependence on a single supplier.

So the bottom line isn’t that AMD is going to overtake Nvidia. It’s that AMD is steadily assembling the capabilities required to become the indispensable second platform for AI infrastructure – and in a market growing this quickly, that has been enough to create enormous value; and there’s potentially much more to come.

What does winning look like for AMD?

Everything we’ve discussed leads to this somewhat obvious conclusion. AMD does not need to beat Nvidia. In our view, that’s the wrong way to think about it.

The AI infrastructure market is expanding so rapidly that there is room for more than one successful platform. Winning doesn’t mean becoming the category leader. It means becoming the industry’s most trusted second platform.

The following points summarize why:

• Hyperscalers don’t like single-source dependence;
• Enterprise customers want negotiating leverage;
• So do OEMs like Dell, HPE and Supermicro – plus they want choice in their portfolios; and
• The rapidly emerging Neoclouds like Tensorwave are actively looking for differentiated infrastructure strategies.

At the same time, inference is becoming an increasingly important battleground – one where entry economics, availability, power efficiency and workload optimization can matter as much as peak benchmark performance. AMD’s commitment to openness, industry standards and heterogeneous computing also gives customers an alternative to highly integrated proprietary platforms.

None of these advantages displace Nvidia. But collectively, they create a very credible path to becoming the preferred second platform. And that’s why we believe AMD’s opportunity isn’t to become another Nvidia. It’s to become indispensable to customers who want choice, resilience, and flexibility in what is rapidly becoming the largest infrastructure market in computing history.

Comparing financials of the AI silicon players

What would a Breaking Analysis be without some numbers? Ultimately, the investment case comes down to the financials and the durability of business models.

The first point is one we’ve emphasized throughout this analysis. AMD doesn’t need to displace Nvidia to create significant shareholder value. Investors have rewarded AMD with a nearly $1 trillion valuation as shown below ($800 billion-plus).

If the company captures even a mid-single-digit share of the AI accelerator market (we have them above at 6%) while maintaining its CPU leadership, it can participate meaningfully in one of the fastest-growing infrastructure markets we’ve ever seen; and it’s valuation will have upside assuming execution and the bubble doesn’t burst in the near term.

Second, AMD remains the strongest merchant-silicon alternative. Broadcom is an exceptional company, but its AI business is primarily driven by custom ASICs built for hyperscalers. Broadcom’s customers build full systems with Broadcom proving critical IP. AMD is pursuing a broader merchant platform strategy.

Third – and this is obvious but important to emphasize – Nvidia is not a wounded Intel. Nvidia’s competitive position is fundamentally different. Its software ecosystem, networking leadership, systems integration, margins and cash generation make an Intel-style stumble far less likely.

Finally, valuation remains in focus.

Notably, AMD trades at valuation multiples that assume significant future AI success, while Nvidia’s extraordinary profitability and cash flows suggest it may be comparatively undervalued. Nvidia’s growth rate is much higher than any competitor. It’s margins are better, its free cash flow is far higher. The company has no debt. Yet its forward price-to-earnings ratio is about the same as the S&P 500 despite it growing at five times the collective growth rate of the companies in that index.

Our conclusion is pretty clear, however. AMD has created and can continue to create substantial value with a relatively small slice of a very large market. Investors should at the same time recognize that Nvidia’s leadership today is real – and we think sustainable – which is precisely why AMD’s strategy is centered on becoming the preferred second platform rather than attempting to disrupt the market leader the same way it did Intel.

The strategy is set – it’s now all about execution

Ultimately, the success of AMD’s strategy comes down to execution. The good news is that execution has become one of Lisa Su’s greatest strengths. Over the past decade, AMD has consistently delivered on ambitious roadmaps, regained credibility with customers and investors, and built one of the strongest engineering cultures in the semiconductor industry.

That gives us confidence that the company can continue closing the gaps relative to the leader – at least to the point where it will sell every AI system it can build. But investors should also recognize that the challenges are substantial. CUDA remains one of the strongest software moats in technology. Supply-chain constraints – from high-bandwidth memory to advanced packaging to energy to data center builders – will continue to govern the pace of the buildout. Especially as a big hope for AMD rests on its OpenAI deal to build out six gigawatts of capacity in four to five years.

ROCm is improving rapidly, but it still trails CUDA in ecosystem maturity and developer adoption. And Nvidia continues to extend its lead in networking, rack-scale systems, and integrated AI infrastructure.

Finally, timing as they say, is everything. AI infrastructure spending is growing at an extraordinary pace today, but technology cycles are never linear. Slower execution, changes in customer demand, or a normalization in AI investment could all affect AMD’s trajectory. All that said, overall, we believe AMD’s strengths far outweigh its risks and it is well positioned in an absolutely enormous market.

Success won’t look like AMD becoming another Nvidia. It will be measured by executing well enough to become the preferred alternative for AI infrastructure, delivering compelling economics, high quality inference, genuine customer choice and consistent execution.

Summary and action item for AI operators

Let’s close where we began.

AMD’s first reinvention was a comeback story. It rebuilt the company by beating Intel in a mature x86 market through better architecture, better execution, and disciplined leadership. Its next reinvention is fundamentally different. This isn’t a battle to replace or even beat Nvidia. It’s a race to become the indispensable second platform for AI infrastructure.

If Lisa Su and her team can successfully integrate EPYC, Instinct, ROCm, networking, rack-scale systems and the broader software ecosystem into a coherent platform, AMD doesn’t have to become the market leader to deliver outsized returns. It has to become the trusted alternative that every enterprise, hyperscaler, OEM and neocloud believes they should have in their AI strategy to close supply/demand gaps, keep Nvidia honest and fill seams in the market.

In a market expected to create trillions of dollars of new infrastructure spending over the coming decade, that may be one of the most valuable positions in technology.

So here’s our Breaking Analysis Action Item.

If you’re responsible for AI infrastructure strategy – whether you’re a data center operator, AI architect, platform engineering leader, or CIO – don’t wait until you need a second platform. Qualify one now. Benchmark AMD on real workloads. Validate the software stack. Understand where EPYC, Instinct and Helios fit your architecture.

Not because we believe Nvidia is going away or is under fire.

Quite the opposite.

Preserving strategic optionality today creates negotiating leverage, operational resilience and architectural flexibility tomorrow. In our opinion, that’s the real lesson from AMD’s next reinvention. The mandate isn’t to replace Nvidia everywhere. It’s to build strategic optionality by qualifying AMD where it creates leverage, resilience and economic advantage.

Image: theCUBE Research

Anthropic PBC is reportedly seeking to lease some of Meta Platforms Inc.’s data center capacity.

The New York Times today cited three sources as saying that the deal could be worth $10 billion over two years. However, the report noted that the companies’ discussions are at an early stage and could fall through.

The idea of a data center lease was reportedly floated by Anthropic in June. According to the Times, the company is seeking terms that would give it the option to cancel the contract early.

The artificial intelligence developer added a similar clause to its recently signed infrastructure deal with SpaceX Corp. Anthropic will pay $1.25 billion per month to use the rocket maker’s Colossus 1 and Colossus 2 supercomputers. The contract is structured as a 180-day lease, but both companies can end it early with a 90-day notice. 

Shortly after signing the SpaceX deal, Anthropic raised the rate limits of its application programming interface and Claude Code. A contract with Meta could be followed by a similar increase. However, any rate limit boost would likely be smaller given that lease is worth $416 million per month, or a third of what Anthropic is paying SpaceX.

Today’s report didn’t specify what Meta hardware the AI developer hopes to use. Some of the Facebook parent’s servers contain Nvidia Corp. chips while others use the MTIA 400, a custom accelerator that debuted in March. Anthropic is more likely to pick the former option. Its workloads are already compatible with Nvidia chips and adapting AI workloads to Meta’s silicon would involve a significant amount of work.

Leasing AI chips to other companies could help Meta recoup some of its heavy infrastructure spending. This week, the Facebook parent committed more than $50 billion to a data center campus in Louisiana. The sprawling development spans 3,650 acres and will be supported by 10 power plants.

Meta faces heavy competition in the AI infrastructure market. Buyers can choose among the offerings of not only the industry’s top cloud providers and SpaceX but also numerous well-funded data center startups. In theory, signing up a high-profile customer such as Anthropic could make it easier for Meta to stand  out.

The companies’ lease discussions are particularly notable because they compete with each other in the large language model market. Last week, Meta debuted an LLM called Muse Spark 1.1 that is optimized for programming tasks. The company plans to sell access to the model through an API that will cost 75% less than Claude.

Photo: Meta Platforms

Databricks Inc. is in the process of finalizing a funding round that will value it at $118 billion.

The company announced the deal on Thursday without disclosing the amount that it’s raising. According to the Wall Street Journal, the round will add $3 billion to Databricks’ balance sheet. Coatue is leading the investment with contributions from new and existing backers.

Databricks operates a cloud data platform that enterprises use to store and analyze business information. Over the past few years, the company has added numerous artificial intelligence features to speed up common analytics tasks. Databricks will use its new funding to further expand its AI capabilities.

The initiative will place particular emphasis on improving Genie, a suite of AI assistants that the company debuted in March. One of the tools in the lineup helps developers generate code such as SQL queries. Another makes it possible to create custom AI agents optimized for specific data science tasks.

Databricks debuted the newest addition to the suite, Genie One, earlier this year. It enables users to query records stored in Databricks and external platforms using natural language prompts. Genie One rolled out alongside a second new tool, Genie Ontology, that automatically organizes business records into a form that lends itself to analysis.

Databricks stated that its engineering push will also prioritize two other products: Lakebase and Unity AI Gateway.

Lakebase is a managed relational database that AI agents can use to store their information. Unity AI Gateway, in turn, is a governance tool for ensuring that AI agents meet quality requirements. It applies cybersecurity guardrails, blocks harmful prompt responses and performs related tasks. The tool also identifies opportunities to lower agent-related hardware costs.

“Enterprises are moving from tokenmaxxing to valuemaxxing,” said Databricks co-founder and Chief Executive Officer Ali Ghodsi, Co-founder and CEO of Databricks. “They don’t want to burn expensive tokens on the smartest model for every task – they want the best outcome per dollar. That means having the freedom to choose the right AI for the job.”

Databricks will use some of the proceeds from the new round to make acquisitions. Lakebase is the fruit of a $1 billion startup purchase  that the company inked last May. A few months later, Databricks expanded the platform’s capabilities by buying another startup called Mooncake Labs Inc. The deal bought the company technology that reduces the need to move data between applications and lowers the associated costs.

Databricks has already signed a term sheet for its latest funding round and expects to close it later this year.  

Image: Databricks

Google LLC said today it’s letting users link some of their favorite, most often-used applications with AI Mode, its conversational search tool powered by artificial intelligence, initially supporting Canva, Instacart and YouTube.

The update, announced in a brief blog post by Google, marks a significant expansion of AI Mode, which now does much more than just answering questions. With these app integrations, AI Mode now can complete some basic tasks for users. The company seems to be hoping that this will be enough to persuade more people to start using AI mode for things such as shopping and planning. It also positions Google to compete better with rivals like OpenAI Group PBC and Anthropic PBC, whose ChatGPT and Claude chatbots already integrate with third-party apps.

Google offered a number of examples of what users can now do with AI Mode. For instance, people planning a barbecue can ask AI Mode to create a list of groceries they need, then connect to their Instacart account and add all of those items to their shopping cart before paying on the Instacart app.

Alternatively, if they’re working on a project, such as designing a flyer, they can ask Canva for a selection of templates. In a third example, users could also ask AI Mode to curate a playlist based on their favorite artists and then save it on YouTube Music.

The update will come to users in the U.S. first of all, and Google said it will add support for more applications in the coming weeks and months.

Google already offers a similar capability with Gemini, which was launched earlier this year at Google I/O. It lets Gemini connect to apps including Canva, Instacart, OpenTable, Spark and others and complete tasks within the.

Though Gemini remains the focus of Google’s AI efforts, the company has been enhancing AI Mode continuously since launching it in early 2025. AI Mode is different from the standalone Gemini app. It’s an interface inside Google Search that relies on Gemini to deliver summaries and explanations based on user’s internet searches.

In a recent update, Google said AI Mode gained the ability to check whether an item is in stock in nearby stores. It also made it possible for users to explore the web side-by-side with AI Mode, so they can compare details on various websites and ask follow-up questions without losing the context of their initial search. Google also launched a “Personal Intelligence” feature in AI Mode, which enables it to search through people’s Gmail inboxes and Google Photos to generate more personalized responses to some searches.

Image: Google

Switzerland-based Mimic Robotics AG today announced it’s launching a new specialized robotic hand that looks and behaves like a human hand, introducing a new architecture for embodied robotics.

Physical AI, or embodied artificial intelligence, combines AI models and sensors that can perceive the world with robotic frameworks that can intelligently plan and execute tasks.

Mimic Robotics not only develops AI models that provide high-mobility robotic hands with the architecture to understand and interact with the world, but it also builds the hands themselves. As part of today’s announcement, the company released the Mimic hand M1.0 for industrial automation.

The M1.0 is designed and manufactured in Switzerland, in-house, and weighs about 4 pounds. It uses bidirectional pulley-guided tendons, somewhat similar to a human hand, allowing the hands to flex fingers and grip objects. To prevent crushing or bruising of fragile or soft objects, it includes tactile fingertip sensors that provide information about different force angles. The hands also come with a variety of gloves for different applications.

Mimic argues that the robotics domain, especially when it comes to embodied AI, suffers from two massive challenges: Robotics lacks the vast corpus of internet data that large language models use as a growth engine — sheer volumes of human-generated text — and the current robotic trend toward two-finger grippers is a mismatch for human hand dexterity.

These two problems collide when it comes to physical AI attempting to mimic, as it were, human poses, grips and fine activity. Although it’s possible to train AI systems based on video, giving them trajectories, a physical understanding of how the world looks and how human hands approach and grab objects. The two-finger design doesn’t operate like a hand; its grip is completely different from how humans grab and manipulate objects.

To solve this, Mimic designed a system that blends human video pretraining to provide robotic embodiment during the understanding phase, then uses physical data from wearables during mid-training to deliver the last mile.

Mimic’s M1.0 hands exhibit tremendous dexterity. In a video, the company showed one using a pair of tweezers to pick up a packaged integrated circuit and lay it on a printed circuit board, then carefully tap it into place. Another picked up a bolt between two fingers and passed it off to another hand. The hands have enough grip strength to lift and hold over 25 kilograms, about 55 pounds.

Including the company’s training pipeline, the hands can be trained to perform highly fine work, as mentioned above. The company also demoed the hands moving fingers independently and forming hand signals – such as the peace sign. Although they are designed for industrial applications that require fine movements, heavy lift and robustness, this means that they could be operated gently for care situations or even potentially used to display sign language.

For industrial purposes, Mimic’s AI models and hands can provide powerful capabilities for assembly, packaging and sorting, and numerous other complex manual tasks. Coupled with cameras and touch sensors, they can operate in unstructured environments, such as opening and closing boxes, placing items, handling irregular objects, moving and placing wiring, and performing fine control tasks.

Images: Mimic Robotics

Monumental BV, a Dutch startup that operates a fleet of bricklaying robots, has raised $35 million in funding. 

Khosla Ventures led the Series B investment. Monumental stated in its funding announcement today that existing backers Plural and Hummingbird chipped in as well.

Monumental uses robots to construct the walls of homes, schools and other buildings. It uses two robots called Petra and Panama to deliver bricks and mortar, respectively, to the section of a construction site where they’re needed. A third system called Pisa carries out the bricklaying work.

“Every robot we deploy expands the industry’s capacity to build, bringing a future of beautiful, affordable, bespoke buildings and infrastructure closer to reality,” said Monumental co-founder and Chief Executive Officer Salar al Khafaji (pictured, left, with co-founder Sebastiaan Visser).

The company coordinates its robots using a platform called Atrium. According to Monumental, the software also automates several of the other tasks involved in construction projects.

Atrium automatically generates blueprints based on high-level details inputted by an architect. Users can customize parameters such as the manner in which bricks should be arranged. Most construction projects use bricks that lie on their longer side. Atrium can also place bricks vertically, an arrangement that is mainly used for decorative purposes.

The first version of a building facade plan often contains inaccuracies that can complicate construction. Monumental corrects errors using a method called photogrammetry. The company takes photos of a construction site, assembles them into a three-dimensional virtual replica and compares the replica to the blueprints. Monumental says that it can catch millimeter-scale inconsistencies.

Once a blueprint is ready, Atrium translates it into a bricklaying plan for the company’s robots. The software then uses sensor data from the robots to monitor for mistakes. Atrium can coordinate multiple robot teams on the same site, which makes it possible to speed up construction by parallelizing work.

Monumental developed a custom programming language for Atrium. The syntax, which is known as Plan, enables the company’s developers to optimize how the company’s robots should go about their work. Software teams only have to provide a relatively high-level description of the project workflow. The more granular details are defined automatically by a component called an interpreter.

Monumental had to solve multiple technical issues to make its robots work reliably. 

Robot sensors sometimes malfunction, which means that the data they generate doesn’t reflect the parameters of a construction site. Monumental avoids such issues by calibrating its sensors before each project. The calibration is carried out at a dedicated section of the company’s headquarters that is equipped with motion cameras.

Monumental’s robots pick up the bricks they use one by one. If a robot picks up a brick a few millimeters off center, it may incorrectly place the brick in the building under construction. The company addressed the challenge by training a custom neural network that automatically offsets such errors. 

Monumental uses its robots to deliver a bricklaying service with outcome-based pricing. According to the company, its billing approach removes the need for construction firms to buy pricey machinery. The company has so far helped customers construct over 100 homes, schools and other buildings.

Monumental will use its new funding to hire more engineers. Additionally, the company plans to grow its international presence and the number of construction tasks that its robots can automate.

Photo: Monumental

TerraFirma Inc., a startup with a platform that enables construction teams to remotely operate heavy machinery, has raised $115 million in funding.

The bulk of the capital arrived in the form of a $100 million Series A round led by Kleiner Perkins. TerraFirma disclosed today that Bain Capital, Definition and more than a half-dozen other institutional investors chipped in as well. The company’s investor roster also includes employees at several major tech firms including SpaceX Corp.

TerraFirma was launched in 2024 by former SpaceX engineers Noah Schochet and Noah McGuinness. Its platform enables construction teams to remotely pilot excavators, tractor loaders and other vehicles using joysticks. TerraFirma says that removing the need for users to sit in a construction vehicle’s cab boosts safety while significantly improving productivity.

Customers don’t have to manually coordinate every construction task. According to TerraFirma, its software provides the ability to create “autonomous workflows” that automate repetitive chores. The platform also promises to ease several related tasks.

Creating a price estimate for a construction project is a highly technical task. Engineers have to account for factors such as a construction site’s elevation, the type of building that it will host and the amount of soil that must be added or removed. The task can take weeks in some cases.

TerraFirma says its software speeds up the process. Furthermore, engineering firms can use it to plan each step of the construction workflow. They can then test the workflow for inefficiencies with the help of a built-in simulation tool.

Once construction is underway, TerraFirma’s platform uses an obstacle avoidance algorithm to keep remote-controlled vehicles from colliding. The company says that the software makes it possible to operate multiple vehicles on the same construction site.

Autonomy features such as automated collision avoidance require real-time data from vehicle sensors. TerraFirma’s website doesn’t specify how it collects sensor data. However, the company does note that it retrofits existing machinery to work with its software. That suggests TerraFirma equips customers’ vehicles with a sensor module.

Atoms Inc., a newly launched robotics startup led by Travis Kalanick, equips trucks with autonomy features by adding in a ruggedized computer. The system combines a GPS module, cameras and radar sensors in a case designed to withstand severe weather.

“Making construction truly faster and cheaper requires innovating on operations and technology together across the full stack,” said McGuinness. “We believe autonomy is a part of the solution, but driving real change requires building a whole ecosystem of technology that is directly informed by rapid iteration and lessons from the field.”

CNBC reported that TerraFirma will use its new funding to hire 300 workers over the next year. Additionally, the company will build a new facility from which it plans to orchestrate customer vehicles. 

Photo: Unsplash

The European Union today moved closer to a ban on social media for children across its 27 member states in what will become the most meaningful effort so far to protect young people from the dangers of spending too much time online.

Ursula von der Leyen, the president of the bloc’s executive arm, the European Commission, cited a recent report by the child psychiatrist, Professor Jörg Fegert, and epidemiologist, Dr. Maria Melchior, which pointed to harmful features of social media such as the infinite scroll, autoplay, recommendation algorithms and persistent notifications.

The study revealed that across Europe, the average child spends four to six hours on social media daily, with about 60% of those children experiencing what the authors wrote were “socio-emotional development and susceptibility to mental health issues.” This has resulted in widespread sleep and concentration problems, and increased rates of depression and anxiety.

The authors recommend that the EU block social media for children under 13 unless they are under the supervision of a parent or teacher. For adolescents, children aged 13 to 18, the authors recommend that access should only be granted if the platforms have built-in safety mechanisms that limit the child on features such as infinite scroll. They also recommended that social media be blocked for all toddlers.

“Childhood is a period of extraordinary and delicate brain development,” von der Leyen said in a statement. “During this stage, our children need time in the real world. Time to play, to build friendships face-to-face, to make mistakes. Time to shape their own identities, their own personalities, before an algorithm shapes them instead. I believe we need to give our children this time.”

The move comes after a slew of countries have already blocked social media for children under 16, with Australia being the first and a number of other countries, including the U.K., following. More than 20 countries in total now have either blocked social media for young people or are in the process of developing the legislation that will profoundly affect how young people use various platforms.

Critics have argued that this is government overreach or that kids will find ways around the blocks, which seems to have happened in Australia. Those against blanket bans have argue that the onus should fall on parents to teach children how to use social media responsibly while children should be empowered to make better decisions when online.

Photo: Unsplash

Cloudflare Inc. today opened general availability for Precursor, a bot detection system that tracks how a visitor behaves across an entire browsing session instead of testing them once on arrival.

Precursor runs inside the browser. It streams interaction signals back to Cloudflare’s edge, where servers score them in real time for evidence of automation. The target is the CAPTCHA. A challenge page tests the visitor once, at the door. Everything the visitor does after that is assumed good.

Cloudflare puts bot traffic at roughly 57% of web requests. By its count, automation now outweighs people on the internet. Cloudflare’s argument is that a point-in-time check is easy to fake. A bot can fake a single action. Faking an entire session, with the timing irregularities of a real person, costs real engineering effort.

“Instead of just checking an ID at the gate, we are looking at behavior over the entire visit,” said Chief Technology Officer Dane Knecht.

Customers turn Precursor on with a single click and no code changes. Cloudflare injects a small script into pages already passing through its network and the script logs mouse movement, scrolling rhythm, typing cadence, clipboard activity and how long a page stays visible in the browser tab.

What happens next is a coherence check. Cloudflare’s analysis engine unpacks the telemetry and looks for internal contradictions, such as pointer activity recorded while the page was hidden or typing events fired at a moment when no text field held focus. Suspicious sessions accumulate context rather than resetting, feeding a running Bot Score that follows the visitor through a site or single-page application.

That closes off a standard evasion. Under per-request challenges, an automated agent can wipe its behavioral signature by reloading the page. Precursor keeps scoring.

Cloudflare said the script records aggregate patterns and not the inputs themselves. Keyboard activity is stored as timing rhythm and cadence. The characters typed are never captured, according to the company.

The launch extends a long run of bot and crawler products out of Cloudflare. The company began blocking artificial intelligence scrapers by default for new customers last year, built the Pay Per Crawl marketplace so publishers can charge AI firms for access and shipped AI Crawl Control for per-crawler allow and block decisions. Bot management is also a wholesale business. WP Engine Inc. built the bot controls in its Global Edge Security service on top of Cloudflare’s network.

Knecht said Cloudflare already protects users billions of times a day at login and checkout and described the stretch between those moments as “a black box” the company is now filling in.

Precursor is generally available now. Cloudflare did not disclose pricing.

Image: SiliconANGLE/Ideogram

After eight years as the chief executive of VMware Inc. and nearly four more leading Intel Corp., Pat Gelsinger suddenly found himself retired. Then the phone rang.

“Less than two seconds after the Intel departure was announced, Scott called,” Gelsinger (pictured) recalled. “Whether it was opportunistic on his part or God ordained on his part we may not fully know.”

The caller was Scott Beck, co-founder and CEO of the faith-based technology company Gloo Inc. Gelsinger had been an investor in Gloo over the years and, as he told SiliconANGLE in an exclusive interview, he wanted time to reflect on his next move following his departure from Intel near the close of 2024. In the end, Gelsinger decided to become a general partner of the venture capital firm Playground Global, and executive chair/head of technology for Gloo.

In his leadership role at Gloo, Gelsinger oversees a portfolio of brands and services that provide text and email messaging services, data-driven metrics to manage volunteers and improve donor engagement, and artificial intelligence applications to automate administrative tasks. There is an AI Studio where developers can create new tools and assistants while integrating top large language models. Churches, faith-based universities and nonprofits are Gloo’s primary customers.

The company, which went public last November, reported 238% growth in revenue year-over-year for the most recent quarter, and raised its revenue guidance to $195 million in 2026. On Monday, the firm announced a proposed public offering of 7 million shares of its Class A common stock be be used for further acquisitions and investments.

When asked what had driven Gloo’s value proposition in appealing to the faith-based world, Gelsinger described the issue facing many religious institutions today.

“Every time you put a dollar in that offering plate, or send your check in, or do a transaction to support, it’s a bad dollar because you don’t have good technology behind it,” Gelsinger said. “The mission of Gloo is enabling those who serve, powering those who serve, shaping technology for good.”

Serving a diverse ecosystem

To turn that dollar into a good one, Gloo has built a platform that helps missional organizations amplify impact by powering their technology. The company offers a values-aligned AI portfolio that modernizes systems, workflows and data, with marketing and donor solutions to expand reach and long-term giving.

“The ecosystem we serve is diverse,” Chief Product Officer Benjamin Gauthier said in an interview with SiliconANGLE. “But every organization that we work with is going after impact.”

The company’s use of AI has focused on ensuring that biblical quotations are accessible and accurate. Gloo found that as new versions of AI models were deployed, the accuracy of biblical citations could vary significantly.

“The major models are 60% to 70% accurate,” Gelsinger said. “We’ve developed the technology to take the major models and add accurate Bible quotations on top of it. We guarantee that we do 100% accurate biblical quotation.”

In addition to accuracy, Gloo is also seeking to help organizations leverage AI models in keeping with faith-based values. This can occasionally come into conflict with major providers such as OpenAI Group PBC, which announced and then paused plans to develop an “erotic” mode for ChatGPT.

“The power that’s being released for these mission-oriented organizations is really quite stunning,” Gelsinger noted. “On the other side, OpenAI is announcing erotica. Not OK. In the middle of that, how do you harness that energy and make it good for this community? If you were to come to our AI Studio, pick whatever model you like. But we are going to put guardrails and systems around it that make it good.”

Strengthening parishioner communication

Along with leveraging the power of artificial intelligence, Gloo’s clients are also using the platform for basic functions such as communication and outreach. At St. Mark Missionary Church in Mishawaka, Indiana, Discipleship Pastor Johnny Bennett has found that having a resource for key administrative tasks allows him to focus more on his role as a spiritual leader.

“My role as a pastor is to be with people,” Bennett said in an interview. “There are so many demands of this position that pull you away from that. Gloo can take care of those nitty-gritty logistics. This is a direction that the church is going in.”

St. Mark has expanded by 400 members over the past two years, according to Bennett. That has created a need for improved communication, especially when visitors attend a service for the first time. Bennett’s use of Gloo’s text and email tools has helped facilitate outreach and grow the church’s congregation.

“For as long as I’ve been around churches, they’ve been trying to figure out how to welcome new people well,” Bennett explained. “[Gloo’s] workflows allow us as a ministry staff to be significantly more direct with our communication. Gloo gives us any easy way to follow up with new visitors.”

One way that St. Mark stays in touch with its regular parishioners is through the “517 Club,” a reference to the apostolic command of “pray without ceasing” in Thessalonians 5:17. Through workflows driven by Gloo’s platform, St. Mark sends out a text reminder every Monday at 5:17 p.m. to pray.

The church’s embrace of modern communications tools is an example of the opportunity that Gloo intends to leverage as it grows its business. The company is offering a way to reach churchgoers that reflects today’s technology, something that has not always been the case among faith-based organizations, according to Gelsinger.

“It’s like when I was running VMware,” Gelsinger said. “Careful which tool you pick today because it’s tomorrow’s legacy. This ecosystem has got 30 to 40 years of legacy.”

Expanding through acquisitions

Gelsinger’s current role at Gloo has enabled him to draw from many years of experience in the technology world. He found that his work at VMware, which transformed from a server-virtualization provider into a hybrid cloud powerhouse during his eight years at the helm, has proven to be especially useful as he seeks to scale up his latest venture.

“I’m a hardware, hard tech guy at heart,” Gelsinger said. “I’ve often asked myself, ‘Why did God have me run a software company for eight years?’ The answer is now extraordinarily obvious, it’s called Gloo. Every day the experience of running an at-scale software company has come to benefit. There are things that I can now be very confident and rapid in putting in place because I ran VMware.”

One of the things that Gelsinger has moved swiftly to do is acquire other companies. Gloo purchased Masterworks (marketing and fundraising), Igniter (creative media) and XRI Global (voice and multilingual AI) in 2025, and has added Midwestern Interactive (application building), EnterpriseMarketdesk (AI-enabled services) and Westfall Group (donor engagement) so far this year.

Asked about Gloo’s acquisition strategy, Gelsinger noted that the faith-based technology market has not been as active as others in the tech industry.

“There’s not a lot of people out there acquiring these assets in this industry, so we can be judicious,” Gelsinger said. “We can be efficient acquirers. We haven’t lost the founders, they’re committed to this journey as part of the Gloo family.”

As Gloo brings smaller companies into the fold, inherited customers often are unaware that there is now a parent firm in the mix. Gelsinger cited his company’s decision to make a strategic investment in Barna Group, a Texas-based research firm focused on how spiritual and cultural trends affect Christian organizations, as an example. He noted that some loyal customers of Barna were not aware of Gloo’s investment, but the platform has been transformed nevertheless.

“Barna was almost 100% analog,” Gelsinger said. “Today, they are about 90% digital and AI-fueled. We upgraded their technology and accelerated their growth rate.”

Striving for profitability

The challenge ahead for Gelsinger is to make the company profitable. During the company’s earnings call in June, executives said Gloo ended the quarter with $33 million in cash with an expectation to approach breakeven in the third quarter and turn profitable in the fourth quarter.

“We’re building this scalable engine and solidifying the different pieces here, so I feel very good about that,” Gelsinger told SiliconANGLE. “Now the next phase is to get it to profitable, scalable growth. There’s a lot of emphasis on getting to profitability in the second half of the year.”

According to Gelsinger, more than a trillion dollars of the U.S. economy is driven by the faith and flourishing ecosystem, with hundreds of billions of dollars funneled through church donations every year. It is a big, fragmented, underserved market and Gloo’s technology platform is geared towards bringing it into the modern AI age.

He noted that Gloo now claims 12 of the 14 organizations in the U.S. that do Bible translation as clients, and a majority of campus ministry organizations are on the company’s platform.

Yet beyond the numbers and market potential is a larger calling for Intel’s former CEO. Gelsinger says he is driven by what he terms “the community,” his lifetime spent inside the faith ecosystem where technology supports a higher purpose.

“You are talking to people who have invested decades for a mission,” Gelsinger said. “These are like the best human beings on earth. If you make them better, you just sleep so good at night.”

Photo: Gloo

On the four dimensions of real sovereignty, the fifth dimension every chief financial officer is learning about the hard way, and why open source isn’t a preference — it’s the architecture.

Something is wrong with the way the industry is talking about sovereign AI. Not slightly wrong. Not grammatically wrong. Structurally wrong.

The term has landed in board decks, vendor marketing and government procurement documents — and in almost every case, it means something far narrower than what’s at stake. Sovereign AI has become a synonym for data residency. For picking the right Amazon Web Services region. For a geographic configuration that provides legal comfort (think GDPR) without addressing any of the underlying dynamics that create the actual exposure.

This is not a semantic complaint — it’s a strategic one, and it’s baked in fallacies. Organizations that define sovereignty incorrectly are building on a false foundation. By the time it becomes obvious, the window to fix it has already closed: vendor lock-in, compliance penalties, P&L nightmares and — pick your poison.

So let’s define it properly. Because the definition is the strategy (think AA for AI — the first step is admitting you have a dependency).

Why this is happening now — and everywhere at once

The forces pushing sovereign AI from concept to operational imperative are converging at a speed most enterprise planning cycles weren’t designed to absorb.

Start with Stargate — a $500 billion U.S. commitment to own the global AI supply chain, announced January 2025 by OpenAI, SoftBank, Oracle and MGX, with $100 billion to deploy immediately. ^[1] The message was unmistakable: AI compute is strategic national infrastructure, treated with the same intentionality as energy grids and satellite networks. Then DeepSeek demonstrated that frontier AI could be trained on chips U.S. export controls were specifically designed to restrict. The theory that raw compute dominance wins the AI race dissolved in a matter of weeks.

Europe responded structurally, not rhetorically. The Summit on European Digital Sovereignty in Berlin on Nov. 18, 2025 — Macron and Merz co-chairing — convened more than 900 policymakers, industry leaders and member state representatives, and produced the Declaration for European Digital Sovereignty alongside more than €12 billion in announced investment and partnership commitments. ^[2] That declaration was then formally presented to the EU Telecom Council on Dec. 5, 2025. ^[3] Mistral raised $830 million in institutional debt from a seven-bank consortium — BNP Paribas, Crédit Agricole CIB, HSBC, MUFG and others; no U.S. bank participation — to fund a sovereign graphics processing unit data center in Bruyères-le-Châtel, outside Paris, with 13,800 Nvidia GB300 GPUs and 44 megawatts of capacity. The company’s stated objective: 200 MW of sovereign capacity across Europe by end of 2027. ^[4] These are not policy positions. These are buildings being built — with reportedly much larger IPCEI on AI and IPCEI on Compute Infrastructure Continuum programs now entering matchmaking, set to commence by early 2027 across 17-plus member states. ^[5]

The Gulf is running at the fastest pace. Saudi Arabia’s HUMAIN — launched May 2025 under the Public Investment Fund and now with a minority stake from Aramco — is building the full AI stack: compute, models, cloud, applications. Eleven data centers under construction at 200 MW each; a $10 billion Advanced Micro Devices partnership for 500 MW of compute; agreements with xAI and AirTrunk. This isn’t a data center play — it’s a national intelligence infrastructure play, with a stated ambition to become the world’s third-largest AI provider behind the U.S. and China. ^[6] The UAE’s G42 and TII’s open-weight Falcon models — including Falcon Arabic and Falcon-H1 — reflect the same logic at the model layer. ^[7] Malaysia has established its National AI Office. India launched a sovereign large language model initiative. Canada committed an AI Compute Access Fund. None of these is a coincidence.

The numbers confirm what the politics are signaling. The CNAS Sovereign AI Index tracks 130-plus national sovereign AI initiatives, with more than 80% of total disclosed investment concentrated in the Middle East and East Asia. Infrastructure projects account for 59% of all initiatives tracked; Nvidia supplies the GPUs in 52% of them. ^[8] More than 60 nations have published formal AI strategies. More than 30 have committed specific domestic funding. This is not a wave — it is a continental shift.

Gartner’s November 2025 survey of 241 Western European chief information officers found 61% will shift toward local cloud providers specifically because of geopolitics — not performance, not cost. ^[10] Accenture’s 2025 survey of 1,928 organizations across 28 countries found 62% of European organizations actively seeking sovereign solutions in direct response to geopolitical uncertainty, with Danish (80%), Irish (72%) and German (72%) firms leading. ^[11]

What sovereignty actually means — and what it doesn’t

Here’s the thing about data residency: it’s not wrong. It’s just one-fourth of the answer — and most organizations have mistaken it for the whole thing.

Sovereign AI is about your data, your infrastructure, your stack, your rules.

Genuine sovereignty has four core dimensions (plus a fifth bonus pillar — financial — which we’ll get to). All four. Failure in any one breaks the chain.

1.  Territorial: Where data and compute physically reside

Yes, it matters. Regulatory compliance, latency, baseline legal protection. All real. But this is the beginning of the analysis, not the conclusion. Organizations that stop here have answered 25% of the question and are filing it under “done.”

2.  Operational: Who actually manages and secures the environment

This is where most “sovereign” deployments quietly break down. Sovereign servers operated by a foreign-headquartered managed service provider, under foreign employment law, governed by foreign corporate policy — are not sovereign. (Think Huawei managing infrastructure for Telefónica.) ^[12] The questions that expose this are simple and uncomfortable: Who holds the encryption keys? Who gets paged at 3 a.m. during a breach? Which jurisdiction’s law enforcement can walk in and demand the audit logs? If the answers point outside your organization’s control boundary, the territorial dimension was just window dressing.

3.  Technological: Who owns the underlying stack and IP

This is the one that hits hardest in AI specifically. A sovereign deployment built entirely on proprietary model application programming interfaces, proprietary orchestration frameworks and proprietary policy engines hasn’t eliminated vendor dependency — it has relocated it from the cloud infrastructure layer to the software layer, which is harder and more expensive to migrate. If you can’t fork the orchestration layer, audit the policy engine or modify the runtime without going through the vendor, you don’t own the system. You’re licensing it — categorically different things.

4.  Legal: Which jurisdiction governs access

Jurisdiction follows the company, not the data center. A U.S.-headquartered entity’s foreign subsidiary is subject to US law. A European company running US-headquartered software is subject to U.S. legal process through that vendor relationship. The CLOUD Act is the most visible expression; Mutual Legal Assistance Treaties and national security orders create parallel pathways. Real sovereignty requires the legal framework governing access to your data, models and AI systems to be deliberately chosen and architecturally enforced — not assumed based on a map.

McKinsey’s December 2025 survey of 300 executives, investors, and government officials found 71% characterize sovereign AI as an “existential concern” or “strategic imperative.” The same research found most organizations lack a detailed strategy, action plan, or budget to execute on it. ^[13] The intent is there. The architecture is not.

Sovereign cloud is a geography question. Sovereign AI is broader: how intelligence is created, trained, governed and deployed — across infrastructure, models, applications and organizational control surfaces. It requires a different posture entirely.

The fifth dimension: Financial sovereignty

Bonus pillar: Sovereignty from vendor lock-in

There is a fifth dimension of sovereignty that doesn’t appear in policy white papers — but that enterprise technology leaders are learning about this year, in real time, on live budgets. Call it financial sovereignty: the ability to own, predict and control what your AI actually costs, without being subject to unilateral vendor pricing changes, usage-based billing surprises or the forced obsolescence of the models your entire workflow depends on.

This is not theoretical. This is the defining enterprise AI story of the first quarter of 2026 — and three companies are now the cautionary case study every CFO is being shown.

Uber: The budget bonfire

Uber gave Claude Code to roughly 5,000 engineers and encouraged adoption aggressively — going as far as ranking staff on internal leaderboards based on usage. Seemed like a great idea. Adoption ran from 32% of engineers in February to 84% “agentic users” by March. Average monthly spend per engineer landed in the $150 to $250 range; heavy users hit $2,000. By April 2026, the company had burned through its entire 2026 AI budget in four months. Chief Technology Officer Praveen Neppalli Naga told The Information the company was “back to the drawing board” on AI budgeting; the chief operating officer began questioning return on investment. ^[14] A productivity win that produced invoices competing directly with headcount budgets — and nobody had modeled for it.

Microsoft: Beloved by users, killed by the CFO

Microsoft ran the same experiment in parallel. It rolled out Claude Code in December 2025 to its Experiences and Devices division — the group behind Windows, Microsoft 365, Outlook, Teams and Surface. Engineers loved it. They preferred it over GitHub Copilot by a wide margin. On May 14, 2026, Microsoft began canceling those licenses, directing thousands of engineers back to Copilot CLI by June 30 (conveniently, the last day of Microsoft’s fiscal year). The tool wasn’t canceled because it didn’t work. It was canceled because it worked so well that token-based billing consumed the annual AI budget in months ^[15] — a tool beloved by the people using it, cut by the people paying for it.

The pricing model itself: Altman’s vision is your invoice

Anthropic, following the industry’s broader direction, moved away from flat fees toward usage-based token pricing for agentic workloads. As agentic AI runs multistep tasks autonomously — in the background, without a human triggering each call — token consumption becomes nonlinear, continuous and nearly impossible to forecast with traditional enterprise budgeting methods. At BlackRock’s 2026 Infrastructure Summit, Sam Altman described the direction as follows: “We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter.” ^[16] Your CFO heard that too. There’s lots of talk out there about outcome-based pricing. This ideally aligns pricing with customer objectives where pay for value (versus token consumption) becomes the preferred model. We’re seeing this today with the way some firms price forward deployed engineers based on milestones. Think of this as more “services-as-software” where clients pay for value realized.

Google: The deprecation treadmill

Google illustrates the third risk: forced migration. Enterprises that built production workflows on Gemini API model versions have faced a rolling deprecation cycle — Gemini Code Assist legacy tools removed Oct. 14, 2025; the original Gemini Python SDK reached end-of-life Nov. 30, 2025; multiple preview models deprecated on short notice throughout early 2026. ^[17] Each cycle pulled engineering teams off product work and into emergency rewrites. The vendor’s roadmap — not the enterprise’s own priorities — was driving the engineering calendar.

The pattern across all three is identical. Organizations that did not own their AI stack became price-takers, timing-takers and migration-takers. Their AI strategy was shaped not by their own decisions but by the commercial and engineering priorities of vendors they had no leverage over.

Sovereign AI resolves this. Self-hosted open-weight models make inference a compute cost — predictable, owned, optimizable. When you control the orchestration layer, you choose when to upgrade and on whose schedule. When your runtime is open source, a model deprecation is a parameter change, not a project. When you own the stack, your AI program budget is yours to manage.

Cost sovereignty is not about being cheap. It’s about being in control — which, as these three companies just demonstrated, is worth a great deal more than the token savings.

Open source is not optional

Stated plainly: You cannot build genuine sovereign AI on a closed proprietary stack. It is definitionally impossible.

Once you accept the five dimensions above, the conclusion is logical. Technological sovereignty requires the ability to audit, fork and operate the system independently. A proprietary stack doesn’t allow this. The orchestration layer is a black box. The policy engine is vendor-controlled. The inference API can be modified, restricted or repriced without your input. You cannot fork what you cannot read.

And the CLOUD Act problem applies to software IP just as much as it applies to data. A U.S.-headquartered software vendor is subject to U.S. legal process — which means the software they control, the keys they hold, the telemetry they collect and the modifications they can be compelled to make are all within reach of U.S. law enforcement. No hosting geography resolves this. Proprietary AI platforms create legal, operational, strategic and financial exposure that a well-chosen data center cannot fix.

The only credible path runs through open source — not as a developer-culture preference, but as an architectural requirement. An open-source-first architecture means the orchestration engine can be audited, forked and self-hosted. Model weights can be inspected and fine-tuned locally. The policy engine is code you own, version-controlled, readable. Connectors, pipelines, inference runtimes — owned, not licensed.

And open source enables genuine portability: the ability to move workloads between providers, across jurisdictions, across hardware generations — without being captive to a vendor’s release cycle or pricing changes. The EU Data Act — which prohibits cloud switching charges and data egress fees from Jan. 12, 2027 — is the policy layer of the same structural shift. ^[18] Open-source-first architectures are what make an organization genuinely fluid in that market. Without them, you’ve traded cloud lock-in for software lock-in — a different constraint, not the absence of one.

The next wave of AI will be decided here

The first wave was about capability: Which models could do what, who could ship fastest.

The second wave — the one we’re in now — is about control: Who owns the infrastructure that produces intelligence, who governs access, who controls the costs, who sets the rules.

This wave won’t be won by whoever has the best models. Those are commoditizing faster than anyone predicted. It will be won by whoever owns the sovereign layer that governs how models are deployed, governed and integrated into the systems of intelligence that run the world and will define the AI software stack.

The money confirms it. The global sovereign cloud market is projected at $195 billion in 2026 (Fortune Business Insights), growing to $1.13 trillion by 2034. ^[19] McKinsey projects the sovereign AI market at $500 billion to $600 billion by 2030, representing 30% to 40% of all AI spending. ^[20] These aren’t venture bets. They’re infrastructure commitments — the same category as telecommunications networks and energy grids in prior generations.

And the timeline is real. McKinsey found sovereign AI migrations take three to four years — not because the technology isn’t ready, but because organizations need time to decide where sovereignty actually matters and restructure their operating models accordingly. That clock starts from the architecture decision — not from the compliance deadline, not from the crisis. From the decision.

The organizations that start building for sovereignty now will have genuine AI capability — and genuine AI independence — when the regulatory, geopolitical and competitive pressures of the next three years arrive simultaneously. The ones that wait will discover the migration window has already closed, and they’re operating on someone else’s infrastructure, someone else’s pricing terms, and someone else’s deprecation schedule.

The sovereignty wars have begun. The territory is forming. The next wave of AI will be decided here.

What this series is going to do

The world is fragmenting into sovereign AI islands — regionalized, regulated, increasingly disconnected deployment environments, each with its own rules, requirements and governance expectations. Every global enterprise faces the same challenge: building AI capability that’s genuinely sovereign within each boundary while staying coherent across the whole.

We’re going to map that territory. Next in the series: the six-layer architecture of sovereign AI readiness — what it actually takes to build a production-grade sovereign AI system from infrastructure foundation to governance apex. Then: a vendor market map (who passes the sovereignty test and who is applying the label to a software-as-a-service product that doesn’t), deep-dives on sovereign AI infrastructure programs, and profiles of the regulated verticals where sovereign AI is a compliance clock, not a roadmap item.

We will interview the architects, CTOs, government program leads, and enterprise technology officers doing the actual building — not the press releases.

We will do awards purpose-built for sovereign AI — recognizing the organizations and people doing the hard, unglamorous work of production deployment.

And we will hold the line on rigor. Sovereign AI washing is already loud. Part of our job is giving the industry the vocabulary and the criteria to tell the difference.

The sovereignty wars have begun. The territory is forming. This is where we start mapping it.

About the lead authors

Next in series: The Architecture of Sovereign AI Readiness: the six-layer stack that separates genuine sovereignty from sovereignty theater — layer by layer, from infrastructure foundation to governance apex.

References and sources

All claims, figures and quotations referenced in this paper are linked below. Where the underlying primary source differs in scope from the in-text claim, both are cited.

[1]  OpenAI, “Announcing The Stargate Project,” January 21, 2025 — $500B over four years; $100B immediate; OpenAI, SoftBank, Oracle, MGX. openai.com/index/announcing-the-stargate-project

[2]  Élysée, “Summit on European Digital Sovereignty in Berlin,” November 18, 2025 — Macron/Merz co-chaired; >900 participants; >€12B in commitments; Declaration signed. elysee.fr — Summit on European Digital Sovereignty

[3]  Council of the EU, Transport, Telecommunications and Energy Council (Telecommunications), meeting of 5 December 2025 — information point on the Declaration. consilium.europa.eu — TTE Council, 5 Dec 2025. Declaration text: Declaration for European Digital Sovereignty (PDF)

[4]  Data Center Dynamics, “Mistral AI raises $830m in debt financing for data center in Paris, France” — 13,800 Nvidia GB300 GPUs, 44 MW, Eclairion site in Bruyères-le-Châtel. datacenterdynamics.com — Mistral $830M raise. Bank consortium named at: thenextweb.com — Mistral seven banks

[5]  Covington Global Policy Watch, “Important Projects of Common European Interest (IPCEIs) on Artificial Intelligence and Compute Infrastructure Continuum,” May 2026. globalpolicywatch.com — IPCEI AI and CIC

[6]  Public Investment Fund of Saudi Arabia, “HUMAIN” portfolio page; CNBC, “Saudi AI firm Humain is pouring billions into data centers,” August 27, 2025; Aramco/HUMAIN minority-stake term sheet announcement. pif.gov.sa — HUMAIN · cnbc.com — Saudi AI firm Humain

[7]  Computer Weekly, “UAE’s TII challenges big tech dominance with open source Falcon AI models”; G42 and TII coverage. computerweekly.com — TII Falcon

[8]  Center for a New American Security (CNAS), “Sovereign AI Index” — 130+ initiatives; >80% of investment in Middle East / East Asia; infrastructure = 59% of projects; Nvidia in 52%. interactives.cnas.org — Sovereign AI Index

[9]  LexisNexis, “Cloud Act vs GDPR: Data Protection for EU Companies” — extraterritorial reach explained. lexisnexis.com — CLOUD Act vs GDPR

[10]  Gartner press release, November 12, 2025: “Geopolitics Will Drive 61% of CIOs and IT Leaders in Western Europe to Increase Reliance on Local Cloud Providers” — survey of 241 Western European CIOs/IT leaders, May–July 2025. gartner.com — Western European CIO survey

[11]  Accenture, “Europe Seeking Greater AI Sovereignty,” November 3, 2025 — survey of 1,928 organizations across 28 countries and 18 industries, July–August 2025. newsroom.accenture.com — Europe AI sovereignty study

[12]  RCR Wireless News, “Huawei to deploy Telefónica’s first commercial 5G-A network in Spain,” August 22, 2025. rcrwireless.com — Huawei / Telefónica

[13]  McKinsey, “The sovereign AI agenda: Moving from ambition to reality,” December 18, 2025 — global survey of 300 executives, investors, and government officials. mckinsey.com — Sovereign AI agenda

[14]  Fortune, “Uber burned through its entire 2026 AI budget in four months,” May 26, 2026 — CTO Praveen Neppalli Naga statement; COO ROI questions; ~5,000 engineers; spend distribution. fortune.com — Uber COO AI spending

[15]  Windows Central, “Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI,” May 2026 — Experiences + Devices division; May 14 rollback; June 30 cutover; FY-end alignment. windowscentral.com — Microsoft cancels Claude Code

[16]  Gizmodo, “Sam Altman Says Intelligence Will Be a Utility, and He’s Just the Man to Collect the Bills” — BlackRock 2026 Infrastructure Summit, March 11, 2026. gizmodo.com — Altman “on a meter”

[17]  Google for Developers, “Gemini Code Assist feature deprecations” (legacy tools removed October 14, 2025) and “Gemini API deprecations” (legacy Generative AI Python SDK EOL November 30, 2025). developers.google.com — Code Assist deprecations · ai.google.dev — Gemini API deprecations

[18]  European Commission, “Data Act explained” — switching charges including data egress prohibited from January 12, 2027. digital-strategy.ec.europa.eu — Data Act explained

[19]  Fortune Business Insights, “Sovereign Cloud Market Size, Share, & Growth Report [2034]” — $195.35B in 2026; $1,133.3B by 2034 (CAGR 24.60%). fortunebusinessinsights.com — Sovereign Cloud market

[20]  McKinsey, “Sovereign AI ecosystems for strategic resilience and economic impact” — projects 30–40% of AI spending will be sovereignty-influenced ($500–600B by 2030). mckinsey.com — Sovereign AI ecosystems

© 2026 SiliconANGLE Media Inc. and Agentcy Labs.  All rights reserved.

Image: SiliconANGLE/ChatGPT

Apple Inc. today sued OpenAI Group PBC for allegedly stealing intellectual property related to its consumer devices.

The iPhone maker filed the complaint with the U.S. District Court for the Northern District of California.

OpenAI entered the consumer electronics market last year when it bought io Products Inc., a startup founded by former Apple executives. The Information reported in February that the ChatGPT developer was working on a smart speaker and smart glasses. Additionally, it’s believed that OpenAI could launch a smart lamp with artificial intelligence features.

The company has reportedly hired more than 400 former Apple employees to support its hardware push. According to today’s lawsuit, OpenAI has been instructing candidates to bring blueprints and device prototypes from the iPhone maker to interviews. The ChatGPT developer allegedly also asked some Apple staffers to share engineering methodologies.

One of the individuals allegedly involved in the effort is io Products co-founder Tang Tan, who is named as a defendant in the lawsuit. Tan spent more than two decades at Apple before launching the startup. He is currently OpenAI’s chief hardware officer.

According to the lawsuit, Tan instructed Apple staffers recruited by OpenAI to stay as long as possible at the iPhone maker. He allegedly provided them with information about Apple’s security procedures. The company charges that the information came from a confidential internal document improperly retained by Tan after his departure.

The other individual that the lawsuit names as a defendant is Chang Liu, a former Apple engineer who joined OpenAI this year. The iPhone maker alleges that Liu failed to return a work laptop after his departure. Furthermore, Apple claims that Liu used the machine to log into its internal network and download dozens of files that contained information about upcoming products.

The iPhone maker is seeking damages from OpenAI. Additionally, it has asked the court to block the ChatGPT developer from using its trade secrets without permission. 

“At every level, from members of its Technical Staff to its Chief Hardware Officer, and in coordination with business partners, OpenAI has been stealing Apple’s trade secrets and confidential information,” Apple wrote in the lawsuit.

OpenAI said in a statement responding to the complaint that “we have no interest in other companies’ trade secrets. We remain focused on building innovative technology that empowers people everywhere.”

The lawsuit is particularly notable because Apple inked a software partnership with OpenAI in June 2024, a year before the io Products acquisition. The companies developed an integration that enables Siri to route complex prompts to ChatGPT. This past January, Apple announced plans to release a new version of the AI assistant that will use Google LLC’s Gemini model series.

Photo: Pixabay

The European Union has tentatively found that Meta Platforms Inc. breached the bloc’s DSA tech industry law.

The European Commission, the EU’s executive arm, published its conclusions today.

The DSA, or Digital Services Act, is a piece of legislation that went into effect in early 2024. It prohibits tech giants from using interface design tactics that can make consumers addicted to their platforms. EU officials believe that Meta ran afoul of that rule with Facebook and Instagram.

The European Commission has taken issue with the two social networks’ content recommendation features. In particular, officials pointed to feed personalization settings that make users more likely to continue browsing. They also raised concerns about Meta’s infinite scroll and video autoplay features.

The EU’s findings are the product of a probe that kicked off shortly after the DSA went into effect. The investigation placed particular emphasis on the impact of Meta’s interface design choices on minors.

The company provides teen accounts that ask users to stop browsing Facebook and Instagram after one hour of daily use. Additionally, the accounts mute notifications between 10 p.m. and 7 a.m. According to the EU, those settings “do not lead to a meaningful reduction” in social media use because they can be easily disabled.

Meta provides a tool that enables parents to apply stricter settings to teen accounts. In particular, the feature can disable Facebook and Instagram access after a daily use limit is reached. The EU found Meta’s controls to be lackluster because they’re “only effective if parents and guardians possess adequate technical expertise.”

Officials also flagged several other issues during the investigation. According to the EU, Meta disregarded information about the way its interface design practices factor into excessive social media use. Additionally, the European Commission found the mental health resources that Meta provides via its Safety Center portal to be lackluster.

The company could face a fine equal to up to 6% of its annual revenue if the EU confirms its preliminary findings. In addition, the European Commission has signaled that it will require Meta to change some of its interface components. The Facebook parent has the option to challenge the findings.

Meta told CNBC that “we disagree with these preliminary findings, which don’t accurately take into account the significant steps we’ve taken to protect teens.”

The company also faces a second EU probe over its practices in the social media market. Its terms of service specify that children under the age of 13 may not access Facebook and Instagram. In April, EU officials tentatively found that Meta doesn’t enforce the rule effectively. The decision could lead to a fine and an order requiring the company to implement stricter controls.

Image: Unsplash

Data scraping startup Oxylabs UAB has broken into unicorn territory after raising $130 million in funding from the private equity firm Warburg Pincus LLC.

The round is the first time the bootstrapped company has ever sought outside funding, and lifts its valuation to a cool $3.6 billion, it said today.

The Lithuanian startup provides artificial intelligence model developers with the comprehensive infrastructure they need to collect publicly available internet data at the enormous scale required to train advanced frontier models. Founded in 2015, it was originally a premium web proxy service provider, a kind of tool that routes requests through many different URLs in order to collect prices, listings and security data without getting blocked.

Nowadays, Oxylabs positions itself as more of a “web intelligence” platform, capable of handling billions of requests per day on behalf of its customers. Its proxy service has access to a pool of over 175 million “ethically sourced” consumer device IPs, which allow it to scrape internet data at scale without being hindered by IP bans, CAPTCHA tools and geo-restrictions. It also provides a collection of tools to developers, such as its fully-managed Web Scraper API data collection tool, and Web Unblocker, an AI-native proxy manager that automatically bypasses anti-scraping systems.

Besides scraping data to build vast training datasets for AI models, Oxylabs can also help autonomous AI agents that do work on behalf of humans to navigate the internet at scale. With its stealthy Headless Browser, agents have the perfect tool for extracting data from heavy dynamic websites.

Oxylabs co-founder and Chief Executive Vytautas Savickas said his company’s tools are essential to enable the agentic web. He points out that AI agents already browse the web more than humans do. “They need a live feed of what is out there, not a stale index,” he said. “The next generation of AI won’t be powered by static indexes.”

The amount of money raised is impressive for a company that was entirely bootstrapped until now. The company reckons it sees big demand for its web scraping services, counting more than 350,000 customers globally, which have helped it grow its annual recurring revenue to more than $350 million. Oxylabs will use the funds to expand its global network and build the next generation of web scraping tools. “The future belongs to the live infrastructure that grounds these systems in real-time, interruption free knowledge,” Savickas said.

Ethical questions

The capital may also help Oxylabs navigate the general controversy that surrounds the practice of web scraping. AI companies such as OpenAI Group PBC and Anthropic PBC have come in for heavy criticism for the way they have basically just helped themselves to the world’s biggest dataset, gobbling up billions of articles, images and movies posted online, without offering a cent in compensation to the people who created all of that data.

OpenAI, Anthropic and other leading AI model makers have been sued many times for their web scraping practices, but the companies that aid them in this haven’t gone unnoticed. Last year, Reddit Inc. slapped Perplexity AI Inc. and three data-scraping service providers — including Oxylabs – with a lawsuit, accusing them of trawling through its copyrighted content. It famously compared Oxylabs and the others to “bank robbers,” saying that they will do almost anything to access its data – except pay for it.

Fact is, large-scale web scraping is a bit of a legal gray area, and Oxylabs recognizes this, trying to position itself as one of the industry’s most ethical scraping services providers, if such a thing is possible. The company strives to maintain compliance with European Union laws on data collection, which are some of the strictest in the world, and it’s also a founding member of the Ethical Web Data Collection Initiative. Known as EWDCI, it’s an international industry-led consortium and working group that’s trying to establish ethical standards and best practices for the data aggregation industry. It sits under the umbrella of the Internet Infrastructure Coalition, and acts as a kind of regulatory framework for data collection firms that want to maintain digital trust.

Regardless of the controversy around web scraping, there is no denying that companies like Oxylabs are emerging as critical players within the broader AI industry. While AI agents get most of the headlines for coding, creating graphics, making movies and automating business tasks, they wouldn’t be able to do any of that without the data infrastructure that Oxylabs provides.

The money comes from Warburg’s $4 billion Capital Solutions Founders Fund, which has previously backed another Lithuanian startup, Nord Security UAB, the creator of NordVPN. It now has a sizable stake in Lithuania’s two biggest technology firms. “Oxylabs has established itself as a leader in web intelligence through its sophisticated technology and expansive network,” said Warburg Pincus Principal Allison Ross.

Image: Oxylabs

Secure remote privileged access company Cyolo Ltd. today announced Cyolo CPS Segmentation, a microsegmentation product that extends its platform beyond human-to-machine access into controlling how machines and systems communicate across critical infrastructure.

The offering targets a gap industrial operators have long acknowledged but rarely closed. Segmentation caps how far an intruder can travel after breaking into a network. Putting it in place has been the hard part. Older methods call for network redesigns and infrastructure changes and both bring downtime.

In plants and utilities that run around the clock, that’s a dealbreaker. Many operators put it off and relied on detection and response instead.

That tradeoff is getting harder to justify. Operations keep adding connections and artificial intelligence is speeding up attacks. Cyolo’s pitch is to box threats in before they move sideways and to do it without the rebuilds that sank earlier segmentation efforts.

CPS Segmentation runs the full segmentation lifecycle. It discovers assets, maps how they communicate, then helps teams build, test and enforce policy. The feature set includes shadow-access discovery, visual flow mapping, blast-radius analysis, intelligent grouping and policy recommendations. A simulation mode lets teams try changes before they go live. Everything sits on Cyolo’s zero-trust architecture, which the company says keeps sensitive data under customer control.

“We’ve pursued a single goal since Cyolo was founded, to help the organizations that keep our world running control every connection across their critical infrastructure,” said co-founder and Chief Executive Almog Apirion. “We started with human-to-machine access and now we’re bringing the same OT-native security principles to machine-to-machine communication control. As AI helps attackers move faster and farther, the ability to stop lateral movement and reduce blast radius has never been more crucial.”

Cyolo describes CPS Segmentation as the first secure connectivity platform built for critical infrastructure. The company positions it as the latest addition to a broader platform that also spans remote privileged access, identity security, malware detection, secure web access and endpoint security.

The launch continues Cyolo’s push deeper into operational technology. The company introduced Cyolo PRO, its OT-focused remote access product, in 2024 and expanded its privileged access platform with vendor-connection monitoring last year.

Cyolo is a venture capital-backed startup that has raised $85.2 million in funding over four rounds. Investors include National Grid Partners Inc., Glilot Capital Partners, Flint Capital, Differential Ventures Inc. and Merlin Ventures.

Image: Cyolo

Chip startup SambaNova Inc. today announced that it has raised $1 billion in funding at a $11 billion valuation.

General Atlantic led the Series F round with contributions from more than a dozen others. Intel Capital, Vista Equity Partners and JPMorgan Chase & Co. were among the participants. The investment follows a $350 million round in February. 

SambaNova debuted its flagship product, an inference chip called the SN50, in conjunction with the February raise. The company says the accelerator can provide more than three times as much throughput as Nvidia Corp.’s B200 graphics card. The SN50’s top speed, in turn, is described as being five times faster.

Artificial intelligence models generate prompt responses through an iterative process. A set of artificial neurons analyzes the user’s prompt, generates a preliminary response and saves the response to memory. A second set of artificial neurons then retrieves the preliminary response, refines it and saves the results back to RAM. The workflow is repeated numerous times until the AI arrives at an answer.

The repeated movement of data to and from memory accounts for much of the time required to generate prompt responses. According to SambaNova, the SN50’s performance is the result of an architecture that speeds up on-chip traffic. 

The SN50 comprises dozens of modules called tiles that each combine processing circuits with high-speed SRAM memory. The close proximity of the memory to the processing circuits reduces the amount of time it takes data to travel between them, which speeds up inference. When a tile completes the calculations assigned to it, it sends the results to the next tile for further processing.

The SRAM memory in the SN50’s tiles is one of three RAM varieties used by the chip.

According to SambaNova, the SN50 includes an HBM memory pool that can store an AI model’s active model weights and KV cache. Those are the components that perform the bulk of the processing involved in answering prompts. The SN50 also features a DRAM module that enables it to store multiple inactive neural networks. If an application needs to swap the AI model it uses to process prompts, the chip can load a neural network from DRAM to HBM in a few milliseconds.

SambaNova ships the SN50 as part of a 16-chip appliance called the SambaRack SN50. The system uses about 20 watts of power, which means that it produces a relatively limited amount of heat. That enables the system to use standard air cooling equipment instead of the more complicated liquid cooling systems that many high-end graphics card servers require.

JPMorgan, one of the contributors to SambaNova’s new funding round, today announced plans to adopt the SN50. It also intends to use the company’s previous-generation SN40 chip. The bank will integrate the processors into its on-premises inference infrastructure.

SambaNova will use the proceeds from the round to enhance its chip lineup, rack design and software. The company also plans to accelerate its go-to-market efforts.

Image: SambaNova

Meta Platforms Inc. today debuted an image generation model that can write code and search the web.

Muse Image is the second algorithm released to date by Meta Superintelligence Labs, the company’s artificial intelligence research group. The first is the Muse Spark large language model that made its debut in April. Both algorithms are available through the Meta AI chatbot.

Muse Image can generate images based on highly detailed, multi-sentence user prompts. It’s also capable of editing existing photos. Users can have the model remove elements such as fog, change the camera angle and perform other changes. The Meta AI chatbot provides the option to refine Muse Image’s output by uploading a sketch that explains what edits it should make.

Under the hood, the algorithm shares several similarities with reasoning-optimized LLMs. The perhaps most notable is that it features a tool use capability.

If a prompt doesn’t contain all the details that Muse Image requires to generate a file, it can use a search tool to retrieve the needed information from the web. Another built-in tool enables the model to generate code. For example, it could write a Python script to turn the contents of a spreadsheet into a graph. Scripts help increase the accuracy of complex visualizations.

Muse Image can loop in Muse Spark, Meta’s reasoning model, when its coding capabilities are insufficient to process a user request. The latter LLM is capable of turning images generated by Muse Image into websites and video games.

Many image generators refine their output using a method called best-of-N, or BoN. The technique consists of generating multiple media files and picking the one that best aligns with the user’s prompt. 

Muse Image takes a different approach. Meta says the model engages in “deliberate reasoning” before it starts generating an image. According to the company, that method enables the model to make better use of the underlying infrastructure than BoN. Muse Image supports a test-time compute feature, which means that increasing the amount of hardware at its disposal boosts output quality.

The model reviews the images that it generates before displaying them to users and makes tweaks when necessary. Meta says the model’s self-refining behavior, as it calls the feature, emerged on its own during the reinforcement learning phase of training. Reinforcement learning hones an AI model’s reasoning capabilities through trial and error.

On launch, Muse Image is available via the Meta AI chatbot in a limited number of markets. It also powers a set of new image effects in Instagram stories. Meta plans to bring the model to Facebook, Messenger and more parts of Instagram in the future.

Further down the line, the company will launch a clip generator called Muse Video. The current early iteration of the model is in third place on the popular Arena AI ranking of video generators.

Meta plans to make Muse Image available to advertisers through its Advantage+ suite of marketing tools in the coming weeks. The company will likely take the same approach with Muse Video. Given that Meta reportedly plans to launch an AI infrastructure service, it’s possible the two models will eventually become available to developers via an application programming interface. 

Image: Meta

OpenMatter Network Inc. today launched a platform it says lets organizations collaborate, run sensitive workloads and deploy artificial intelligence agents across computing environments they do not fully control, using cryptography to prove what happens rather than trusting that it happened.

The Melbourne, Florida-based startup is pitching the product as a “verifiable trust layer” built on a premise it sums up as “Don’t Trust Data. Prove It.” The pitch is that organizations should be able to mathematically verify how their data is used and how AI systems behave, rather than assuming the underlying systems are secure.

The launch comes as enterprises increasingly run AI agents that act on their own across applications and organizational boundaries. Its argument: trust-based security falls apart once data and AI workloads land in systems a company does not own.

The platform does not replace anything. It sits on top of the cloud, data and AI tools a company already runs, adding cryptographic verification and tighter control over how workloads execute. The company says it pairs that verification with enforceable policy controls and a distributed architecture.

OpenMatter’s platform centers on three components. Masked Compute allows execution and computation across organizations without exposing the underlying data. QuantumGuard handles policy enforcement and governance for AI agents working across different systems. Datavizor acts as a visibility layer, producing what the company describes as a cryptographically provable record of execution and AI activity.

“For decades, organizations have been asked to trust the systems they rely on,” said co-founder and Chief Executive Renee Davis. “We believe the next generation of digital infrastructure will be built on proof. They don’t need to replace the infrastructure they already rely on, they need the ability to verify what happens across it.”

Co-founder and Chief Technology Officer Ada Anderson framed the gap in governance terms, arguing that “policy prompts and assumed compliance are not enough” and that organizations cannot govern what they cannot prove happened.

The target market is anywhere sensitive data changes hands between parties. OpenMatter said the platform suits healthcare collaboration, secure AI model training, financial analytics and distributed scientific research. It has an early partnership with Dara AI Ltd., a privacy-focused health data platform, to explore whether the technology can generate healthcare insights without compromising the privacy of the individuals behind the data.

OpenMatter did not disclose pricing, or general availability. Founders Davis and Anderson have backgrounds in secure systems architecture, distributed computing and cryptography.

Image: OpenMatter

Build Inc., an artificial intelligence-driven startup that automates complex industrial real estate development management projects, today announced it has raised an $8.5 million seed round led by Index Ventures.

Pebblebed, Puzzle Ventures and Tiny.vc also participated in the round. A range of industry-focused angels joining the round included OpenAI Group PBC Chief Financial Officer Sarah Friar, Blackstone Inc. Chief Technology Officer John Stecher and senior figures from OpenAI, Meta AI Research and Google Maps.

Headquartered in New York, Build provides a platform that combines architectural and AI expertise. The company provides an AI system named Dougie, which automates complex infrastructure workflows including site sourcing, technical due diligence, power assessment and early design, helping customers reduce project timelines.

Build says Dougie compresses what would normally take humans over four weeks into about 75 minutes, rapidly prototyping the research, planning, diligence and curation for industrial real estate projects. The company said it pulls from more than 1,600 data sources and has been deployed across more than 100 projects in 15 countries for governments, Fortune 500 companies and institutional real estate groups.

“The industries shaping the physical world have spent decades trapped in process instead of creativity,” said co-founder and Chief Executive James Stirrat-Ellis. “By removing that operational burden, we can help teams move faster, make better decisions and deliver better infrastructure.”

In a blog post, the company laid out the potential for a fleet of AI agents, all working on that vast body of data, information, documents and other information that would take a number of expert human minds hours across many days to retrieve, read, critically examine and finally come to conclusions. This includes land due diligence for everything from solar farms to data centers, on regulatory interests, where water and power lines come from, how the local ecosystem is affected, what the neighboring cities and roads look like and how that affects the ability to build.

If mishandled, the company stressed, this could cost millions of dollars.

One customer, Tishman Speyer, a leading global real estate owner, developer and operator, has pivoted from building mostly office buildings to joining the recent rush to construct hyperscale data center infrastructure. In May 2024, TS entered the data center market with a 32-megawatt campus in Frankfurt, expandable to 70MW, a joint venture with data center builder Mainova WebHouse GmbH.

The explosive expansion of generative AI, high-performance computing and hyperscale cloud computing has led to rapid growth in the data center market, often described in superlative terms as a massive infrastructure supercycle. Global data center capacity is on track to double, according to a report from market analyst Jones Lang LaSalle IP Inc., which estimates that up to $3 trillion in real estate and information technology will be required by 2030 to add nearly 100 gigawatts of capacity.

That metric only examines the compute infrastructure cycle, which will drive and parallel the need to upgrade and retrofit energy grids and water delivery systems.

Photo: Unsplash

Agentic ecosystem security startup Vorlon Inc. today launched Guardian, a real-time enforcement gateway that aims to block risky actions by artificial intelligence agents before a transaction completes rather than flagging them after the fact.

The product targets a gap Vorlon argues most agent security tools leave open. AI agents do not log in. They authenticate with OAuth tokens and application programming interface keys, chain actions across systems, and move sensitive data at machine speed with no human in the loop. As the company puts it, with agents, the activity is the threat, not the access.

Guardian sits at the protocol layer between agents and the systems they connect to, applying policy before any transaction goes through. Security teams can block actions that violate policy, mask sensitive data in transit before it reaches an unauthorized destination and downgrade an agent’s access to read-only where write access is not warranted. Policies are set both at the agent platform level and at each connected system, so a data lake holding personally identifiable information can run under stricter rules than a project management tool.

Vorlon says the gateway covers any application or data store with an API or a Model Context Protocol server and can be turned on in minutes. That includes homegrown and citizen-developed apps, legacy systems and tools built with AI coding assistants such as Anthropic PBC’s Claude Code and OpenAI Group PBC’s Codex. Many competing tools, the company contends, either watch the user prompt instead of the agent’s actions or work only in environments with native MCP support.

The launch comes amid sharp unease among security leaders. Vorlon’s “2026 Agentic Ecosystem Security Gap Report” found that 75.4% of chief information security officers rate AI agents a critical or significant risk, 99% are worried about an AI or software supply chain breach this year and fewer than 1% feel adequately protected.

The company points to a concrete failure to make its case. In April, an AI coding agent deleted the entire production database and all backups of a service called PocketOS in nine seconds, ignoring explicit rules against destructive operations. Guardian’s read-only enforcement would have stopped the writes regardless of what the model decided, Vorlon said.

Guardian draws on DataMatrix, the patented simulation engine Vorlon introduced in April 2025 that maintains a live behavioral model of every agent, app, identity, integration and data flow in an environment. The engine discovers shadow agents and tools automatically, which Vorlon says lets teams add enforcement without building an inventory by hand.

“Enterprises have spent years building security programs around the assumption of governing access equaling governing risk,” noted co-founder and Chief Executive Amir Khayat. “AI agents broke that assumption. They operate through legitimate access, at machine speed, across systems no single team fully owns.”

Guardian, he said, is the enforcement layer the agentic era has been missing. “It acts before the transaction completes, not after the damage is done,” he added.

Vorlon, founded in 2022 and backed by Accel, integrates with security information and event management, security orchestration, information technology service management and data loss prevention platforms, including Netskope Inc., Microsoft Purview and Google DLP. Guardian is available now.

Image: Vorlon

Perforce Software Inc. today expanded its Perforce Intelligence lineup with an agentic gateway for managing artificial intelligence agents, an autonomous testing platform driven by natural language and a unified compliance tool that turns written security policies into continuous enforcement.

The DevOps company is pitching the releases at enterprises and regulated industries trying to move AI from experiments into production without losing control of cost, governance and quality. Perforce frames the problem as a control and trust gap that widens as AI agents touch more of how software gets built and shipped.

At the center is Perforce Agentic Gateway, an orchestration layer that sits in front of the Model Context Protocol, the standard for connecting AI agents to tools and data. The gateway gives organizations a single install and guided setup to access Perforce’s MCP portfolio across code, intellectual property, data, infrastructure and testing. It can also govern third-party MCPs and the company said it reduces token consumption, a recurring cost as agent usage scales. The gateway is available now through GitHub.

Perforce Autonomous Testing lets non-testers describe what they want to validate in plain language through a single chat interface, with AI then carrying out the work. The platform spans functional, performance and mobile testing and a single test can run across desktop web, iOS, Android, accessibility and performance. It is built on BlazeMeter and Perfecto, two testing products already in Perforce’s portfolio, with a Delphix integration for test data planned for later releases.

The third release, Perforce Intelligence Unified Compliance, sits above existing Puppet deployments and translates internal and external policies into code. It routes enforcement to the relevant infrastructure, including Kubernetes environments, monitors for drift, remediates violations and keeps the audit evidence that regulators require. Initial capabilities include natural language policy controls, cloud cost and financial compliance enforcement and executive dashboards showing compliance posture.

Perforce is positioning the bundle as a control plane for what it calls the AI-driven development lifecycle, citing pressure on executives to show returns on AI spending. The company pointed to a Harris Poll survey commissioned by Dataiku Inc., in which 80% of more than 900 chief executives said their roles would be at risk if their companies failed to deliver measurable business gains from AI by the end of 2026.

“Scaling AI across the software delivery lifecycle or AI lifecycle introduces new challenges around orchestration, governance and compliance,” said Jim Mercer, program vice president at International Data Corp. “There’s a lot of money being spent with AI, but the next evolution needs to be on ROI of AI, which can only be measured if you have control mechanisms in place and visibility into what it’s doing.”

Perforce said its software is used in more than 80 countries, including by more than 75% of the Fortune 100 and half of the Global 500.

Image: Perforce

A consortium of more than 140 financial, payments and technology companies, including Visa Inc., Stripe Inc. and BlackRock Inc., is backing a new stablecoin called Open USD, taking direct aim at the market leaders Tether and Circle Internet Group Inc.

The token, also known as OUSD, was unveiled by Open Standard LLC, an independent company that will operate it with a board drawn from its partners. Open USD will launch later this year. The group did not say which blockchain it will run on, though Tempo, the Stripe-aligned payments chain, said OUSD will be natively issued on its network from day one.

The pitch to business is built on economics. Companies can mint and redeem Open USD with no fees and no volume limits and partners keep most of the revenue earned on the reserves backing the token, minus a management fee that covers operating costs. That structure breaks from the dominant model, in which the issuer pockets the interest income on reserves.

“Existing stablecoins have great strengths, but to use them at scale, businesses need something that’s open, low-cost, high-throughput, broadly accessible and aligned to their interests,” Zach Abrams, founding chief executive of Open Standard, said in the announcement. Abrams co-founded the stablecoin startup Bridge Ventures Inc., which Stripe acquired for $1.1 billion in 2025.

The lineup pulls from across finance and tech. Payment networks Mastercard Inc., American Express Co. and Discover Financial Services Inc. signed on, as did banks and asset managers BlackRock, The Bank of New York Mellon Corp. and Standard Chartered plc. On the technology side are Google LLC, Shopify Inc. and IBM Corp.

Crypto firms feature heavily too, among them Coinbase Global Inc., Ripple Labs Inc., OKX and MetaMask. Klarna Group plc, Affirm Holdings Inc., DoorDash Inc., Western Union Co. and MoneyGram International Inc. round out the early partners.

Notably absent are Tether and Circle, the two issuers who dominate the market. Tether’s USDT accounted for roughly 62% of stablecoin supply in April and Circle’s USDC about 25%, according to CoinGecko. Shares in Circle fell about 13% after the announcement.

“We welcome continued innovation and competition in the space,” Circle Chief Executive Jeremy Allaire wrote on X.

Open USD is the latest entrant in a stablecoin rush that followed the Genius Act, the regulatory framework President Donald Trump signed into law in July 2025. Klarna launched its own KlarnaUSD in November and Amazon.com Inc. and Walmart Inc. have both signaled interest in issuing tokens.

Consortium efforts are not new to the sector. Paxos launched USDG through its Global Dollar Network in late 2024 with backers including Mastercard and Robinhood, and a group of large banks led by JPMorgan Chase & Co. unveiled a shared on-chain deposit network earlier this month.

Open USD goes further on both scale and economics. Its launch list runs past 140 names, and its revenue-sharing model gives partners a financial stake in the token’s growth rather than leaving the interest income with a single issuer. Tether and Circle have spent the past decade competing against each other. They now face a coalition.

Image: Open Standard

Anthropic PBC today debuted Claude Sonnet 5, a midrange large language model that outperforms its predecessor in several areas.

The LLM will be the default option in the consumer tiers of the company’s Claude chatbot service.

Anthropic’s commercially available LLMs are organized into three product families: the entry-level Haiku series, Sonnet and the high-end Opus lineup. In April, the company debuted two LLMs called Mythos 5 and Fable 5 that are even more capable than Opus. However, they’re not yet broadly accessible.

Anthropic measured Sonnet 5’s coding capabilities using two benchmarks called SWE-Bench Pro and Terminal-Bench 2.1. The model improved upon the scores of its predecessor by 5.1% and 13.4%, respectively. GPT-5.6 Terra, OpenAI Group PBC’s competing midrange LLM, outperformed Sonnet 5 on Terminal-Bench 2.1 by about 4%.

Anthropic also tested how its new LLM fares in other areas. The model set a score of 1,618 on GDPval-AA v2, a benchmark that includes knowledge work tasks spanning 44 professionals. Sonnet 4.5 earned 1,395 points.

One of the contributors to Sonnet 5’s increased output quality is that it’s more autonomous. According to Anthropic, users who tested the model before its release reported that it sometimes double-checks its output without instructions to do so. Additionally, Sonnet 5 can perform tasks that are too difficult for its predecessor.

Increased LLM autonomy can create cybersecurity risks in some cases. According to Anthropic, Sonnet 5 is better than Sonnet 4.6 at fending off such risks. In particular, it’s more adept at blocking malicious requests and prompt injection attacks. A prompt injection attack is a malicious instruction hidden in the data analyzed by an LLM.

Sonnet 5 includes guardrails that prevent hackers from using it to launch cyberattacks. According to Anthropic, the model poses a limited cybersecurity risk because it can’t develop working exploits.

Sonnet 5 is the new default model in the free and consumer-focused Pro tiers of the Claude chatbot service. It will also become available in the Max, Team, and Enterprise plans. Developers, meanwhile, can access the model through Anthropic’s application programming interface. Sonnet will be priced at $3 per million input tokens and $15 per million output tokens starting in September, slightly more than OpenAI’s midrange Terra.

Anthropic customers will soon also gain access to two other new models. Earlier this month, the company debuted a flagship LLM called Mythos 5 and a scaled-down version known as Fable 5. The latter model includes stricter guardrails that block potentially risky prompts.

The U.S. government imposed export controls on Mythos 5 and Fable 5 a few days after their introduction. In response, Anthropic paused the models’ rollout. The company disclosed today that the controls have been lifted, which will allow it to start restoring access on Wednesday. Anthropic plans to make Fable 5 broadly available, while Mythos 5 will be accessible only to a limited number of trusted organizations.

Image: Anthropic

French multinational energy giant Schneider Electric SE said today it’s buying an industrial artificial intelligence company called Cognite Holding B.V. in a $3.1 billion all-cash deal, with an eye to strengthening its AI software and industrial data business.

The deal highlights the evolving nature of industrial AI, which for years simply followed what was happening industrial facilities such as power plants, recording outputs and flagging problems that required maintenance. Nowadays, industrial AI is becoming far more autonomous, able to make decisions and take actions on its own, and Schneider has decided it needs to own that.

“Cognite has built something rare, a truly industrial-grade AI platform,” said Schneider Chief Executive Olivier Blum. He said the acquisition will help put his company at the center of “the next phase of industrial intelligence.”

The company has developed AI software that pulls messy industrial data from across facilities into one place, so that intelligent agents can act on it autonomously. It was founded in 2017 and has grown to employ more than 800 staff across the Americas, Europe, the Middle East and Asia-Pacific.

Its platform marries a unified data model and knowledge graph with agentic AI agents that help to clean and connect the data that pours out of industrial machines. Other agents can then analyze that data and automate workflows based on what it tells them.

There are two key products that do this. Data Fusion is the data infrastructure that handles the messy work of contextualizing and operationalizing industrial data. Then Atlas AI provides the agentic layer that automates tasks and speeds up decision-making. What this means is that an agent might be able to spot a failing pump, order the parts required to fix it, and then schedule the repair, with a human signing off on that.

Schneider said it will integrate these products with the Aveva CONNECT platform. Aveva is a Cambridge, U.K.-based software unit previously acquired by the energy giant that’s focused on the design, operation and optimization of industrial assets.

Cognite’s AI tools have already been deployed in industries such as power generation, oil and gas and manufacturing. These sectors sit on masses of data that has been accumulated over decades but rarely put to much good use. Cognite’s AI agents finally change that, which explains why it has grown so much since its founding.

In 2025, its revenue exceeded $170 million, with recurring bookings increasing by 36%. Cognite’s backers will profit immensely from the sale. For instance, the Norwegian investment firm Aker, which helped to found the company back in 2017, is expecting to receive around $1.48 billion in cash from the sale, Bloomberg reported.

The deal comes at a time of growing interest in industrial AI, especially in Europe, where manufacturers are striving to increase their efficiency and reduce waste. Blum said Europe is also keen to switch to cleaner forms of energy and reduce its reliance on Russian oil and gas, and that transition requires intelligence.

In turn, intelligence needs data, and to unlock that data, AI is required too. Schneider, he said, already sells the hardware that powers industry, and Cognite gives it the software needed to make that hardware smarter.

“By bringing Cognite into Schneider Electric and Aveva, we unite the world’s most comprehensive energy management and automation infrastructure with the software and AI capabilities to make it natively intelligent,” Blum said. “Together, we go beyond connecting systems. We give them the ability to think, adapt and act. This is what industrial intelligence looks like at scale.”

Schneider said the transaction is subject to regulatory approval, but it expects the deal to close in the “coming quarters.”

Image: Cognite

Etched Inc., a developer of artificial intelligence inference chips, launched today with $800 million in funding.

The startup raised the capital over multiple rounds. The most recent investment, which closed in December, valued Etched at $5 billion. It included the participation of VentureTech Alliance, a startup fund associated with Taiwan Semiconductor Manufacturing Co. It was joined by more than a dozen other backers including Geoffrey Hinton, Fei-Fei Li and Andrej Karpathy.

Etched plans to produce its inference chips using TSMC’s N4P process. It’s an enhanced version of the chip giant’s five-nanometer node that provides 11% better performance than the original. According to Etch, its first prototype chips rolled off TMSC’s N4P production line earlier this year.

Nvidia Corp.’s flagship Rubin graphics processing unit is optimized for both AI training and inference. Focusing solely on inference, the approach that Etched has taken, makes it possible to reduce power usage by removing training-optimized circuits. Alternatively, engineers can add more inference circuitry to boost processing speeds.

Etched has also equipped its chip with several other performance optimizations.

The more calculations a GPU performs per second, the more power it draws, which in turn increases its operating temperature. Past a certain threshold, the extra heat can cause malfunctions. Graphics cards mitigate the issue by lowering their clock rate when they start approaching their peak speeds. Thermal throttling, as the practice is known, slows down inference. 

Etched has developed a technology called LVI that reduces the need for thermal throttling. According to the company, its chip can run a trillion-parameter AI model at “80%+ peak FLOPs” without lowering its clock rate. The result is a significant inference speedup. Etched says that its chip’s FLOP density, a measure of performance, is several times higher than the existing AI processors on the market.

The company plans to ship its silicon as part of a rack-scale inference appliance. The system features multiple chips installed on custom circuit boards. Etched has also developed custom cold plates, components that play an important role in liquid cooling systems. A cold plate is a flat piece of metal that channels heat generated by chips into a rack’s coolant.

Etched’s appliance includes a mix of SRAM and HBM memory. AI chips use SRAM, the fastest RAM variety on the market, to store their workloads’ most important data. Other information is sent to HBM memory, which trades off some speed for significantly increased capacity.

The AI chips in a rack often require the ability to access data in one another’s memory. Etched’s appliance uses a custom interconnect to facilitate such data movement. According to the company, the machine features a system-wide shared memory pool that can process requests with less latency than earlier technologies.

Etched is currently in the process of ramping up chip production and plans to ship its first racks this summer. The company disclosed today that it has received more than $1 billion worth of customer orders.

Image: Etched

Chip component supplier Stathera Inc. today announced that it has closed a $35 million funding round.

Semiconductor-focused fund Maverick Silicon led the Series A deal. It was joined by chipmaker MediaTek Inc.’s venture capital arm, Celesta Capital and several others.

Every modern processor includes an oscillator, a component that generates a signal at specific time intervals. That signal sets the pace at which processing is performed. Without an oscillator pulse, circuits can’t accurately determine when they should start a new calculation, which mixes up the order of computing operations.

The oscillator that a processor uses to coordinate computations is known as its clock. It generates a signal billions of times per second. Processors also include a second, slower oscillator that the host computer uses to track what time it is. The second oscillator generates a signal precisely 32,768 times per second.

Montreal-based Stathera has consolidated the features of a processor clock and a timekeeping oscillator into a single device. The technology enables engineers to equip their chips with one oscillator instead of two, which reduces the associated space requirements by up to 85%. It also reduces power use in the process.

Standard oscillators are built around a tiny quartz crystal. An electric current causes the crystal to vibrate, producing movements that each correspond to a single processor clock signal.

Stathera’s two-in-one oscillator swaps the quartz crystal for a structure composed of single-crystal silicon, the material from which chip wafers are made. That makes it possible to produce the device at chip fabs. When a fab operator adopts a new manufacturing process, Stathera can use the technology to miniature its oscillator. Such advances are not possible with quartz devices.

Stathera says its technology also provides several other benefits.

The vibrations that an oscillator uses to generate signals can damage it over time. Silicon-based oscillators such as Stathera’s device have a smaller mass than quartz devices, which decreases the strength of vibrations and thereby lowers the risk of malfunctions. To further boost its devices’ durability, Stathera places them in a vacuum-sealed enclosure. That isolation protects  oscillators from debris and other sources of damage.

Silicon oscillators are less prone to certain reliability issues than their quartz-based counterparts. However, they experience drift, or a decrease in accuracy, when the temperature of the host computer changes. That drift amounts to 30 signals out of a million every time there’s a change of one degree Celsius. According to Stathera, its oscillator design includes features that mitigate temperature-related drift.

The company’s product portfolio currently comprises two devices. The STA320 is designed to power medical devices, wearables and other low-power systems while the ST156 is geared towards more advanced products such as smartphones. 

Stathera will use the proceeds from its funding round to ship a new generation of oscillators. Additionally, it will start developing a third-generation device series optimized for artificial intelligence chips. Stathera will hire more engineers and go-to-market professionals to support the effort.   

“AI data center performance is increasingly limited not by raw compute, but by how quickly and coherently data can move and remain synchronized across tens of thousands of processors and the networking connecting them – synchronization that must run on precision timing,” said co-founder and Chief Executive Officer George Xereas. “For decades that timing has meant quartz, but Stathera is using semiconductor technology to move this foundational hardware to a new era of silicon.”

Image: Stathera

Amazon Web Services Inc.‘s next-generation silicon is working its way deeper into the company’s cloud compute infrastructure offerings with the debut of the new Amazon EC2 C9g and EC2 C9gd instances today.

The new virtual machines are powered by the company’s flagship AWS Graviton5 processors, which are central processing units that have been designed to deliver superior compute performance for diverse cloud workloads, including artificial intelligence applications.

The C9g and C9gd instances debut just weeks after AWS launched its general-purpose M9g and M9gd instances, and make the new CPUs available for a range of higher-performance workloads.

In a blog post, AWS Principal Developer Advocate Sébastien Stormacq said the C9g and C9gd instances are designed for compute-intensive and data-heavy tasks including video encoding, scientific modeling and distributed analytics, where performance and efficiency are vital. He explained that the Graviton5 chips provide a major leap forward in this regard, with up to 25% higher performance per vCPU compared to the previous-generation C8g instances that were powered by Graviton4.

The advantage of Graviton5 is that it delivers a massive increase in memory and cache, Stormacq said. Both C9g and C9gd feature DDR5 8800MT/s DIMMs, which is the fastest memory of any cloud-based instance currently available. They also pack in five times more L3 cache and three times higher packet-processing performance than Graviton4, dramatically reducing the time workloads sit idle while waiting for the data they need.

According to Stormacq, this architecture means the C9g and C9gd instances are perfectly suited for next-generation “agentic AI” workloads, where artificial intelligence models complete complex, multistep tasks on behalf of humans. Such workloads increase the computational strain on the underlying processors, which is why Graviton5’s higher core count and bigger caches become necessary.

The C9g instances are optimized for workloads that employ the Amazon Elastic Block Store for storage, including video encoding pipelines and batch processing jobs. Meanwhile, the C9gd instances are designed for applications that rely on high-speed and low-latency local NVMe SSD storage, Stormacq said. In this case, they deliver up to 30% higher throughput and input/output operations compared to the previous-generation instances designed for local storage. They’re a great fit for workloads such as temporary caches, ad-serving engines and high-performance computing simulations, he added.

Customers have plenty of options for scaling up the new instances, with 11 sizes available. These range from the medium-sized VMs that feature one virtual CPU and 2 gigabytes of memory, all the way up to 48xlarge, which pack 192 vCPS and 384 gigabytes of memory.

There are also bare-metal server options, and there are massive gains in network and Elastic Block Store throughput, which have increased by 15% and 20% respectively across instance sizes. In the case of the largest 48xlarge instances, these bring 100 gigabytes per second of network bandwidth and 72 gigabits per second of EBS bandwidth to the table, which is almost twice the previous limits.

Stormacq also discussed a critical new security feature that comes with the instances. The C9g and C9gd families are the first from the company to support the new AWS Nitro Isolation Engine, which is designed to enforce isolation between virtual machines. It works by restricting access to the VM’s memory, CPU states and I/O devices to a minimal set of application programming interfaces to guarantee data confidentiality and integrity.

The new C9g and C9gd instances are generally available now in the US East (Ohio, N. Virginia), US West (Oregon), and Europe (Frankfurt) regions, with additional regions to follow later in the year. They’re accessible via the AWS Management Console, the Command Line Interface and also AWS’ software development kits.

Photo: SiliconANGLE

Agentic AI is driving key technology providers to rethink the computing architecture required to run rapidly expanding autonomous systems.

In response to this challenge, two leading tech companies have recently unveiled a key milestone involving system-level validation for an entire rack-scale architecture. In June, CoreWeave Inc. and Nvidia Corp. announced the first bring-up and validation of Nvidia Vera Rubin NVL72 on CoreWeave Cloud.

The announcement involved a fundamentally different approach to infrastructure, designed to provide an environment where workloads reason continuously, scale unpredictably, and operate in production around the clock. This is testing the limits of data bandwidth, as described by Chen Goldberg (pictured), executive vice president of product and engineering at CoreWeave.

TheCUBE’s John Furrier and CoreWeave’s Chen Goldberg talked about the latest announcements during the event.

“Vera Rubin is not an incremental upgrade – 72 Rubin GPUs, 36 Vera CPUs, 260 terabytes per second of NVLink 6 bandwidth inside a single rack, which is more data bandwidth than is used by the entire global internet,” Goldberg said. “The world is shifting from asking AI questions to having AI actually do things continuously at scale without stopping to close your laptop, agents writing code, running experiments and executing multi-step reasoning loops. This is exactly what Vera Rubin was architected for.”

Goldberg spoke during “Scaling the Agentic Era With Nvidia Vera Rubin NVL72 on CoreWeave Cloud,” a virtual event hosted by theCUBE, SiliconANGLE Media’s livestreaming studio. Executives from CoreWeave, Nvidia and Dell Technologies Inc. spoke with theCUBE about the engineering and operational requirements needed to support production-scale inference and agentic AI workloads, as well as what it would take to build accelerated computing infrastructure for this next phase of AI. (* Disclosure below.)

Rack-controlled computing architecture

CoreWeave’s announcement and theCUBE’s event highlighted the important role of Vera Rubin NVL72 in supporting large-scale inference, persistent reasoning sessions and production AI workloads that require more than raw GPU density. Nvidia’s ability to supply advanced chip architecture has allowed CoreWeave to provide multiple systems solutions, including liquid cooling, rack control, networking and secure multi-tenant operations.

These innovations include CoreWeave’s liquid cooling solution – Valvey – which monitors flow rate, temperature, pressure and leak detection in real time.

“It manages the liquid cooling system in a software-defined way,” Goldberg said. “We can control a single valve at a sub-second timescale, so if we detect any leak, as an example, we take action immediately.”

CoreWeave’s Peter Salanki spoke with theCUBE about emerging trends in computing architecture.

The latest release also included Racky, a new unified rack control appliance specifically designed for aggregating power, cooling and environmental sensors into a standardized management surface. This allows each Vera Rubin rack to be managed as a cloud resource rather than a custom one-off build, which provides system administrators with a bigger picture, according to Peter Salanki, chief technology officer of CoreWeave.

“It takes in telemetry from the GPUs themselves, takes in telemetry from the power systems, takes in telemetry from different leak sensors and from the building management system and allows us to tie all these things together,” Salanki told theCUBE. “It’s all deployed locally in the pod. It will interact with Valvey, it will interact with other systems upstream and downstream to make decisions.”

Leveraging Spectrum-X for GPUs

With multiple CPUs and GPUs in a single rack, communication becomes particularly important. CoreWeave’s announcement includes multi-rail and multi-plane networking, with support for both Nvidia Quantum-X800 InfiniBand and Nvidia Spectrum-X Ethernet with RDMA over Converged Ethernet RoCE.

“The genius of this rack scale system is that it allows you to scale memory, scale compute, scale all the fabrics so that it can talk to each other at full line rate so that GPU number 1 can talk to GPU number 72 at the exact same speed,” said Dion Harris, product leader at Nvidia. “This is what gives you a very consistent, reliable way to scale your workload across the entire rack. When you scale out … that’s where you start to leverage our Spectrum-X co-packaged optics network. That allows you to scale out efficiently across the racks so that as workloads scale across NVL72s, you can run those efficiently as well.”

CoreWeave’s Harshdeep Banwait (center) and Nvidia’s Dion Harris (right) spoke with theCUBE about the integration of processor technology in the latest offerings.

CoreWeave is also leveraging Nvidia BlueField-4 DPUs or data processing units to enable secure, multi-tenant AI cloud operations. The goal is faster data access and lower latency. BlueField-4 allows tenants to run workloads across the full Vera Rubin computing platform while preserving control and security.

“From our standpoint, the hardest thing to do was, when all of it comes together, how does it look and feel in practice?” said Harshdeep Banwait, director of product at CoreWeave. “We brought it all together, went through the validation flow to make sure that the Vera CPUs work with the Rubin chips, to work with ConnectX NICs, to work with the BlueField-4 DPUs. It was just making sure that all of these components talk to each other together and act as a system to essentially unlock that level of performance.”

Integration with Dell PowerEdge

To implement a rack-scale platform such as the Vera Rubin NVL72, CoreWeave drew from its partner ecosystem. Dell provided the architectural backbone for the platform through its high-performance PowerEdge XE9812 servers.

Dell’s involvement is based on a belief that as AI models expand at trillion-parameter scale and context windows encompass millions of tokens, compute density is going to grow in importance, according to Ihab Tarazi, senior vice president and chief technology officer of Dell.

Dell’s Ihab Tarazi and CoreWeave’s Jacob Yundt talked with theCUBE about inference performance and computing density.

“The density metric is going to become very important. How much can you squeeze out of the density and performance?” Tarazi said. “First of all, all the new models that really matter to people are trillion parameter models. So, they no longer fit for the most part. If you want the full performance and you want some of the use cases, they’re not going to fit on an 8-way GPU standard server. They really need those NVL72 GPU systems.”

CoreWeave’s validation of the Vera Rubin NVL72 also underscores the need for inference performance that can support agentic AI in production. According to theCUBE Research, the journey toward a meaningful return on investment from AI is entering a new phase. It has moved from model innovation to operationalizing inference at scale in key business processes.

“As we’ve seen, the inference market has grown exponentially over even just the last couple of years,” said Corey Sanders, senior vice president of product at CoreWeave. “It’s the opportunity for Vera Rubin to now play this really interesting role of both supporting this massive buildup of training while also supporting a huge opportunity for massive inferencing at a cost and performance that I think before would have been impossible. Brand new workloads are coming to life as part of it.”

These new workloads are being driven by increased adoption of agentic AI, and CoreWeave has taken a series of actions over the past year to build a cloud infrastructure that supports it. This has included the acquisition of the AI model development firm Weights & Biases Inc. in 2025.

CoreWeave’s Corey Sanders and Shawn Lewis spoke with theCUBE about new tools to drive agentic AI.

A number of new Weights & Biases agentic AI tools have been added to the CoreWeave platform since then.

“For the first time, there is an agent inside of Weights & Biases that helps those AI users train AI models and build AI applications,” said Shawn Lewis, founder and chief technology officer of Weights & Biases at CoreWeave. “There’s also a feature in Weights & Biases called W&B Launch that connects the Weights & Biases toolkit, which you can think of as a UI that users spend time analyzing data in, back to infrastructure. So it allows W&B users to launch and execute jobs on CoreWeave infrastructure from the UI. Now an agent can, instead of a human … launch experiments onto the infrastructure.”

Redefining rack-scale computing

The partnership between CoreWeave, Nvidia and Dell illustrates an important step in the evolution of software and hardware design. As CoreWeave’s Senior Director of Compute Architecture, Jacob Yundt noted in his interview with theCUBE, the company’s approach to building its platform is shaped by scale and changes in how the computer itself is defined.

An example of the Nvidia Vera Rubin Superchip was on display during the event.

“The NVL72 products have really changed the landscape,” Yundt explained. “They change how you approach engineering. The rack has this high-speed interconnect, this ultra-high bandwidth, ultra-low latency connection. All of the CPUs, all the GPUs, they’re all in one rack-level package. You’re no longer thinking of things as just an individual server. I have to think of them now as racks and then that has a cascading effect, I have all these racks that are working together. Now the rack is the computer.”

In redefining the rack, CoreWeave is also validating how innovation at the hardware and software layers is unlocking new levels of performance, efficiency and scale for AI-driven applications. This will power agentic AI and the systems needed to implement them throughout the enterprise.

“AI is no longer about isolated models,” said theCUBE’s John Furrier. “It’s about systems, systems that bring together compute, networking, storage, software, data, security, and operations into a unified platform capable of delivering real-world outcomes. Today’s discussions highlight an industry moving beyond proof of concept and into execution. The winners in this next phase won’t simply have access to AI, they’ll be the organizations that can operationalize it, scale it, govern it, and continuously innovate around it.”

(* Disclosure: TheCUBE is a paid media partner for the “Scaling the Agentic Era” event. Neither CoreWeave, the sponsor of theCUBE’s coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE

Queue, a company building an autonomous “robotic” pharmacy kiosk that dispenses medication, today announced that it has raised $12.6 million in seed funding led by AlleyCorp.

The company is launching today with $18.6 million under its belt, which includes today’s seed round and $6 million in pre-seed funding led by Riot Ventures less than a year ago. Additional backers bringing capital investment include House Capital, Ubiquity Ventures, Grep Ventures and Banter Capital.

Queue provides a robotic kiosk that supports about 280 medications. It’s designed to make fulfilling a prescription as simple as walking up and displaying a QR code on a phone to verify the script. The idea is to make going to the pharmacy quick and accessible.

“Pharmacy in America is structurally broken,” said co-founder and Chief Technology Officer Josh Liu. “Queue isn’t an incremental fix; it’s a complete reimagining of how medications get dispensed, verified and delivered.”

According to the company, Queue can deliver most medications at 96% lower cost than a traditional pharmacy and can be deployed across retail locations, hospitals, rural communities and other locations where quick access to a drugstore might be difficult.

Although the kiosks have limited space, the selection of 250 medications includes the most commonly prescribed medications in the United States. Each machine would most likely have a curated selection based on the expected needs of the community it serves. A predictive artificial intelligence system tracks the usage and calls a pharmacy technician to refill it when certain meds get low.

Aside from targeting underserved markets, Queue’s machine presents a second opportunity: providing pharmacists in existing venues with relief when they are understaffed and overwhelmed. Customers who arrive at the pharmacy seeking these common meds need not wait in long lines while pharmacy staff handles less common cases, Queue says; they can simply interact with the machine and have their prescription filled quickly.

The company said it has secured a major national pharmacy chain as a customer and deployed a working prototype to work out early market validation. Queue also noted that the retail-facing pharmaceutical industry is currently facing a labor shortage, with schools graduating 3,000 to 4,000 fewer pharmacists than will be needed in the years to come. A figure backed by the trade journal Drugstore News, which also noted growing dissatisfaction in the industry and that 70% of independent pharmacy managers had difficulty filling open staff positions.

According to analyst Persistence Market Research, the retail pharmacy market size was estimated to be $670.6 billion in 2025 and projected to reach $901.4 billion by 2032. Throughout human history, pharmacies have evolved from providing bespoke remedies to communities to becoming the core of the healthcare industry’s logistics, delivering patient-focused care by distributing and dispensing medications. Their role has expanded beyond that origin, especially in the wake of treatments for chronic disease and COVID-19, and this has driven an expansion in services for more than 131 million Americans.

Pharmacies, from smaller independent shops to large chain operations, have also continued to adopt technology in front of customers and behind the scenes. This includes incorporating AI solutions to improve efficiency and patient engagement. It’s an industry that could readily extend into spaces where pharmacies don’t usually go, without brick-and-mortar buildings or their digital counterparts that deliver to doorsteps.

The company said it expects to launch the first wide-scale rollout at the beginning of next year, around January.

Queue said it would use the funding to build product development, expand deployments with existing pharmacy customers and grow the company’s engineering team. The company currently employs 20 engineers in Silicon Valley and is actively hiring for robotics, hardware, software and pharmacy operations.

Image: Queue

Couchbase Inc. is trying to solve one of the hardest problems in enterprise artificial intelligence today: turning brittle, chat-style pilots into production-grade agents capable of remembering, reasoning and acting on live operational data. With the launch of its AI Data Plane, the company is betting that the real bottleneck for “agentic” AI isn’t the model — it’s the underlying data architecture.

Industry discussions about what’s holding back AI often focus on security, graphics processing unit availability and other infrastructure-level issues, but in most cases, data is what keeps chief information officers up at night. At the recent HPE Discover event, I asked Dallas Cowboys CIO Matt Messick about the challenge of bringing disparate data sets together, and he said it’s a major challenge and the thing he thinks about most right now – all day, every day, as he put it.

Couchbase’s announcement aims to address this headache with a scalable data platform.

Couchbase takes the covers off an AI data layer for agents