OpenAI and Broadcom this morning unveiled their first custom AI accelerator chip named “Jalapeño,” positioning it is as a purpose-built processor for large language model (LLM) inference, rather than the more general GPUs offered by the likes of Nvidia or AMD.
According to its creators, Jalapeño is designed to support workloads behind ChatGPT, Codex, the API and future agentic products, though notably, both OpenAI‘s and Broadcom’s news releases position it as a product that could be made available to external AI firms as well — “built from the ground up for current and future LLMs across the industry.” [Emphasis mine.]
Jalapeño’s engineering timeline set a blistering pace for the semiconductor industry, moving from early schematics to fabrication readiness within a brief nine-month window, when new processor development cycles are typically measured in years. Indeed, the OpenAI and Broadcom partnership itself was only publicly announced in October 2025.
The companies attributed this speed to a deep software-hardware co-development process that actively used OpenAI’s own models to accelerate parts of the chip design. Greg Brockman, OpenAI’s president and co-founder and Broadcom CEO Hock Tan appeared on CNBC this morning to discuss the news, and Brockman noted in the interview that the development process relied on prior generation OpenAI models, not even the cutting-edge GPT-5.5, though a company spokesperson declined to specify exactly which when asked by VentureBeta.
After receiving an early physical model on Wednesday, OpenAI outlined plans to begin rolling out these processors across active data centers by the end of this year. OpenAI says it has already begun testing running at least one of its prior generation models, GPT‑5.3‑Codex‑Spark, on the chips at a production workload, though in a test environment.
The release marks a major strategic expansion for the ChatGPT creator as it attempts to build the full computational stack required to make advanced AI faster, more reliable, and more accessible.
There remain, of course, many outstanding questions — including how the new Jalapeño chip performs compared to direct competitors, its costs, and its manufacturing viability. Sources close to the company said the initial performance itself was (ironically): “outstanding.”
On X, Brockman himself wrote that “Perf[ormance] per watt looking incredible.”
To understand why OpenAI is moving into chip design, it helps to look at the architecture. Jalapeño is an Application-Specific Integrated Circuit, or ASIC.
Unlike a GPU, which can handle many types of workloads, an ASIC is tuned for narrower uses, as industry experts note. That narrower focus can make it cheaper and more efficient for specific AI tasks, though less adaptable than Nvidia-style GPUs.
In Jalapeño’s case, OpenAI is starting from a clean design focused on modern LLM serving, instead of adapting a broader accelerator to fit its needs. The company says the architecture is shaped by its experience running large-scale AI products and is meant to reduce unnecessary data movement while better matching compute, memory and networking resources.
Broadcom is contributing core silicon implementation and networking technology, including Tomahawk networking silicon, while Celestica is helping with board, rack and system integration. The goal is to move the chip closer to its practical performance ceiling in real workloads, not just improve theoretical benchmarks.
However, OpenAI’s pivot into proprietary hardware is not just as a quest for technical supremacy: it may also make its core unit economics far more sustainable.
Audited financial documents posted recently by AI critic and AI public relations specialist Ed Zitron revealed that while OpenaAI generated an impressive $13.07 billion in revenue throughout 2025, its total operational expenses for the year ballooned to $34 billion, resulting in an operating loss of nearly $20.92 billion.
The primary culprit behind this cash hemorrhage involved pure compute requirements, though more is likely due to training than inference.
In 2025 alone, research and development costs—driven largely by the infrastructure required to train and serve massive language models—accounted for $19.18 billion, or approximately 56 percent of the company’s entire spending footprint. Furthermore, OpenAI reportedly paid Microsoft over $10.59 billion just for R&D and compute infrastructure last year.
Still, as OpenAI lays the groundwork for a heavily anticipated public offering in 2026, the Jalapeño inference chip may offer some reassurance to private investors and public markets that OpenAI has a plan for digging itself out of the financial hole and moving toward profitability. If it can drive down the costs of AI inference, then maybe it can recoup some of the losses spent on costly training runs.
“By designing more of the stack ourselves, we can serve more intelligence with greater efficiency and keep pushing advanced AI toward broader access,” said Brockman included in Broadcom’s release.
The introduction of Jalapeño immediately raises questions about OpenAI’s strategic positioning within the fiercely competitive semiconductor and GPU market.
Since kicking off the generative AI boom in late 2022, OpenAI has remained one of the largest customers of GPU market leader Nvidia’s premium products, but has also taken billions in investment dollars from the firm (engendering accusations of “circular dealing”), and expanded to work with other rival chipmakers to fuel its appetites.
Nvidia: In February 2026, Nvidia finalized a $30 billion direct investment into OpenAI as part of a massive $110 billion funding round.This deal secured an agreement to deploy 10 gigawatts of computing systems—including 3 gigawatts of dedicated inference capacity and 2 gigawatts of training capacity—utilizing Nvidia’s next-generation Vera Rubin platform. Sources close to the companies tell VentureBeat Nvidia will remain central to OpenAI, particularly on the model training and development side.
Amazon Web Services (AWS): As part of the same February 2026 funding round, Amazon invested $50 billion into OpenAI. This deal included a commitment for OpenAI to consume approximately two gigawatts of AWS’s proprietary Trainium computing capacity over the next eight years.
Advanced Micro Devices (AMD): OpenAI signed agreements with Nvidia’s chief hardware rival, AMD for the former’s usage of the latter’s AMD Instinct™ MI450 Series GPUs.
Cerebras: The company also struck a pact with Cerebras, an AI chipmaker that executed its initial public offering in May 2026.
Sources with knowledge of these deals said at present, they currently remain in place, unaltered.
Before the introduction of Jalapeño, OpenAI operated at a distinct structural disadvantage compared to the world’s vertically integrated technology empires.
Tech giants like Google and Amazon have for years utilized their own mature custom silicon programs— Google’s Tensor Processing Units (TPUs) and Amazon’s Trainium lines—to serve massive computational workloads at drastically lower margins.
Microsoft, OpenAI’s primary cloud provider and single biggest financial backer, aggressively entered the bespoke silicon market by launching the Azure Maia 100 accelerator in late 2023.
Microsoft subsequently escalated this effort in January 2026 by introducing the Maia 200, an inference powerhouse built on TSMC’s 3-nanometer process that already actively powers OpenAI’s GPT-5.2 models within Azure data centers.
Similarly, Meta has aggressively expanded its Meta Training and Inference Accelerator (MTIA) portfolio in recent years, debuting the MTIA 300, 400, 450, and 500 series to power its recommendation engines and generative artificial intelligence features without relying solely on Nvidia.
Jalapeño provides OpenAI with the opportunity to match and offset the hyperscaler advantage. By baking its software architecture directly into a proprietary processor, OpenAI has the chance to replicate, at least in part, the playbook used by Google, Amazon, Microsoft, and Meta — transitioning from a captive cloud customer into a more independent AI infrastructure provider.
The timing is ripe amid a rapidly escalating global silicon arms race. Driven in part by United States export restrictions, Chinese tech heavyweights are pursuing more of their own custom AI chip hardware, too:
In May, Alibaba’s semiconductor division, T-Head, unveiled the Zhenwu M890, a proprietary processor expressly engineered for autonomous AI agents that require massive memory bandwidth and long-running context windows.
Huawei is reportedly gearing up to release its new Ascend 950DT chip next month
ByteDance, the corporate parent of TikTok, reportedly entered active negotiations with Qualcomm in June 2026 to design custom application-specific integrated circuits for its data centers to escape third-party dependency.
By successfully finalizing the Jalapeño design, OpenAI is seeking to move beyond the traditional confines of a software laboratory and stand shoulder-to-shoulder with international cloud and infrastructure titans.
This sprawling web of vendor agreements highlights the sheer scale of OpenAI’s infrastructural ambitions. The ultimate goal of the OpenAI and Broadcom partnership involves deploying gigawatt-scale data centers with Microsoft and other partners beginning in 2026 — that is, data centers with compute requiring energy on the order of cities.
For Broadcom, the partnership acts as a massive reputational catalyst. The company has been among the biggest beneficiaries of the generative AI boom, helping hyperscalers and frontier labs engineer custom silicon.
Broadcom shares reflect this momentum, demonstrating an 18% year-over-year increase in the first part of 2026 and a nearly 7X boost since the end of 2022, according to CNBC.
Ultimately, Jalapeño confirms that OpenAI believes it is ready to move beyond software and code into the realm of real-world, custom hardware.
By controlling the physics of its inference pipeline—while simultaneously leveraging the capital and hardware of Nvidia, Amazon, AMD, and Cerebras—OpenAI is attempting to rapidly rewrite its future unit economics of AI.
Anthropic on Tuesday launched Claude Tag, a new product that embeds its most advanced AI model directly inside Slack as a persistent, shared teammate that anyone on a team can delegate work to by simply typing @Claude.
The product, available today in beta for Claude Enterprise and Team customers, replaces Anthropic’s existing Claude in Slack app and represents the company’s most aggressive move yet to colonize the enterprise collaboration layer — the place where decisions get made, work gets assigned, and institutional knowledge accumulates in real time.
For enterprise technology leaders who have spent the past two years evaluating where AI fits into their operational stack, Claude Tag reframes the question entirely. This is not a chatbot, a coding assistant, or a search tool bolted onto a messaging platform. It is an AI agent designed to function as a standing member of a team — one that builds memory, takes initiative, works asynchronously, and interacts with every person in a channel rather than serving a single user. The implications for enterprise workflow, governance, and vendor strategy are significant.
Anthropic says 65% of its own product team’s code is now created by its internal version of Claude Tag, and the company runs internal support and data insight channels through the same system. The claim is striking: Anthropic is asserting that the majority of its own product engineering output already flows through the tool it just put in customers’ hands.
At its core, Claude Tag works like this: an administrator pairs it with a Slack workspace, grants it access to specific tools and data sources, sets spending limits, and defines which channels it can operate in. From that point on, any team member in those channels can tag @Claude with a request — write a pull request, pull sales numbers, run a data analysis — and Claude will break the task into stages, execute them using the tools it has access to, and respond in a Slack thread with the result. The product runs on Claude Opus 4.8, the model Anthropic released less than a month ago.
Four capabilities differentiate Claude Tag from its predecessors and from competing integrations. First, it is multiplayer. Within a given Slack channel, there is one Claude that interacts with everyone, not a separate instance per user. Anyone can see what it is working on, and anyone can pick up the conversation where the last person left off. This is a direct contrast to most existing AI integrations in Slack, which tend to operate as single-player tools.
Second, it learns over time. As Claude follows along with its channel, it accumulates context about the work happening there. Users do not need to re-explain projects from scratch. If granted permission, Claude can also pull context from other Slack channels and data sources, though Anthropic says it will not report from private channels. Third, it takes initiative. With ambient behavior enabled, Claude will proactively surface relevant information from across the channels it monitors and the tools it is connected to, and will follow up on threads or tasks that have gone quiet without resolution. This is a notable expansion of agency: Claude is not just responding to requests but monitoring the information environment and deciding what its human teammates need to know. Fourth, it works asynchronously, pursuing projects autonomously over hours or days. Anthropic says its own teams “now spend much more of our time delegating tasks to many Claudes in parallel.”
Anthropic has designed the system with enterprise-grade isolation at its center. System administrators define separate Claude identities for different uses, scoped to specific channels with specific tools and data access. Everything, including Claude’s accumulated memories, stays within those boundaries. A Claude configured for sales work will not share memories or data access with one configured for engineering.
Administrators can set token-spend limits at both the organizational and channel level, and can review a complete log of every action Claude has taken and which user requested each task. For organizations managing compliance, audit, or regulatory requirements, this logging and scoping architecture is table stakes — and its absence has been a dealbreaker for many enterprises evaluating AI collaboration tools over the past year.
Migration from the existing Claude in Slack app requires an administrator opt-in within 30 days, and Anthropic says it is issuing introductory launch credits to eligible Enterprise and Team organizations. The four-step setup process — pair with Slack, connect tools, set spend limits, test in a private channel — is designed to reduce friction for IT teams already managing sprawling SaaS portfolios.
Claude Tag arrives in the middle of what has become the most fiercely contested territory in enterprise AI: the Slack channel. Slack itself has been aggressively positioning the platform as an “agentic operating system,” and the major AI players have responded by racing to plant their flags.
Salesforce, which acquired Slack for $27.7 billion in 2021, announced more than 30 new capabilities for Slackbot in March — the most sweeping overhaul of the platform since the acquisition — transforming it from a simple conversational assistant into a full-spectrum enterprise agent. OpenAI introduced “Workspace Agents” in April, allowing enterprise subscribers to design agents that take on work tasks across third-party apps including Slack, Google Drive, Microsoft apps, Salesforce, and Notion. Perplexity launched its enterprise “Computer” agent with direct Slack integration, letting employees query @computer directly inside Slack channels. Cognition’s Devin, the autonomous AI software engineer, has been built around Slack as a primary interface since its early days. Even Microsoft has brought GitHub Copilot into Teams.
The logic driving this convergence is straightforward: the average enterprise juggles over 1,000 applications, and employees waste countless hours on context switching, draining productivity by up to 40%. Whichever AI system becomes the default presence in the communication layer where work is coordinated gains an enormous distribution advantage — and, critically, an enormous data advantage. The AI that lives in the channel where work happens absorbs the institutional context that makes it increasingly difficult to replace.
To understand Claude Tag’s strategic significance, it helps to trace the product arc that led to it. Anthropic first integrated Claude with Slack in October 2025, offering two-way connectivity: users could invoke Claude from within Slack or connect Slack as a data source for Claude’s chatbot. The initial integration was focused on individual productivity — direct messages, AI assistant panels, and thread participation. In January 2026, Anthropic expanded Claude’s Slack presence when it launched interactive Claude apps, which included workplace tools like Slack, Canva, Figma, Box, and Clay.
In parallel, Anthropic was building out its enterprise infrastructure stack. In August 2025, the company bundled Claude Code into enterprise plans, a move its product lead Scott White called “the most requested feature from our business team and enterprise customers.” In April 2026, Anthropic launched Claude Managed Agents, a suite of composable APIs for building and deploying cloud-hosted AI agents at scale, with early adopters including Notion, Rakuten, Asana, and Sentry.
Then came Claude Opus 4.8 in late May, which Anthropic described as “a more effective collaborator” with “sharper judgement, more honesty about its progress, and the ability to work independently for longer than its predecessors.” Benchmark improvements included a jump in agentic coding scores from 64.3% to 69.2% and a knowledge work score increase from 1753 to 1890. Claude Tag is the synthesis of all of these threads — combining the Slack channel presence, the enterprise security architecture, the Managed Agents infrastructure, and the Opus 4.8 model’s improved agentic capabilities into a single product that Anthropic frames as “the beginning of an evolution of Claude Code.”
The financial stakes behind this launch are enormous. Anthropic raised $65 billion in Series H funding in late May at a $965 billion post-money valuation, and its run-rate revenue crossed $47 billion earlier this month. Claude Code’s run-rate revenue alone has grown to over $2.5 billion, more than doubling since the beginning of 2026, and enterprise use has grown to represent over half of all Claude Code revenue.
Those numbers explain why Anthropic is investing so heavily in channel-level presence. Every enterprise customer who grants Claude persistent access to a Slack channel — with connected tools, accumulated context, and ambient monitoring enabled — represents a dramatically deeper integration than a chatbot conversation or an API call. The usage patterns become stickier, the token consumption grows, and the switching costs rise. Deloitte’s deployment of Claude across more than 470,000 employees in 150 countries — reportedly its largest-ever enterprise AI deployment — illustrates the scale at which these dynamics play out.
The broader market trajectory reinforces the bet. Fortune Business Insights projects the global agentic AI market will grow from $9.14 billion in 2026 to $139 billion by 2034, and Gartner forecasts that 40% of enterprise applications will feature task-specific AI agents by 2026, up from less than 5% in 2025. Anthropic is not alone in seeing this future, but with Claude Tag it is making one of the most direct plays yet to own the enterprise agent layer.
Claude Tag raises several questions that enterprise buyers will need to evaluate carefully. The first is vendor dependency. As VentureBeat reported when analyzing Claude Managed Agents earlier this year, once an organization’s agents, operational configurations, and monitoring run on Anthropic’s managed infrastructure, switching costs increase significantly. Claude Tag deepens this dynamic: a Claude that has accumulated months of channel context and institutional memory becomes very difficult to replace. Enterprise procurement teams accustomed to negotiating multi-cloud flexibility will need to think hard about what it means to give a single vendor’s AI persistent access to the communication layer where institutional knowledge lives.
The second is governance around ambient monitoring. The proactive behavior mode — in which Claude monitors channels and surfaces information it decides is relevant — represents a meaningful expansion of what enterprise AI systems do. Organizations will need to develop clear frameworks for an AI agent that is not just responding to requests but actively surveilling information flows and making editorial judgments about what humans need to know. For regulated industries, this raises questions that existing AI governance policies may not yet address.
The third is pricing. Anthropic has not published detailed pricing for Claude Tag beyond noting that it runs on token-based spending with administrative controls. For an agent that monitors channels continuously, builds memory, and works asynchronously over hours or days, the token consumption profile could look very different from traditional AI usage. And the fourth is reliability: Anthropic has been candid in recent months about infrastructure strain caused by surging demand, and for a product positioned as an always-on team member, downtime carries a different kind of cost than it does for a tool invoked on demand.
Anthropic says its goal is to expand Claude Tag beyond Slack “so that teams can tag @Claude in the many other places they work.” The company is clearly eyeing the full collaboration surface — Microsoft Teams, email, project management tools, and beyond. If Claude Tag succeeds, it will validate a model of enterprise AI that looks less like a tool and more like a new category of worker: one that never sleeps, never forgets what was discussed in the channel last Tuesday, and never needs to be onboarded twice.
But the deeper significance of this launch may be what it reveals about the competitive dynamics reshaping enterprise software. For decades, the most valuable real estate in business technology was the system of record — the database, the CRM, the ERP. The current AI arms race suggests that the next era of enterprise value will be captured not by the system that stores the data, but by the agent that sits in the room where the work happens and understands what to do with it. Anthropic just gave that agent a name, a permanent seat in the channel, and permission to speak up when it thinks it has something to say. The question for every enterprise technology leader is no longer whether that agent will arrive. It is whether they are ready to manage it when it does.
Alibaba Cloud on Sunday released HappyHorse 1.1, a major upgrade to its AI video generation model that the company says delivers production-ready video synthesis across core content creation scenarios. The model is now live on Alibaba Cloud Model Studio with full API access for enterprise customers and developers, accompanied by a 40% sitewide launch discount for the first two weeks.
The release arrives at a moment of remarkable upheaval in the AI video generation market — and Alibaba appears keenly aware of the timing. OpenAI discontinued Sora after it proved financially unsustainable. ByteDance indefinitely shelved the international rollout of Seedance 2.0 following a barrage of copyright complaints from Hollywood studios. For enterprise procurement teams that had been evaluating or integrating those tools into marketing, advertising, and content production workflows, the competitive landscape has contracted sharply in a matter of months.
That contraction creates both an opportunity and a test for Alibaba. HappyHorse 1.1 is not a research demo or a consumer toy — it is an API-first product built for integration into enterprise software stacks, priced for volume, and backed by a $52.7 billion global infrastructure buildout. Whether it can convert technical capability into enterprise adoption, particularly in Western markets navigating intensifying U.S.-China tech tensions, will determine whether Alibaba can establish itself as a serious player in the generative video market that analysts expect to reach tens of billions of dollars by the end of the decade.
HappyHorse first appeared in early April as an anonymous submission on the Artificial Analysis Video Arena, an independent benchmarking platform where real users compare model outputs in blind, side-by-side evaluations. The model immediately claimed the top position in both text-to-video and image-to-video rankings. Alibaba was subsequently confirmed as the creator, revealing it was built by the company’s ATH (Alibaba Token Hub) AI Innovation Unit — a team previously part of the Future Life Lab under the Taobao and Tmall Group before a strategic organizational restructuring.
According to Arena.ai, HappyHorse 1.0 now holds the No. 2 position across all three Video Arena leaderboards. The platform noted the model scores 1,444 in both text-to-video and image-to-video categories, leading Google’s Veo-3.1 (with audio) by 69 points in text-to-video and xAI’s Grok-Imagine-Video by 23 points in image-to-video. In Elo-based ranking systems like Arena’s, models gain or lose points based on whether users prefer their outputs in head-to-head comparisons, meaning persistent double-digit leads reflect a consistent quality gap as perceived by human evaluators — not a statistical fluke.
The model’s architecture helps explain why. According to community-compiled technical documentation, HappyHorse is built around a 15-billion-parameter unified self-attention Transformer that processes text, image, video, and audio tokens within a single token sequence. Unlike many competitors that stitch together separate models for video and audio, HappyHorse operates as a unified system that handles all modalities in a single generation pass, eliminating the need for third-party dubbing or post-processing audio tools. For enterprise buyers evaluating total cost of ownership, that architectural simplicity translates directly into fewer integration points, fewer vendor dependencies, and faster time to production.
The 1.1 upgrade targets a set of pain points that enterprise video production teams know intimately. Alibaba Cloud described the release as “systematically optimized across core content generation scenarios,” and the specific improvements reveal a model that has been tuned for commercial deployment rather than viral social media demos.
The most consequential upgrade is multi-image reference capability, which Alibaba calls R2V (Reference-to-Video). The feature allows users to upload multiple character reference images and maintain consistent identity across generated video — directly addressing one of the hardest problems in AI video production, where subjects tend to drift in appearance between frames or shots. For brands producing advertising campaigns, product videos, or serialized marketing content, identity consistency is not a nice-to-have; it is a requirement that has historically forced teams back to traditional production methods.
Motion quality receives a significant overhaul, with what Alibaba describes as “strengthened motion modeling” that addresses prior limitations in speed and fluidity. The company also made targeted improvements to visual texture, specifically calling out the elimination of “facial oiliness,” “over-sharpening,” and “unnatural textures” — artifacts that have plagued commercial AI video since the technology emerged and that immediately signal to viewers that content is machine-generated.
Two additional upgrades round out the release. HappyHorse 1.1 improves audio-visual synchronization, including what Alibaba claims is “zero-drift lip sync” for dialogue scenes and context-aware speech pacing — building on the 1.0 version’s already notable ability to generate up to 15 seconds of 1080p video with synchronized audio output. The model also improves instruction-following for long and complex prompts, a critical differentiator for enterprise users who need to specify precise camera movements, lighting conditions, and narrative beats in a single generation pass rather than iterating through dozens of attempts.
The competitive context surrounding this launch is unusually favorable for Alibaba, and it is worth understanding why.
OpenAI’s Sora web and app experiences were discontinued on April 26, with the Sora API set to follow on September 24. The shutdown came after the product proved financially untenable: Sora cost roughly $1 million per day to operate but generated only about $2.1 million in total revenue, while active users dropped from a peak near 1 million to under 500,000. For enterprise teams that had integrated Sora into production pipelines, the abrupt withdrawal underscored the risks of depending on AI products that lack a sustainable business model — a cautionary tale that procurement officers are unlikely to forget quickly.
ByteDance’s Seedance 2.0, which many considered Sora’s most formidable successor, ran into a different kind of wall. Netflix, Warner Bros., Disney, Paramount, and Sony sent legal threats to ByteDance over allegations of systematic copyright infringement after users generated viral clips featuring Hollywood intellectual property. ByteDance indefinitely postponed the international launch, and the global rollout remains suspended.
That leaves Google’s Veo 3.1 as the primary Western competitor in the enterprise video generation space. But Alibaba’s Arena rankings suggest HappyHorse is outperforming Veo on user-perceived quality, and the 40% launch discount on Alibaba Cloud Model Studio could make HappyHorse significantly cheaper at scale. At the 1.0 level, pricing through third-party API platforms ran roughly $1.82 per 10-second clip at 720p and $3.12 at 1080p. With the promotional pricing, HappyHorse 1.1 could bring production-quality AI video generation within reach of mid-market companies and agencies that previously considered the technology too expensive for anything beyond experimentation.
HappyHorse 1.1 does not exist in isolation. It sits atop a global infrastructure offensive that distinguishes Alibaba from pure-play AI model companies that build impressive technology but lack the physical and commercial machinery to serve regulated enterprise customers at scale.
Just five days before the HappyHorse 1.1 launch, Alibaba Cloud opened its first data centers in France, establishing its third European hub after Germany and the United Kingdom. The Paris region features two availability zones, bringing the company’s global footprint to 105 availability zones across 32 regions. “The expansion of our cloud infrastructure into France reinforces our ongoing commitment to empowering European businesses with sovereign, secure, and intelligent solutions,” said Dr. Feifei Li, Alibaba Cloud’s CTO and president of international business, in the company’s announcement. In Japan, the company opened its fifth data center in Tokyo on June 19.
As reported by Data Center Dynamics, CEO Eddie Wu has committed to investing $52.7 billion in building a “unified global cloud network,” with the company later considering increasing this to $69 billion. This year alone, Alibaba has launched new regions in Mexico, Thailand, Malaysia’s Johor, and France. The France deployment is also part of Alibaba Cloud’s plan to roll out enterprise-grade agentic AI services across Europe in the second half of the year, including AgentRun (a development platform for AI agents), STAROps (an intelligent operations platform), and ACS Agent Sandbox (which provides hardware-level security isolation for agent workloads).
The infrastructure buildout serves a dual purpose for a product like HappyHorse. Running a 15-billion-parameter video generation model with integrated audio is extraordinarily compute-intensive, and having local infrastructure reduces latency for enterprise API calls while keeping customer data within regulatory boundaries. For European buyers operating under the European Commission’s new tech sovereignty framework — published June 3 with the explicit goal of protecting the bloc’s “digital independence” — the ability to run AI video generation workloads on locally hosted infrastructure is not a luxury. It is increasingly a compliance requirement.
Alibaba’s global push is unfolding under significant geopolitical headwinds that enterprise buyers cannot afford to ignore. The Pentagon added Alibaba, along with BYD and Baidu, to its list of Chinese military companies on June 8, preventing them from securing U.S. defense contracts. Alibaba rejected the designation, saying it is “not a Chinese military company nor part of any military-civil fusion strategy.”
The listing does not automatically trigger sanctions, and it does not directly restrict commercial transactions between private U.S. companies and Alibaba. But it adds a layer of reputational and regulatory complexity to procurement decisions, particularly for companies with U.S. government exposure, defense supply chain connections, or transatlantic operations. Enterprise technology purchases are rarely evaluated on technical merit alone — vendor risk assessments, board-level compliance reviews, and geopolitical scenario planning all factor into buying decisions for cloud infrastructure and AI tooling.
For European customers specifically, the calculus is layered in a different way. The continent’s growing emphasis on digital sovereignty cuts in two directions simultaneously: it creates demand for alternatives to the dominant U.S. hyperscalers (Amazon Web Services, Microsoft Azure, and Google Cloud control roughly 70 percent of European cloud infrastructure revenue, according to Synergy Research Group), but it also raises questions about whether a Chinese provider represents a meaningful improvement in strategic autonomy. Alibaba’s strategy of building sovereignty-compliant infrastructure in-market is a direct attempt to answer that question — but the Pentagon listing ensures it will be asked repeatedly.
The practical implications of HappyHorse 1.1 for enterprise teams are substantial. HappyHorse supports four modes of generation — text-to-video, image-to-video, subject-to-video, and the newly added video editing — covering the full spectrum of commercial video needs from ideation through production to post-production, all with integrated audio at no additional cost. That breadth of capability, delivered through a single API endpoint, simplifies what has historically been a fragmented and expensive production pipeline.
The question going forward is whether Alibaba can convert benchmark dominance and competitive timing into durable enterprise relationships. The company plans to release HappyHorse through Alibaba Cloud Model Studio with full enterprise SLAs, security certifications, and regional compliance — the table stakes that separate research breakthroughs from production-grade services. Watch for customer disclosures, usage metrics, and whether third-party platforms like fal.ai and Atlas Cloud (which already host HappyHorse 1.0) update to the 1.1 version quickly, which would signal genuine developer demand beyond Alibaba’s own ecosystem.
The AI video generation market entered 2026 with three credible enterprise contenders. One is dead. One is frozen. And the one still standing is a Chinese company backed by $52.7 billion in infrastructure spending, ranked No. 2 across every major independent benchmark, and offering a 40% discount to anyone willing to place the bet. In enterprise technology, the best product does not always win — but it rarely loses when the competition has already left the field.
When Anthropic quietly released Claude Design in April as a “research preview,” it generated the kind of instant traction most product teams dream about: more than one million users in its first week. It also generated a problem. The tool consumed tokens so voraciously that a PCWorld reviewer burned through 80 percent of his weekly Claude Pro allowance in roughly 25 minutes, producing just three variations of a single webpage prototype. “We’re talking another token-hungry Claude product here,” the reviewer wrote, “one that Pro users in particular will barely be able to use before burning through their usage limits.”
Two months later, Anthropic is shipping a substantially overhauled version of Claude Design that attempts to fix the consumption issue while simultaneously repositioning the product from a flashy demo into something far more strategically important: a design system compliance layer that connects to code, connects to the tools enterprises already use, and — critically — keeps everything on brand.
The update, announced Wednesday, arrives at a moment when Anthropic is executing one of the most aggressive product expansions in the AI industry’s brief history. In the past ten weeks alone, the company has launched Claude Opus 4.8, released (and then suspended) the Mythos-class Fable 5 model, shipped ten agent templates for financial services, announced a multi-year alliance with DXC Technology to embed Claude inside the IT infrastructure of the world’s largest banks and airlines, rolled out Claude for Small Business with integrations into QuickBooks and PayPal, and published research showing that Claude Code users now average 20 hours per week on the tool.
Claude Design’s transformation from prototype toy to enterprise platform is the latest move in a company-wide strategy to make Claude not just an assistant people talk to, but a worker embedded in the systems where work actually happens.
The headline feature in Wednesday’s update is not the new drag-and-resize editor, nor the expanded list of export destinations, though both matter. The feature that signals where Anthropic is heading is the rebuilt design system import.
Users can now bring one or several design systems into Claude Design from a GitHub repository, design files, or raw uploads. Once imported, Claude builds with those components, checks its output against the design system, and auto-corrects before the user ever sees the result. For larger organizations, a new admin role can approve a single standard system and lock down edits, ensuring that every asset Claude produces conforms to company guidelines.
This is a meaningful departure from the tool’s original positioning. In April, Claude Design was a blank canvas: give it a prompt, and it would generate something visually impressive but stylistically arbitrary. Business Insider tested it against Canva AI for a photography workshop slide deck and found that Claude Design “anticipated my needs” and “identified its own errors and corrected them without prompting.” But the output reflected Claude’s aesthetic judgment, not the user’s brand. For an individual freelancer or a startup founder sketching ideas, that was fine. For a 10,000-person enterprise with a 200-page brand standards document, it was a non-starter.
The design system import changes that equation. By ingesting a company’s actual components — its buttons, typography, color tokens, spacing rules — and then validating output against them before surfacing results, Claude Design is attempting something that most human designers struggle with: consistent brand compliance at speed and scale. The admin lockdown feature, which prevents individual users from overriding the approved system, is a direct play for the enterprise procurement conversation, where “can we control what it produces?” is often the first question.
The second major update is the bidirectional integration between Claude Design and Claude Code. Users can now run /design-sync in Claude Code to import their local codebase’s design system into Claude Design, ensuring that prototypes start from real components rather than approximations. When a design is ready to ship, it hands off to Claude Code, which picks up exactly where the designer left off — no screenshot, no rebuild. The integration works in reverse, too. From a Claude Code terminal, the /design command lets developers create, edit, and sync design projects without leaving their workflow.
This matters because the handoff between design and engineering has been one of the most persistent friction points in software development for decades. Tools like Figma’s Dev Mode and Zeplin have tried to bridge the gap by generating specifications and code snippets from design files, but the translation has always been lossy. A designer’s prototype and an engineer’s implementation inevitably diverge, creating a cycle of visual QA, redlines, and “that’s not what the mockup looked like” conversations.
Anthropic is betting that if the same AI system both designs and codes — and if both modes share the same underlying component library — the gap disappears. It is, in effect, arguing that the design-to-code problem was never really about better specification formats or smarter handoff tools. It was about the fact that two different humans (or two different tools) were interpreting the same intent. A single AI system that operates on both sides of the workflow doesn’t need to interpret; it just continues.
The timing of this integration is also significant in light of Anthropic’s own research. Just yesterday, the company published an analysis of roughly 400,000 Claude Code sessions showing that domain expertise — not coding proficiency — is the primary driver of successful outcomes. Every major occupation succeeded at coding tasks at nearly the same rate as software engineers. If designers can now move fluidly between visual prototyping and code implementation through a single AI system, the research suggests they will succeed not because they learned to code, but because they deeply understand the design problems they are solving.
The token consumption issue that dogged Claude Design’s launch was not just a user experience annoyance — it was a structural threat to the product’s viability. If a $20-per-month Pro subscriber could exhaust their entire weekly allowance in a single 30-minute session, the tool was effectively inaccessible to the individual users and small teams who drove its initial viral adoption.
Anthropic’s response is twofold. First, Claude Design now shares usage limits with chat, Claude Cowork, and Claude Code, rather than drawing from a separate, smaller pool. This gives most users significantly more headroom. Second, the company says it has reduced the average token consumption per turn while maintaining output quality, and that error rates have dropped sharply.
Whether this is enough remains an open question. The fundamental tension is architectural: generative design is inherently token-expensive. Every variation Claude produces requires the model to reason about layout, typography, color, spacing, responsiveness, and content simultaneously, then generate a complete, functional artifact. That is a fundamentally different workload than answering a question in chat, and it consumes tokens accordingly. Anthropic’s efficiency improvements may push the breaking point further out, but they do not eliminate the underlying economics. For enterprise customers on Team and Enterprise plans with higher limits, this may be a non-issue. For Pro subscribers, the math is still likely to be tight.
The new editor helps mitigate this somewhat by giving users direct control over individual elements — drag, resize, and align — without burning a model turn for every small adjustment. Hundreds of stability fixes also mean fewer wasted turns on errors and regenerations, which were a significant source of token drain in the original release. These are not glamorous improvements, but they are the kind of grind work that separates a research preview from a daily-use tool.
The update’s third pillar is an expanded set of export destinations. Claude Design now sends work to Adobe, Base44, Canva, Gamma, Lovable, Miro, Replit, Vercel, and Wix, in addition to PDF and PowerPoint. The breadth of this list reveals a deliberate positioning strategy: Anthropic is building Claude Design not as a place where work is finished, but as the place where it begins.
The partner quotes tell the story. Replit’s president Michele Catasta frames the integration as meeting “builders wherever ideas begin.” Canva’s Anwar Haneef describes the flow from Claude Design as turning “a first draft” into “a finished asset — kept on-brand, personalized for the moment.” Vercel’s Andrew Qu talks about pushing a concept “straight to Vercel to ship.” In each case, Claude Design is the origin point, and the partner tool is where polish, collaboration, and deployment happen.
This hub-and-spoke model also serves as a defensive moat against the open-source alternative that has emerged with surprising speed. Open Design, a community-built project tracked by Augment Code, reached 57,400 GitHub stars and 310 contributors in just eight weeks after Claude Design’s launch. It offers local-first operation, model flexibility supporting 16 different coding agents, and 259 skills with 142 design systems — all without cloud lock-in. Augment Code’s Paula Hingel noted that for “teams that need to self-host, use their own API keys, or swap models, Open Design is currently the only local-first option with this level of skill and design system coverage.”
Anthropic’s answer to this competitive pressure is not to match Open Design on self-hosting or model flexibility — those are philosophical concessions the company is unlikely to make. Instead, it is building an integration ecosystem that open-source projects cannot easily replicate. A native Adobe Express connector, a verified Canva export pipeline, a first-party Vercel deployment path — these are partnerships, not features, and they require business relationships that community projects cannot forge at the same pace.
To understand why Claude Design’s evolution matters, it helps to zoom out. Anthropic is building a product surface that now spans creative work (Design), code (Code), knowledge work (Cowork), and enterprise operations (Managed Agents) — all unified by the same underlying models and, increasingly, by shared context that carries across tools.
The trajectory of the past quarter makes the pattern unmistakable. In May, Anthropic launched Claude for Small Business with connectors to QuickBooks, PayPal, and HubSpot, putting Claude inside the tools that small business owners already use for payroll, invoicing, and marketing. The same month, the company released ten agent templates for financial services covering everything from pitchbook creation to KYC screening, with connectors to FactSet, S&P Capital IQ, and Morningstar. Claude Opus 4.8 shipped on May 28 with a “dynamic workflows” feature enabling hundreds of parallel sub-agents in a single Claude Code session. Then came the Fable 5 and Mythos 5 launch on June 9, followed almost immediately by a US government export control directive that suspended access to both. DXC Technology announced a multi-year alliance to train tens of thousands of Claude-certified engineers to embed Claude inside the systems it operates for major banks, airlines, and insurers.
The design system you import into Claude Design is the same component library that Claude Code uses to implement. The financial model you build in Claude for Excel can flow into a pitchbook created in Claude Design and exported to PowerPoint. The brand assets a small business owner creates through Claude Design can be pushed directly to Canva for team collaboration. This is not a chatbot strategy. It is a platform strategy, and the Claude Design update — with its design system imports, code round-trips, and export ecosystem — is one of the clearest expressions of it yet.
Anthropic also published an engineering deep-dive last month detailing how it contains Claude across products using sandboxes, virtual machines, and egress controls — infrastructure that becomes more critical as tools like Claude Design gain access to proprietary design systems and brand assets. The containment architecture reveals both the ambition and the risk: the more deeply Claude embeds into enterprise workflows, the higher the stakes when something goes wrong, and the more sophisticated the security envelope must become.
Three questions will determine whether Wednesday’s update delivers on its ambitions. First, whether the token economics actually work for the broadest user base — shared limits and efficiency gains help, but generative design remains expensive. Second, whether the design system import proves robust enough for real enterprise use, because ingesting a GitHub repository of React components and faithfully using them across dozens of design variations is a genuinely hard technical problem. And third, whether the Claude Code round-trip actually eliminates the design-engineering gap or merely shifts it.
Claude Design launched two months ago as a thing people tried once and marveled at. Anthropic is now trying to make it a thing people use every day — and more than that, a thing their entire team trusts to stay on brand while they do. In the AI industry, the distance between a viral demo and an indispensable tool has swallowed more products than it has produced. Anthropic just bet that design systems, not just design prompts, are the bridge across.
On Sunday, a team of nine researchers at Sina Weibo — the Chinese social media giant better known for its microblogging platform than for cutting-edge artificial intelligence — quietly posted a 14-page technical report to arXiv that sent shockwaves through the AI research community. Their claim: a language model with just 3 billion parameters can match or exceed the reasoning performance of flagship systems from Google DeepMind, OpenAI, Anthropic, and DeepSeek that are hundreds of times larger.
The model, called VibeThinker-3B, scored 94.3 on AIME 2026 — the American Invitational Mathematics Examination, one of the most demanding standardized math competitions in the world. That figure places it alongside DeepSeek V3.2, a model with 671 billion parameters, and ahead of Gemini 3 Pro, Google’s high-performance flagship reasoning system, which scored 91.7. With a test-time scaling technique the team calls Claim-Level Reliability Assessment, the score climbs to 97.1, edging past virtually every system in the public record.
Within hours of publication, the paper had drawn 62 upvotes on Hugging Face’s daily papers feed, the model repository had accumulated 130 likes, and the GitHub repository had reached 685 stars. But the reaction on social media was not uniformly celebratory. It was, in many cases, deeply skeptical.
“WHAT THE HELL is happening in AI?” wrote the user @orcus108 on X, in a post that accumulated over 161,000 views. “A 3B parameter model just put up coding benchmark scores in the same league as Claude Opus 4.5… I genuinely don’t know if this is a breakthrough or if the benchmarks are broken.”
That tension — between genuine scientific advancement and the growing suspicion that AI benchmarks have become gameable to the point of meaninglessness — sits at the heart of the VibeThinker-3B story. And the answer matters enormously, not just for academic bragging rights, but for the multibillion-dollar question of whether the AI industry’s relentless push toward ever-larger models is the only path to intelligence.
The results reported in the technical report are, by any conventional standard, extraordinary.
On the mathematics side, VibeThinker-3B achieved 91.4 on AIME 2025, 94.3 on AIME 2026, 89.3 on HMMT 2025 (the Harvard-MIT Mathematics Tournament), 93.8 on BruMO 2025 (the Brown University Math Olympiad), and 76.4 on IMO-AnswerBench, a benchmark comprising 400 problems at the level of the International Mathematical Olympiad. In coding, it posted an 80.2 Pass@1 on LiveCodeBench v6, a benchmark designed to test executable code generation, and achieved a 96.1 percent acceptance rate on unseen LeetCode weekly and biweekly contests from late April through late May 2026. On instruction following, it scored 93.4 on IFEval.
To put the parameter disparity in perspective: DeepSeek V3.2 has 671 billion parameters — roughly 224 times the size of VibeThinker-3B. GLM-5, from Zhipu AI, has 744 billion parameters. Kimi K2.5, from Moonshot AI, exceeds 1 trillion. VibeThinker-3B’s 3 billion parameters could run on a consumer laptop.
The researchers frame this result not as an anomaly but as evidence for a broader theoretical claim. They introduce what they call the “Parametric Compression-Coverage Hypothesis,” which argues that different types of AI capability have fundamentally different relationships to model size. Verifiable reasoning — the kind tested by math competitions and coding challenges, where answers can be definitively checked — is what the paper calls a “parameter-dense” capability: one that can be compressed into a compact core. Open-domain knowledge, by contrast, is “parameter-expansive,” requiring broad coverage across facts, concepts, and edge cases that inherently demands more parameters.
The paper acknowledges this distinction directly. On GPQA-Diamond, a graduate-level science knowledge benchmark, VibeThinker-3B scored just 70.2 — well behind the 91.9 achieved by Gemini 3 Pro and the 87.0 scored by Claude Opus 4.5. The authors write that this gap “is consistent with our claim rather than a contradiction to it: the main finding is not that a 3B model has fully replaced leading general-purpose models, but that a small model can reach first-tier performance on many verifiable reasoning tasks.”
VibeThinker-3B is not built from scratch. It is post-trained on top of Qwen2.5-Coder-3B, a compact foundation model from Alibaba’s Qwen team, through what the Weibo AI researchers call the “Spectrum-to-Signal Principle” — a multi-stage pipeline first introduced in the team’s earlier VibeThinker-1.5B work in November 2025.
The training unfolds in four major phases. The first is a two-stage supervised fine-tuning process that uses curriculum learning: the model first trains on a broad mixture of math, code, STEM reasoning, general dialogue, and instruction-following data, then shifts to a curated subset of harder, longer-horizon reasoning problems. In the second stage, samples with reasoning traces shorter than 5,000 tokens are discarded, and problems that VibeThinker-1.5B can solve more than 75 percent of the time are filtered out, forcing the model to focus on genuinely difficult challenges.
The second phase applies reinforcement learning across multiple domains — mathematics, code, and STEM — using the team’s MaxEnt-Guided Policy Optimization algorithm, or MGPO, which prioritizes training on problems at the model’s current capability boundary rather than problems it already solves easily or finds impossible. Notably, the team found that a strategy that worked well at the 1.5B scale — progressively expanding the context window during RL training — actually hurt performance at 3B. They hypothesize that the stronger starting checkpoint meant that truncating reasoning traces during warm-up was no longer removing noise but disrupting valid reasoning patterns. The solution was to train with a single 64,000-token context window throughout.
Within the math RL phase, the team also introduces what it calls “Long2Short Math RL,” a secondary optimization stage that redistributes rewards to favor shorter correct solutions over longer ones, reducing verbosity without sacrificing accuracy. The technique uses a zero-sum reward redistribution that avoids biasing the overall reward signal while nudging the model toward more efficient reasoning.
The third phase extracts high-quality reasoning trajectories from the RL-trained checkpoints and distills them back into a unified model through supervised fine-tuning. The team uses a “learning-potential score” — essentially the student model’s perplexity on each teacher trajectory — to prioritize traces that are correct but that the student has not yet internalized. The final phase, called Instruct RL, applies reinforcement learning on instruction-following tasks using a combination of rule-based validators for format constraints and rubric-based reward models for open-ended quality assessment.
Francesco Bertolotti, an AI researcher who flagged the paper early on X, described the approach succinctly: “These results were achieved primarily through post-training refinements on Qwen2.5-Coder. The paper doesn’t provide many details, but it appears they distill from RL ckpts and then do a final RL-based instruct RL.” His post drew over 161,000 views.
For every enthusiastic reaction, the paper drew an equally forceful objection. The AI research community in mid-2026 has grown deeply wary of benchmark-driven claims, and VibeThinker-3B arrived in an environment primed for suspicion.
“The benchmarks are literal pattern matching single file coding,” wrote @BigMoonKR on X. “It has no relation to actual coding work. I don’t know how people still don’t get this.”
“Benchmaxxing,” declared @oflu_bedirhan, using a term that has become shorthand in the AI community for models that appear optimized specifically for benchmark performance at the expense of real-world utility.
The most pointed criticism came from users who actually downloaded and tested the model. “Just tried the full precision,” wrote @politilols. “It doesn’t even know what a uv script (so the most popular Python dev tool) is. Haven’t seen that in a single LLM in at least a year now. Benchmaxxed.” When Bertolotti responded that the model seemed more focused on mathematical reasoning than practical coding, the user countered: “They include a livecodebench score. Zero chance that is reflective of the model.”
@Itsdotdev raised a structural criticism: “Look into the benchmarks themselves and it probably won’t be so shocking. Why no DeepSWE? Why none of the standard benchmarks SOTA providers use?” The user @AvenirReym posed a more diagnostic question: “If it holds on a benchmark made after the model’s training cutoff, it’s real. If it only wins on AIME-style sets that have been circulating for years, it’s leakage.”
The paper’s authors appear to have anticipated these objections. The technical report states that training sets “have undergone strict benchmark decontamination,” including n-gram-based filtering to remove “n-gram overlaps with evaluation sets.”
The LeetCode contest evaluation — which covers contests from April 25 to May 31, 2026, dates that postdate any plausible training data cutoff — represents the most robust guard against data contamination concerns. On those contests, VibeThinker-3B passed 123 out of 128 first-attempt submissions, a 96.1 percent rate that exceeded GPT-5.2, Doubao Seed 2.0 Pro, Kimi K2.5, and Claude Opus 4.6 under identical evaluation conditions.
Still, real-world user reports suggest a significant gap between benchmark performance and practical utility — a phenomenon that has become familiar across the industry. “In LM Studio it only responds well to first question, next questions reply to the first question,” reported @luismolinaab.
Even the sharpest critics acknowledged that achieving these benchmark numbers at 3 billion parameters — regardless of how transferable they are to production use cases — is a meaningful engineering achievement. “Even if it’s benchmaxxing doing so with 3B parameters is fascinating, goes to show how fast this field is progressing,” wrote @rohityin.
The observation cuts to a question that has consumed the AI industry since the advent of the scaling hypothesis: Is bigger always better? The conventional wisdom, articulated most famously in the Chinchilla scaling laws and reinforced by the commercial dominance of ever-larger foundation models, holds that more parameters and more training data reliably yield better performance. The economic corollary is stark: training and deploying frontier models costs tens or hundreds of millions of dollars, creating enormous barriers to entry.
VibeThinker-3B challenges that consensus — but only partially. The paper is careful to draw a boundary around its claims, distinguishing between tasks with “clear verification signals” and those that require broad factual knowledge. The Parametric Compression-Coverage Hypothesis explicitly argues that small models cannot replace large ones across the board.
“The true significance of VibeThinker-3B does not lie in proving that a 3B model can replace large-scale generalists,” the paper states, “but rather in providing a concrete empirical signal: the development of compact models is no longer merely a passive compromise for deployment efficiency or cost control; it emerges as a promising research trajectory that is fundamentally complementary to the traditional parameter scaling paradigm.”
Perhaps the most surprising element of the work is its provenance. Sina Weibo — publicly traded on Nasdaq and Hong Kong, with a market capitalization that fluctuates in the single-digit billions — is not a company typically associated with frontier AI research. Yet the VibeThinker series is Weibo’s second major open-source AI contribution in seven months.
VibeThinker-1.5B, released in November 2025, demonstrated that a model with just 1.5 billion parameters could outperform the original DeepSeek R1 on several math benchmarks — a result the team achieved for what it claimed was a post-training cost of just $7,800, compared to the $294,000 estimated for DeepSeek R1.
The research team is compact — nine authors, all listed as Sina Weibo Inc. employees. The model is released under the MIT License, one of the most permissive open-source licenses available, and the weights are freely downloadable from both Hugging Face and ModelScope. Within the first day of release, community members had already created GGUF quantizations and derivative models.
The most honest assessment of VibeThinker-3B may be that it is simultaneously less and more than what the benchmarks suggest. Less, because a model that struggles with basic knowledge of popular developer tools is unlikely to replace any production-grade coding assistant anytime soon. More, because the underlying insight — that reasoning ability and factual knowledge are partially decoupled, and that the former can be compressed far more aggressively than previously assumed — has profound implications for how the industry thinks about model design, deployment economics, and the accessibility of advanced AI capabilities.
If the Parametric Compression-Coverage Hypothesis holds, it suggests a future in which small, specialized reasoning engines operate alongside large knowledge-rich models in hybrid architectures — a vision where a 3-billion-parameter model handles the logical heavy lifting while a larger system supplies the factual grounding. Such an architecture could dramatically reduce the cost of deploying AI reasoning capabilities, potentially bringing competition-level mathematical and coding performance to devices with modest hardware.
“The interesting part is that we’re starting to separate knowledge from reasoning,” wrote @RealLambdaFlux on X. “A small model with strong post-training can punch way above its size on tasks with clear feedback.”
@cmitsakis suggested the practical endgame: “I think small models are the future for agents because they can use tools to get the knowledge and they can run fast and cheap.”
Whether that future arrives through VibeThinker-3B specifically, or through the dozens of teams now racing to reproduce and extend these results, the paper has already accomplished something that no benchmark score can fully capture.
It has forced the AI community to confront an uncomfortable possibility: that for years, the industry may have been spending billions of dollars scaling up parameters to improve a kind of intelligence that could have fit, all along, on a laptop. The weights are public. The code is open. And the most important test isn’t on any leaderboard — it’s whether anyone can make a model this small actually useful in the real world.
For three years, Microsoft’s artificial intelligence story has been inseparable from OpenAI. The partnership — cemented by a cumulative investment exceeding $13 billion — gave Microsoft early access to the most advanced AI models on the planet, catapulting its Copilot products into the enterprise mainstream and adding hundreds of billions of dollars to its market capitalization. To the outside world, Microsoft’s AI strategy was OpenAI.
Mustafa Suleyman wants to change that narrative.
In an exclusive sit-down interview with VentureBeat at Microsoft Build 2026, the CEO of Microsoft AI disclosed that a contractual change with OpenAI roughly six months ago granted his division the formal authority to pursue what he openly calls “superintelligence” — using Microsoft’s own researchers, its own data pipelines, and its own custom silicon.
“We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence,” Suleyman said. “So this is very early days.”
The comment, delivered matter-of-factly backstage at the Fort Mason Center here, offers the clearest signal yet of a strategic inflection point unfolding inside the world’s most valuable public company. Microsoft is not abandoning OpenAI. But it is building something alongside it — and, eventually, something that could stand entirely on its own.
The most tangible evidence of that shift arrived the same day. Microsoft announced a family of seven new AI models developed entirely in-house by its AI Superintelligence Team, spanning reasoning, code generation, image creation, transcription, and voice synthesis. The models — branded under the “MAI” family name — are Microsoft’s most ambitious first-party AI release to date.
The flagship, MAI-Thinking-1, is a 35-billion-active-parameter reasoning model that Microsoft says matches leading models in its weight class on key software engineering benchmarks and demonstrates advanced mathematical reasoning. Suleyman emphasized one point repeatedly: the model was trained from scratch on clean, commercially licensed data, without distillation from third-party frontier models — a direct, if unstated, contrast to the widespread industry practice of using outputs from competitors’ systems to train cheaper alternatives.
“We train our reasoning models from scratch,” Suleyman wrote in a blog post accompanying the announcement. “We don’t distill from other labs and we don’t rely on unlicensed or opaque data.”
The rest of the family fills out a multimodal portfolio designed for enterprise deployment: MAI-Code-1-Flash, a lightweight coding model built specifically for GitHub Copilot and VS Code; MAI-Image-2.5, which supports both text-to-image and image editing; MAI-Transcribe-1.5, which Microsoft claims is the most accurate transcription model available, operating across 43 languages; and MAI-Voice-2, a multilingual speech-generation system. All of the models ship through Microsoft Foundry, the company’s model-hosting and deployment infrastructure, and for the first time, developers can tune model weights themselves through third-party platforms including OpenRouter, Fireworks, and Baseten.
But Suleyman made clear in the interview that the seven models are a proof of concept, not a finished product. The real project is the lab itself.
“Our job is to make sure that when we look out to 2030 and beyond, we have the capacity not just to buy models from third parties, but to build the absolute frontier, the best models in the world,” he said. “That’s a long transition.”
To understand what Suleyman means by “set free,” you need to understand the unusual contractual architecture that has governed Microsoft’s AI efforts for years.
When Microsoft invested billions into OpenAI beginning in 2019, the partnership came with a specific arrangement: OpenAI would build the frontier models, and Microsoft would serve as the exclusive cloud provider, integrating those models into its products and reselling them through Azure. The deal gave Microsoft extraordinary commercial leverage — access to the world’s most advanced AI without having to build it — but it also created a dependency. Microsoft was explicitly barred from pursuing its own AGI research, and the agreement even capped how large a model the company could train, restricting it from building systems beyond a certain computing threshold measured in FLOPS.
That arrangement was formally renegotiated. As Fortune and Axios reported in November, a revised deal with OpenAI removed those restrictions, clearing the way for Suleyman to launch the MAI Superintelligence Team and pursue what he calls “humanist superintelligence.” The result, in Suleyman’s telling at the time, was a “best-of-both environment, where we’re free to pursue our own superintelligence and also work closely with them.”
By the time he sat down with VentureBeat at Build 2026, roughly six months had passed since that self-sufficiency effort formally began. Microsoft had already started shipping in-house models — including MAI-Image-2-Efficient, a lighter-weight image generation model released in April — but the seven MAI models announced at Build are the team’s most ambitious release yet: a full multimodal family spanning reasoning, code, image generation, transcription, and voice.
Even so, Suleyman does not view the shift as a rupture with OpenAI. He described Microsoft’s current position as one of abundance, not scarcity.
“There’s no immediate urgent need to fill a gap in three months’ time or six months’ time,” he said. “We have OpenAI, we have Anthropic, we have thousands of models inside Foundry. So there’s already a huge amount of optionality available to us.”
The framing is telling. Microsoft’s push into first-party frontier models is not born out of a crisis in the OpenAI relationship but out of a strategic calculation: as AI becomes the most consequential technology layer in enterprise computing, the company cannot afford to depend entirely on partners for the foundational capability. “Over the next five years, we have to be able to produce state-of-the-art frontier-scale models,” Suleyman said. “That’s our mission.”
If the seven MAI models represent the technical ambition, a new capability called Frontier Tuning represents the commercial logic. Announced alongside the models at Build, Frontier Tuning allows enterprise customers to customize MAI models using their own proprietary data, workflows, and domain terminology, all within their own secure compliance boundary. The system uses reinforcement learning environments — what Microsoft calls “training gyms for AI” — that let agents learn directly from real workplace tasks without affecting production systems.
The results Microsoft shared are striking. An MAI model tuned for Excel reportedly matches GPT 5.4 performance while operating at up to ten times greater efficiency. Early enterprise adopters are seeing similar gains: when tuned for one unnamed organization’s exacting standards, the MAI model achieved the highest win rate of any model tested at roughly one-tenth the cost.
Suleyman framed Frontier Tuning as part of a broader evolutionary stage — a move from intelligence to action. “We’ve basically moved beyond just conversation,” he told VentureBeat. “Now we’re moving to action.”
He introduced a new framework for thinking about that progression: the shift from IQ (factual intelligence) to EQ (emotional intelligence, or the ability to follow tone and style instructions) to what he calls AQ — the “Actions Quotient.”
Future AI agents, in Suleyman’s telling, won’t just answer questions. They will log into enterprise software, navigate complex multi-application workflows, and execute tasks across Excel, Word, Teams, Jira, Adobe InDesign, and customer relationship management systems — just as a human employee would.
“You should be able to show up on day one and almost provision credentials to a new AI agent,” he said. “The model needs to be able to move across all of these different environments, and that’s actually the great strength of Microsoft.”
The Build 2026 announcements bore this out in concrete product terms. Microsoft Scout, the company’s first “Autopilot” agent, operates as an always-on background assistant built on the open-source OpenClaw technology. It runs with its own governed identity inside Microsoft Entra, so its actions are auditable and attributable. Windows 365 for Agents gives AI agents their own managed Cloud PCs, allowing them to interact directly with applications and browsers inside enterprise environments. And the Foundry platform received major updates — including hosted agents with sub-100-millisecond cold starts, a new Microsoft Agent Framework, and one-click publishing to Teams and Microsoft 365 Copilot.
Suleyman also articulated why he believes Microsoft’s position is uniquely defensible — and the argument has less to do with model architecture than with where work actually happens.
“We’ve sort of hoovered up all of the obvious pools of training data,” he said, referring to the industry’s early scramble to ingest the open web. “In the next phase, we actually want to be able to give these agents to companies to train on their specific tasks with the data that they have inside of their own big workflows.”
The claim is subtle but consequential. The first wave of generative AI was trained on publicly available text — books, websites, Reddit posts, code repositories. That data is now largely exhausted, and its use is increasingly contested in court.
The next wave, Suleyman argues, will be trained on enterprise-specific data: the internal workflows, decision traces, and institutional knowledge that define how real organizations operate. Microsoft, which serves 493 of the Fortune 500 through Azure according to Suleyman, is already embedded inside those workflows through Microsoft 365, Teams, Dynamics 365, and the broader Azure ecosystem. Frontier Tuning is the mechanism that converts that positional advantage into model performance.
“People underappreciate that that’s going to be the next domain,” Suleyman said.
The early partner list for Frontier Tuning reflects the ambition: Mayo Clinic, where Microsoft is co-creating a frontier AI model for healthcare using de-identified clinical data; EY, which is tuning a tax-advisory agent for deployment to 75,000 professionals globally; Land O’Lakes, where Frontier Tuning delivered what the company’s product development scientist called “meaningful improvements in grounded outputs and style compliance”; and Pearson, which is using tuned models to provide learning-science-aligned feedback in its Communication Coach product.
The Mayo Clinic partnership may be the most significant. Microsoft and Mayo Clinic are collaborating to build a healthcare-specific frontier model that combines Mayo’s clinical expertise and longitudinal patient insights with Microsoft’s AI capabilities. The model will be owned by Mayo Clinic and deployed first within Mayo’s own environment before being made available to other organizations through Foundry.
None of this works without an industrial-scale compute infrastructure, and Suleyman was unusually candid about the hardware economics underlying Microsoft’s strategy.
“We are the largest buyer of GPUs on the planet,” he said. “We’re the largest buyer of GB200s and GB300s in the world.”
Microsoft will continue purchasing Nvidia accelerators “for many, many years to come,” Suleyman said. But the company is simultaneously building its own custom silicon. Maia 200, Microsoft’s second-generation AI accelerator, is already running in production across data centers in Iowa and Arizona, with deployments planned for Italy, Australia, and South Korea. According to Microsoft, Maia 200 delivers the best tokens-per-dollar-per-watt in the company’s fleet.
Suleyman put a finer point on the economics in the interview: Maia 200 is 30 percent more cost-efficient than Nvidia’s GB200, he said. And when Microsoft co-optimizes its own MAI models to run natively on Maia silicon, the company sees an additional 1.4x improvement in performance per watt. “It is going to be cheaper in years to come to build on MAI models with Maia 200 and Maia 300 inside of Azure,” he said.
That claim — if it holds at scale — has profound implications for the competitive landscape. It means Microsoft is not merely buying its way to AI dominance through Nvidia; it is building a vertically integrated stack in which its own models, running on its own chips, inside its own cloud, tuned on its customers’ own data, could offer performance and cost characteristics that no competitor can replicate.
Suleyman also pushed back sharply against one of the most popular narratives in Silicon Valley: that AI models are rapidly commoditizing.
“A lot of people are saying models are commoditizing,” he said. “I don’t think that’s true.”
His argument hinges on what he calls “quality tokens” — the proposition that the composition, curation, licensing, and deduplication of training data matter at least as much as raw scale. Microsoft’s new MAI models, he said, were trained on a pre-training mix composed of approximately 50 percent high-quality code, with the remainder drawn from commercially licensed and carefully curated sources.
The result, he argued, is a distinct “lineage” of models optimized for coding, reasoning, and agentic behavior — fundamentally different from models optimized for consumer chat, cultural content, or multilingual breadth.
“We’re going to see very distinct lineages that reflect different training objectives of different companies,” he said. “Quality tokens matter more than just brute-force scale.”
This is a strategically important argument for Microsoft to make. If models are commodities — if any lab can match the frontier within months using cheaper compute and distilled training data — then the model layer becomes a race to the bottom, and Microsoft’s billions in compute investment offer no durable advantage. But if model quality is a function of data discipline, research depth, and institutional patience, then the lab-building approach Suleyman is pursuing becomes a genuine competitive moat.
He used a specific metaphor to describe that approach, one borrowed from optimization theory: the “hill-climbing machine.” The phrase describes a system that continuously improves — cycle after cycle — by applying more compute, better data, and sharper evaluation. “The goal here is to build what we think of as a hill-climbing machine,” he wrote in his blog post. “An organization that can continuously improve, cycle after cycle.” The metaphor is revealing because it describes a process, not a destination. Suleyman is not promising that Microsoft will build the world’s best model next quarter. He is arguing that Microsoft is building the system — the research culture, the data pipelines, the silicon co-optimization, the evaluation infrastructure — that will produce progressively better models over years.
The strategic picture that emerges from Suleyman’s comments — and from the full scope of the Build 2026 announcements — is of a company preparing for a future in which AI capability is not rented from a partner but generated internally, at scale, across every layer of the stack.
Microsoft still needs OpenAI. The partnership continues to power Copilot, Azure AI services, and ChatGPT’s infrastructure. Suleyman acknowledged as much, describing Microsoft’s portfolio of model providers as a source of strength, not a problem to be solved.
But the direction of travel is unmistakable. With its own frontier models, its own custom silicon, its own reinforcement learning environments for enterprise tuning, and its own autonomous agent infrastructure, Microsoft is constructing a parallel path — one that, by 2030, could make the company a fully self-sufficient frontier AI lab embedded inside the world’s largest enterprise software platform.
“Our ultimate goal is what we call Humanist Superintelligence,” Suleyman wrote in his blog post. “That means advanced AI systems designed to serve people and organizations, not replace them.”
Whether that goal is achievable — or even clearly definable — remains one of the great open questions in technology. And Suleyman expressed more confidence than caution when asked about the trajectory of progress. “I really think we’re at the tip of the iceberg,” he said. “The models are so much more powerful than we know how to extract intelligence from them.”
But confidence and execution are different things. Building a frontier lab is not an announcement; it is a decade-long commitment that requires retaining elite researchers, maintaining scientific rigor under commercial pressure, and producing results that justify the staggering capital expenditure.
Google learned this with DeepMind — which Suleyman himself co-founded in 2010, before joining Microsoft — and even that lab, widely regarded as one of the best in the world, spent years navigating the tension between pure research and product delivery.
Suleyman seemed aware of the contradiction. “If you rush it, you’ll screw it up,” he said.
The sticker on his laptop reads: “Patience and urgency.” It is a paradox that Microsoft now has five years — and several hundred billion dollars — to resolve.
Perplexity AI, the fast-growing search startup now valued at $20 billion, unveiled what it calls the first hybrid local-server inference orchestrator at Computex 2026 on Monday night, demonstrating software that autonomously decides — in real time and mid-task — which AI workloads stay on a user’s device and which get routed to frontier models in the cloud.
CEO Aravind Srinivas demonstrated the system onstage alongside Intel CEO Lip-Bu Tan during Intel’s keynote address, using Perplexity’s “Personal Computer” agent to process confidential deal materials. In the demonstration, local models running on Intel Core Ultra Series 3 determined which information should remain on the device and which information could be sent to cloud-based models. Srinivas said the approach balances intelligence, accuracy, privacy, and cost.
The key claim is not that a model can run locally — dozens of tools already do that. It is that Perplexity’s system makes the routing decision itself, task by task, without requiring the user to choose in advance. Sensitive data like financial records or health information stays on the local machine; the heavier reasoning tasks that require frontier-scale models get sent to the cloud. One task, multiple execution locations, automatic orchestration.
“No product has done this before,” a Perplexity spokesperson said in an email to VentureBeat. The product is not yet available to users; according to the company, the hybrid inference feature will launch in the coming weeks.
To understand why the Computex demonstration matters, it helps to trace the product arc Perplexity has been building since early this year.
On February 25, Perplexity launched Computer, a multi-model AI agent that orchestrates 19 different AI models to complete complex, long-running tasks on behalf of users. The system ran entirely in the cloud, breaking goals into subtasks and routing each to whichever model — Claude, Gemini, GPT, Grok, or others — was best suited for the job. Perplexity Computer unified every current AI capability into a single system, functioning as a general-purpose digital worker that operates the same interfaces a user does.
Then, in March, Perplexity introduced Personal Computer at its inaugural Ask 2026 developer conference. That product launched as a new Mac app with support for a hybrid local-cloud AI agent, which Perplexity described as a “personal orchestrator” that hybridizes local and server environments for security and productivity. Personal Computer could access the Mac’s file system and native Mac apps to create and execute entire workflows, with files created in a secure sandbox and all actions auditable and reversible.
What Srinivas demonstrated at Computex extends this architecture in a fundamental way. Previously, even the Personal Computer product divided labor along relatively clear lines: local file access on the device, heavy computation on Perplexity’s servers.
The new hybrid inference orchestrator gives the system itself the ability to reason about where each piece of a task should execute — not just which model to use, but which physical location should process it. The system reportedly asks for user permission before sending sensitive tasks to the cloud, a design choice that addresses one of the central anxieties enterprises have about agentic AI: data governance.
The timing of the demonstration is not coincidental. Computex 2026 has been dominated by a single theme: on-device AI. Just hours before the Intel keynote, Nvidia CEO Jensen Huang unveiled the RTX Spark, a new Arm-based superchip that the company positions as the foundation for a new generation of AI-native Windows PCs.
At full strength, the RTX Spark Superchip offers up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, 128GB of LPDDR5X RAM, and up to 300 GB/s of memory bandwidth — enough power and memory for AI agents and 120-billion-parameter models with context lengths stretching to a million tokens. RTX Spark systems will begin arriving in the fall.
Intel, not to be outdone, used its keynote to showcase Xeon 6+ processors with 288 efficiency cores built on 18A technology for the data center, and positioned its Core Ultra Series 3 as the client silicon that makes hybrid inference possible on the PC.
Perplexity’s hybrid orchestrator sits at the intersection of both strategies. If the system performs as advertised, it creates a direct economic incentive for users — and eventually enterprises — to invest in more powerful local silicon. The more capable the on-device chip, the more inference can run locally, reducing cloud costs and improving latency for sensitive workloads. That dynamic benefits Nvidia, Intel, and every other chipmaker competing for AI PC sockets.
The implications extend well beyond chip economics. “As chips become more powerful, more intelligence moves onto a person’s machine, alongside server inference for the complex tasks that still need frontier models,” a Perplexity spokesperson told VentureBeat. “Sensitive and sovereign work can stay local, which changes the need for massive country-level infrastructure.”
That last claim — about sovereign infrastructure — is the most provocative. Nations from the UAE to France to India have been investing billions in domestic AI compute capacity partly on the assumption that sensitive data must stay within their borders, which means building or buying access to local data centers. If meaningful inference can run on an end user’s device with no data leaving the machine, the calculus changes. It does not eliminate the need for data centers, but it could soften the urgency of the buildout.
Perplexity’s hybrid inference play rests on the same architectural bet the company has been making all year: that the orchestration layer matters more than any individual model. For AI engineers, this signals a fundamental shift — the orchestration layer may matter more than the models themselves.
The key insight is separation of concerns: the orchestration layer handles task decomposition, state management, and tool coordination, while the model layer handles specific computations. This decoupling means teams can swap models as better alternatives emerge without redesigning the entire system.
Perplexity has leaned heavily into this philosophy. The company is doubling down on packaging frontier models in a consumer-friendly user experience, arguing that there is value in orchestrating multiple third-party LLMs to obtain the most cost-effective and accurate answers to queries. Models, in Perplexity’s view, are specializing, not commoditizing.
The hybrid inference extension takes that logic one step further. Perplexity is now orchestrating not just across models but across physical compute locations — choosing which model runs where. A lightweight local model might handle a privacy-sensitive document summarization task while a frontier cloud model tackles the complex reasoning required to analyze that summary against a broader market landscape. The orchestrator manages the handoff.
This is a technically ambitious claim. Making it work reliably in production will require the orchestrator to accurately assess the complexity of each subtask, understand the sensitivity of the data involved, know the capabilities and latency characteristics of whatever local hardware the user has, and manage the state of a task that may be bouncing between environments mid-execution.
It is easy to imagine edge cases where the routing logic fails, sends something sensitive to the cloud, or degrades performance by assigning a task to an underpowered local model. Perplexity says the system will be chip-agnostic, though the initial Computex demo ran on Intel silicon. The company expressed enthusiasm in its communications about the new AI chips announced at Computex this week, suggesting it intends to optimize across vendors.
The hybrid inference announcement arrives at a complicated moment for Perplexity. The company has been on a remarkable growth trajectory: It secured $200 million in new capital at a $20 billion valuation, just two months after raising $100 million at an $18 billion valuation. Since its founding three years ago, the rapidly growing AI company has raised $1.5 billion in total funding, according to PitchBook data.
But the company also faces a mounting stack of legal challenges. Nine organizations have filed active suits against Perplexity for alleged copyright and trademark infringement as of May 31, 2026: CNN, the New York Times, News Corp and Dow Jones, the New York Post, the Chicago Tribune, Encyclopedia Britannica, Merriam-Webster, Reddit, and Japan’s Yomiuri Shimbun. The CNN lawsuit, filed just days ago on May 28, is the most recent, accusing Perplexity of scraping more than 17,000 CNN stories, photos, videos, and other content and using that material to train its products. Perplexity has responded with a consistent message. “You can’t copyright facts,” the company’s chief communications officer Jesse Dwyer said in a statement.
Other publishers have opted for partnership over litigation. Time, Gannett, Le Monde, and Der Spiegel have signed licensing arrangements with Perplexity. The company launched a Publishers Program in mid-2024 in which participating outlets receive a share of revenue generated when their content is cited in Perplexity answers.
According to CNBC, Perplexity’s chief business officer Dmitry Shevelenko confirmed at the time that the flat rate was a double-digit percentage but declined to share specifics. As TechCrunch reported in December 2024, additional publishers including the LA Times, Adweek, The Independent, and Lee Enterprises subsequently joined the program, though not without internal controversy — reporters at some outlets told TechCrunch they were not informed of the deals before they were announced publicly.
The legal risk is not existential, but it is material, and with enterprises increasingly evaluating Perplexity’s tools for sensitive workflows — precisely the use case the hybrid inference system is designed to serve — unresolved intellectual property questions could dampen adoption.
The hybrid inference demo should be read alongside Perplexity’s broader push into enterprise software, a transformation that accelerated dramatically this year. At the Ask 2026 developer conference in March, VentureBeat reported that Perplexity announced Computer for Enterprise, positioning the three-year-old startup as a direct competitor to Microsoft, Salesforce, and the legacy enterprise software stack.
Beyond Computer’s existing 100-plus integrations, enterprise customers gained access to business-grade connectors for Snowflake, Datadog, Salesforce, SharePoint, and HubSpot, with administrators able to install custom connectors via the Model Context Protocol. The package also includes purpose-built workflow templates for legal contract review, finance audit support, sales call preparation, and customer support ticket triage, alongside SOC 2 Type II certification and the option for zero data retention.
Hybrid inference deepens this enterprise pitch considerably. For regulated industries — financial services, healthcare, defense, legal — the ability to keep sensitive data on a local device while still accessing the reasoning power of frontier cloud models is not a nice-to-have. It is a potential compliance requirement.
An investment bank parsing confidential deal documents, for instance, might be unable to send those materials to a third-party cloud under existing data handling agreements. A system that can run the sensitive parsing locally while routing non-sensitive analytical tasks to the cloud offers a middle path. IDC forecasts a tenfold increase in agent usage and a thousandfold growth in inference demands by 2027, and security and governance rank as the top evaluation factor for enterprise agentic platforms, according to a CrewAI survey. Hybrid inference speaks directly to that priority.
Several questions will determine whether Perplexity’s Computex demonstration becomes a landmark product or a compelling prototype.
The actual performance characteristics remain untested outside a controlled stage environment — how the routing logic handles varied hardware configurations, unreliable network connections, and ambiguous data sensitivity classifications is an open question.
The competitive response matters too: Google, Microsoft, Apple, and OpenAI are all building their own local-cloud AI architectures. Apple Intelligence already routes some tasks locally and some to Private Cloud Compute servers, Google’s Gemini Nano runs on-device, and Microsoft’s Copilot+ PCs are designed around local inference capabilities. None of these systems, however, currently offer the kind of dynamic, autonomous task-level routing Perplexity claims.
Even if the technology works as demonstrated, there is the question of whether the business can keep pace with the ambition. At a $20 billion valuation with approximately $200 million in annual recurring revenue, Perplexity trades at roughly 100x revenue, a premium requiring aggressive growth to justify. Management’s $656 million 2026 revenue target implies 230% growth, creating significant execution pressure.
Perplexity has built its business on a bet that the future belongs not to any single model but to the system that orchestrates all of them. At Computex, it extended that bet from the software layer to the physical layer — from which model to which machine. In the AI industry’s relentless race to build bigger data centers and train larger models, Perplexity just argued that the most important computer in the stack might be the one already sitting on your desk.
For the past two years, the technology industry has raced to make AI agents more capable — teaching them to write code, navigate software interfaces, manage files, and orchestrate multi-step workflows with increasing autonomy. What the industry has not done, at least not with any consistency, is answer the question that keeps chief information security officers awake at night: what happens when an agent goes wrong?
On Tuesday at its annual Build developer conference, Microsoft offered what may become the definitive answer. The company introduced Microsoft Execution Containers, or MXC — a policy-driven execution layer, built into the Windows operating system itself, that lets developers and IT administrators declare exactly what an AI agent can and cannot access, with those boundaries enforced at runtime by the OS kernel.
The announcement, buried within a sweeping set of developer-focused updates, is arguably the most consequential platform move Microsoft made at Build this year, and it has the potential to reshape how every enterprise on Earth thinks about deploying autonomous AI software.
MXC is not a product you buy. It is an SDK and a policy model — a foundational primitive embedded in Windows and the Windows Subsystem for Linux — that provides what Microsoft calls a “composable sandbox spectrum.” That spectrum ranges from lightweight process isolation, already adopted by GitHub Copilot’s command-line interface, all the way up to micro-virtual machines, Linux containers, and full cloud instances running on Windows 365.
The system separates an agent’s execution from the user’s desktop, clipboard, user interface, and input devices. Critically, it binds every agent to a strong identity — either a local ID or a cloud-provisioned identity backed by Microsoft Entra — so that every action the agent takes can be attributed, audited, and governed.
The implications are enormous. Until now, the enterprise deployment of AI agents has been stuck in a paradox: the more autonomous and useful an agent becomes, the more dangerous it is to let it operate on a corporate network without guardrails. MXC is Microsoft’s attempt to break that paradox — not by making agents less capable, but by making the environment they operate in fundamentally more controlled.
To understand why MXC matters, consider what an AI agent actually does when it runs on your computer. Unlike a traditional application, which operates within well-understood boundaries — a word processor reads and writes documents, a browser fetches web pages — an AI agent is, by design, unpredictable. It receives a goal in natural language, reasons about how to achieve it, and then takes actions: opening files, executing code, calling APIs, browsing the web, interacting with other software. Each of those interactions creates what security professionals call “attack surface.”
Microsoft’s own blog post framed the challenge in stark terms. The company wrote that “as agents become more capable and autonomous, they’re delivering material productivity gains. But they’re also introducing new risk, and the issue isn’t just the agent. It’s the entire system the agent operates across.” Every interaction between agents and humans, tools, applications, models, and other agents “exposes new attack surface and introduces different failure modes.” Microsoft characterized this as “a multi-layer systems problem.”
This is not a theoretical concern. In the months leading up to Build, security researchers demonstrated numerous ways that AI agents could be manipulated — through prompt injection, through malicious tool calls, through data exfiltration disguised as normal workflow. For enterprises that handle sensitive data, proprietary models, and regulated information, the absence of a trusted execution environment has been the single biggest barrier to moving agents from demo to deployment.
MXC operates on a deceptively simple principle: declare what the agent can do before it runs, and let the operating system enforce those declarations at runtime. A developer or an IT administrator writes a policy that specifies which files, directories, and network resources an agent is allowed to access. MXC then creates a contained execution environment — a sandbox — that enforces those boundaries regardless of what the agent attempts to do.
What makes MXC unusual, and potentially very powerful, is the breadth of its isolation options. Microsoft designed the system so that a single SDK and policy model can map to the appropriate isolation construct for any given workload. For a lightweight coding assistant that just needs to read the current project directory, fast process isolation may be sufficient. For an autonomous agent that executes arbitrary code downloaded from the internet, a full micro-VM may be required. The system is designed to be “dynamically composable based on intent and risk,” meaning that the level of isolation can be adjusted based on what the agent is actually doing, not just what category it falls into.
Session isolation is a particularly important feature. MXC separates the agent’s execution from the user’s desktop, clipboard, UI, and input devices. This directly mitigates several classes of attacks that security researchers have identified as particularly dangerous for AI agents: UI spoofing, where an agent manipulates what the user sees to trick them into approving a malicious action; input injection, where an agent sends keystrokes or mouse clicks to other applications; and cross-session data leakage, where information from one user’s session bleeds into another.
During a pre-briefing with VentureBeat the night before the announcement, a Microsoft developer offered a vivid demonstration of the technology in action. He had set up the open-source agent framework OpenClaw running inside MXC’s sandbox on his personal development machine. He then instructed the agent to delete all the files on his desktop. The agent attempted to comply — but the sandbox prevented it. “If you look at my desktop here, you see how clean my desktop is,” the developer said during the demo. “That’s a lie.” The files, he explained, were completely safe because “the container won’t allow it.”
The demonstration went further, showcasing the granularity of MXC’s controls. Users can mark specific files as read-only for the agent, restrict access to the browser and screen capture, control whether the agent can see location data, and have all of those permissions managed centrally by an enterprise IT department through Intune policies. The agent operates inside what is effectively a one-way mirror: it can do the work it has been asked to do, but it cannot see or touch anything outside the boundaries that its policy defines.
Pavan Davuluri, Microsoft’s Executive Vice President for Windows and Devices, underscored during the pre-briefing that the primitives MXC introduces — security, containment, isolation, and user control — are essential to making AI agents commercially viable.
He emphasized that these capabilities are “not unique to OpenClaw” and that “this pattern repeats itself over and over” for any agent running on a Windows device. The primitives that exist in the operating system now “for the file around security, containment, isolating them, having users in control,” he said, are what will make agents safe enough for ordinary consumers and corporate deployments alike.
For corporate IT departments, the most significant element of the MXC announcement is not the SDK itself but its integration with Microsoft’s existing enterprise security stack through what the company calls Agent 365. Arriving in preview in July, Agent 365 layers Microsoft’s Entra identity service and Intune device management platform on top of MXC, so that IT administrators can govern agent containment centrally while developers choose the level of isolation their workload demands.
The integration goes further: Microsoft Defender will provide runtime threat protection, Entra will handle identity and access management, Intune will enforce device-level policies, and Microsoft Purview will extend its data governance and compliance capabilities to agent activity. This means that an enterprise could, in theory, allow employees to run AI agents on their corporate machines — even powerful, autonomous agents that execute code and manage files — while maintaining the same kind of centralized visibility and control that IT departments currently have over traditional applications.
Microsoft described the identity layer in its official blog: “Windows assigns agents a local ID or a cloud provisioned identity backed by Entra and attributes all activity from the container to that identity, so you can clearly differentiate human from agent.” For regulated industries — financial services, healthcare, government — the ability to produce an audit trail that distinguishes between human actions and agent actions on the same machine could prove to be a regulatory requirement, not merely a nice-to-have feature. Every agent action attributable to a specific identity, every containment boundary enforceable through the same policy infrastructure that already governs hundreds of millions of Windows devices — this is the architecture that could finally move AI agents from pilot programs to production.
Platform announcements at developer conferences are often aspirational. What distinguishes the MXC launch is the breadth and specificity of the partners already building on it. Microsoft named five: OpenAI, Nvidia, Manus, Nous Research (maker of the Hermes agent), and the OpenClaw open-source project. Each is integrating MXC in a distinct way that illuminates a different use case for the technology.
OpenAI’s involvement is particularly striking. David Wiesen, a member of OpenAI’s technical staff, said that “working with Microsoft on the Microsoft Execution Containers (MXC) allows us to explore new patterns for AI agents to safely and efficiently generate and execute code.” He added that by combining Codex’s capabilities with MXC’s execution environment, the goal is “to help developers move from intent to reliable execution faster, while maintaining the security and control enterprises need.” The reference to Codex — OpenAI’s code-generation agent — suggests that MXC could become the default execution environment for one of the most widely anticipated agent products in the industry.
Nvidia is bringing its OpenShell framework to Windows built on MXC, providing what Microsoft described as “an easy-to-deploy package for autonomous, always-on agents safely.” Manus, the Chinese-born AI agent startup that gained viral attention earlier this year, is also integrating. Tao Zhang, Manus’s Chief Product Officer, said that MXC “gives developers a policy-driven way to define what an agent can access and enforce those boundaries at runtime, so more autonomous agents can operate safely in enterprise environments.” And Dillon Rolnick, the CEO of Nous Research, offered what may be the most concise articulation of why MXC matters: “Continuously-running local agents, like Hermes Agent, require intentional isolation. Developers need control over what an agent can access and trust that those controls will hold.”
One of the more revealing stories behind the MXC announcement involves OpenClaw. During the press pre-briefing, a Microsoft developer described how the partnership came together organically — Peter Steinberger, OpenClaw’s creator, sent him a direct message in January expressing interest in collaborating. What began as a casual conversation evolved into a full-fledged platform partnership, with Microsoft developers contributing to the OpenClaw Windows companion app, built as a native WinUI application rather than a wrapped web app.
The OpenClaw integration serves as what Scott called “the ultimate test app for all the stuff that [the Windows platform team] is making.” If OpenClaw — which by its nature gives agents broad autonomy to execute tasks on a user’s machine — can run securely within MXC’s containment boundaries, then the containment system is robust enough for any agent. Scott explained the philosophy driving the work: “Think of OpenClaw Windows as the ultimate test app… If OpenClaw can succeed on Windows, that means that the Linux support is there, the container support is there, the containment is there.”
The companion app demonstrates the full spectrum of MXC’s enterprise controls — file permissions, network access, screen capture restrictions, location data — all manageable centrally through Intune policies. Microsoft donated the project to OpenClaw and plans to continue contributing to it as open source. As one member of the Windows leadership team put it during the briefing: “All agents, all comers, everyone is welcome on Windows… It’s going to run great on Windows, because the primitives are there. The base of the pyramid is solid.”
MXC arrives at a moment when the technology industry is grappling with a fundamental tension. AI agents represent what may be the most significant new category of software since mobile applications, and every major technology company is racing to build them. But the security and governance infrastructure required to deploy these agents responsibly in enterprise environments barely exists. Microsoft’s approach is distinctive because it locates the trust layer at the operating system level rather than in the agent framework, the model provider, or a third-party security product.
This is a deliberate architectural choice. By building containment into Windows itself, Microsoft ensures that the security guarantees hold regardless of which agent, which model, or which framework a developer chooses.
It also means that the hundreds of millions of Windows devices already managed through Intune and secured through Defender can, in principle, become agent-ready through a software update rather than a rip-and-replace deployment.
Apple’s approach to AI agents leans heavily on its walled-garden ecosystem, offering security through restriction — limiting which agents can run and what they can do. Google’s approach, centered on its cloud infrastructure, offers security through centralization. Microsoft’s approach offers security through declaration and enforcement — allowing any agent to run, but containing its impact through OS-level policy.
For enterprises that operate in heterogeneous environments with diverse toolchains and multiple AI providers, the Microsoft model may prove the most practical. The competitive dynamics are already shifting: with OpenAI’s Codex, Nvidia’s OpenShell, and independent agent frameworks like Manus and Hermes all building on MXC, Microsoft is positioning Windows not just as the platform where agents run, but as the platform where agents can be trusted to run.
MXC is available now in early preview, meaning developers can begin building against the SDK and testing containment policies. The Agent 365 integration with Defender, Entra, Intune, and Purview is scheduled for preview in July — a timeline aggressive enough to suggest that much of the engineering work is already done, but far enough out to allow for refinement based on developer feedback.
The real test, however, will come when enterprises begin deploying agents at scale on production networks. Containment is only as good as the policies that govern it, and writing effective agent policies for complex enterprise environments will be an entirely new discipline — one that IT departments have not yet developed and that no vendor has yet figured out how to teach. The technology is promising, but an empty sandbox is just an empty box. Filling it with the right rules, for the right agents, in the right contexts, will require a level of organizational sophistication that most companies are only beginning to contemplate.
Still, the significance of what Microsoft announced on Tuesday is difficult to overstate. For the first time, a major operating system vendor has proposed a comprehensive, kernel-level answer to the question of how autonomous AI software should be contained, identified, and governed on the devices where most of the world’s work actually gets done. The industry spent two years teaching agents to act. Microsoft is now betting that the bigger business — and the harder engineering problem — is teaching the operating system to watch.
Microsoft on Monday unveiled the Surface RTX Spark Dev Box, a compact desktop computer designed to let software developers run large AI models on their desks instead of paying for cloud computing — a move that directly challenges the per-token pricing model that has defined the AI industry’s economics since ChatGPT launched three and a half years ago.
The device, announced at Microsoft Build 2026, packs Nvidia’s new Blackwell-architecture RTX Spark processor and 128 gigabytes of unified memory into a small-form-factor chassis, delivering what Nvidia rates at one petaflop of AI compute. In practical terms, that means a developer can load, run and interact with AI models exceeding 120 billion parameters without sending a single API call to the cloud.
“These class of devices, we think, will get to about 100 billion parameter model running,” Pavan Davuluri, Microsoft’s executive vice president of Windows and Devices, said during a press briefing ahead of the event. He emphasized that raw model size is only part of the equation: “The model size is one thing, but for the model to be effective, it kind of needs to be able to have enough context, because a larger model, you feed it larger context.” At 100,000 tokens of context, he noted, the key-value cache alone can consume 40 to 50 gigabytes of memory — which is precisely why Microsoft and Nvidia engineered the device around a 128-gigabyte unified memory pool shared dynamically between the CPU and GPU.
The machine will be available later this year in the United States, sold exclusively through Microsoft.com. The company did not disclose pricing.
The Surface RTX Spark Dev Box arrives at a moment when the economics of AI development have become a boardroom-level concern. Companies large and small are grappling with cloud GPU bills that scale unpredictably: every fine-tuning run, every inference call, every agentic workflow that loops through a frontier model accumulates cost. For a developer iterating rapidly on a prototype — running the same model dozens or hundreds of times a day — those charges compound fast.
Microsoft is framing the Dev Box as a release valve for that pressure. Andrew Hill, corporate vice president of Surface, wrote in the announcement blog post that the device “changes that equation” by letting developers “reserve frontier model calls for truly frontier problems and handle the rest on their own hardware.” The pitch is not that cloud computing is obsolete, but that much of the work currently being sent to remote data centers does not require state-of-the-art models and would be better served by capable local hardware with predictable, fixed costs.
This is a significant strategic shift for Microsoft, a company that derives tens of billions of dollars in annual revenue from Azure cloud services. By selling hardware that explicitly reduces customers’ cloud dependency, Microsoft is acknowledging a tension that has been building across the industry: the marginal cost of AI inference at scale is unsustainable for many teams, and the market is demanding alternatives. The bet appears to be that developers who prototype locally will still deploy to Azure when they need to scale — and that owning both ends of that workflow is more valuable than owning only the cloud.
The technical architecture of the Dev Box reflects a set of deliberate engineering choices aimed at sustained, not peak, performance — a distinction that matters enormously for AI workloads that can run for hours.
At the center is Nvidia’s RTX Spark system-on-chip, which combines an ultra-efficient ARM-based CPU with a Blackwell-generation RTX GPU. In a traditional Windows PC, Davuluri explained during the briefing, this configuration would require four separate components: a CPU, a discrete GPU, dedicated graphics memory and system RAM. The RTX Spark collapses all of that into a single chip paired with a single unified memory pool.
That unification is the critical design decision. Conventional gaming laptops with high-end Nvidia GPUs top out at roughly 24 gigabytes of GPU-accessible memory. The Dev Box’s 128 gigabytes of unified memory — accessible to both the CPU and GPU through what Nvidia calls its Unified Memory Access architecture — is what makes it possible to load models that would otherwise require cloud GPU instances with specialty high-bandwidth memory configurations.
Microsoft did substantial work at the operating system level to exploit this architecture. The company implemented new memory management logic in Windows that raises the ceiling on how much system memory the GPU can address, introduces smarter page-size allocation for shared memory regions and ensures that heavy GPU workloads do not starve the CPU of the resources it needs for multitasking. The Windows scheduler was also optimized for RTX Spark’s heterogeneous core layout, routing demanding workloads to performance cores while keeping efficiency cores available for background tasks.
The thermal design is equally deliberate. The Dev Box operates within an approximately 100-watt sustained thermal envelope — modest by desktop standards, but meaningful for a device intended to run training jobs and inference workloads continuously. The aluminum chassis itself is engineered to function as a passive heatsink, and the method Microsoft used to build it is among the most striking details about the machine.
The top panel is manufactured using metal 3D printing, a process that enables internal geometries too complex for conventional CNC machining or injection molding. The perforations are not simple through-holes; they are angled in multiple directions around the internal fan to optimize airflow from cold-air intake through heat dissipation. During the press briefing, Harry, a Surface industrial designer, explained the rationale: “The complexity is something other manufacturers wouldn’t be able to do, like CNC, or like any molding, because of the complexity of shape.”
When asked whether 3D printing would constrain mass production, the designer acknowledged the challenge but suggested Microsoft had developed a process robust enough to scale. The result is a machine that runs quietly enough for an open office while sustaining the kind of continuous GPU workloads that would throttle most conventional desktops of similar size. For a device that Microsoft expects developers to leave running overnight on fine-tuning jobs, quiet sustained performance is not a luxury — it is a requirement.
Microsoft is shipping the Dev Box with Windows 11 Pro pre-configured at the image level for development work — a detail that sounds minor but reflects a growing recognition that the out-of-box experience for developer hardware has historically been poor.
The machine boots into a dark theme with a simplified taskbar, widgets removed and Do Not Disturb enabled. Developer Mode is turned on. PowerShell 7 is the default shell. WSL 2 — the Windows Subsystem for Linux — comes pre-installed with GPU passthrough and CUDA support already configured. Visual Studio Code, GitHub Copilot, Git, Python and Node.js are all installed and ready.
“We’ve said, ‘Hey, you know what, we got you, you want to go fast,'” a Microsoft engineer who demonstrated the configuration during the briefing told VentureBeat. The philosophy, he explained, is that developers were going to install all of these tools anyway — the friction was in the hours of setup and configuration that stood between unboxing a machine and writing the first line of code.
The Dev Box also ships with integration points across Microsoft’s AI stack: AI Toolkit for VS Code for model conversion and fine-tuning, Windows ML and Windows Copilot Runtime for local inference, and Microsoft Foundry for connecting local prototypes to cloud deployment pipelines. For enterprises, the device integrates with Entra ID and Intune for identity and device management, and includes Secured-core PC architecture, BitLocker encryption and Microsoft Defender.
The most obvious competitive comparison is Apple’s Mac Mini, which has dominated the compact-desktop category and has been widely adopted by developers drawn to Apple Silicon’s unified memory architecture and power efficiency.
Davuluri addressed the comparison directly during the briefing, saying the Dev Box is “in a different class of performance than Mac Minis, intentionally.” He declined to share specific benchmarks, noting that detailed specifications and performance targets would come closer to the fall launch. But the architectural advantage Microsoft is claiming is clear: while the current Mac Mini with M4 Pro tops out at 48 gigabytes of unified memory and the M4 Max configuration reaches 128 gigabytes, the RTX Spark Dev Box pairs its 128 gigabytes with a Blackwell-class GPU that has a fundamentally different CUDA-based compute model — one that the vast majority of the AI/ML ecosystem’s tooling (PyTorch, TensorRT, llama.cpp, Hugging Face frameworks) is already optimized for.
That CUDA ecosystem advantage is difficult to overstate. While Apple’s Metal framework has made progress, the overwhelming majority of AI training and inference frameworks are built and tested first against Nvidia’s CUDA stack. A developer running models on the Dev Box can use the same code, the same libraries and the same workflows they would use on a cloud GPU instance — a level of portability that Apple Silicon cannot currently match.
The Dev Box is one piece of a three-tier hardware strategy Microsoft laid out at Build. The Surface Laptop Ultra, announced days earlier at Computex, brings the same RTX Spark silicon into a 15-inch laptop form factor for developers and creators who need portability. At the other end of the spectrum, the DGX Station for Windows — built on Nvidia’s GB300 Grace Blackwell Ultra Superchip — targets organizations that need to run frontier models up to one trillion parameters on a deskside system. That machine is expected in the fourth quarter of this year.
The three devices map to a tiered computing model that Microsoft is calling “unmetered intelligence”: small on-device language models (the company’s new Aion 1.0 family) handle lightweight tasks at zero marginal cost; RTX Spark-class hardware runs mid-range models locally for the bulk of development work; and cloud resources are reserved for genuinely frontier-scale problems.
The GitHub Copilot CLI is getting a concrete implementation of this model with a new feature called /fleet, which allows a cloud-based primary agent to build a plan, assess the complexity of each task and route appropriate subtasks to a local model running on the developer’s hardware. The cloud agent handles what requires frontier capability; the local model handles what does not. The result, in theory, is lower cost without lower quality.
Whether Microsoft’s bet pays off depends on questions that will take months to answer. How does the Dev Box actually perform under sustained, real-world workloads? What will it cost? How quickly will the open-source model ecosystem continue to produce capable models in the 70-to-120-billion-parameter range that fit within its memory envelope? And perhaps most critically: will enterprise procurement teams, trained to think of AI as a cloud line item, accept a capital expenditure on desk hardware as an alternative?
The strategic logic, however, is difficult to dismiss. For three years, the AI industry has operated on an implicit assumption: serious AI work happens in the cloud, and the economics of that arrangement are simply the cost of doing business. Microsoft, a company with every incentive to reinforce that assumption, is now selling a machine that undermines it. That is not a contradiction — it is a recognition that the market is moving, and that the company that controls the developer’s local environment and the cloud they deploy to has a more durable advantage than one that controls only the cloud.
Every dollar a developer does not spend on cloud inference is a dollar that can fund another experiment, another iteration, another prototype. For years, the AI industry told developers they needed to rent their intelligence by the token. Microsoft is now asking a different question: what if you could just buy it?
Mistral AI used its inaugural conference on Wednesday to announce a sweeping expansion into industrial manufacturing, a new inference data center south of Paris, and a rebranding of its consumer-facing assistant — moves that collectively signal the three-year-old French startup’s ambition to become the enterprise AI provider of record for companies that refuse to hand their most sensitive data to American hyperscalers.
At the AI NOW Summit, held at a venue in central Paris, co-founder and CEO Arthur Mensch took the stage alongside CTO Timothée Lacroix and Chief Scientist Guillaume Lample to lay out a strategy that stretches from bare-metal GPU clusters to physics simulations for aircraft wings. The company disclosed that it now employs 1,000 people and is targeting €1 billion ($1.17B USD) in revenue for 2026 — a figure that, if achieved, would be an extraordinary growth trajectory for a company that began with 15 employees collaborating with its first customer, BNP Paribas, in 2023.
“We have two convictions at Mistral,” Mensch told the audience. “The first is that in order to deploy AI in the enterprise, you actually need, as an AI provider, to own the full stack.” He described Mistral’s business as fundamentally about “transforming electrons into tokens and intelligence,” arguing that physical infrastructure control matters as much as model quality.
The announcements come at a pivotal moment for Mistral and for the broader European AI ecosystem. The company has raised at least $3.9 billion across nine funding rounds, according to Clay’s funding tracker, including a massive €1.7 billion Series C led by Dutch semiconductor equipment maker ASML in September 2025 at an €11.7 billion valuation, and an $830 million debt financing round in March 2026 from a consortium of seven banks to fund data center construction. Mistral now finds itself in a peculiar competitive position: too large to be dismissed as a research lab, but still dwarfed by the resources of OpenAI, Google DeepMind, and Anthropic.
Its answer, articulated across nearly an hour of presentations Wednesday, is vertical depth — going industry by industry, workflow by workflow, and building the infrastructure to keep everything on premises.
The centerpiece announcement was Mistral for Industrial Engineering, a fully integrated AI stack that combines Mistral’s large language models with physics simulation capabilities acquired through its purchase of Emmi AI, completed earlier in May 2026. The platform targets the aerospace, automotive, and semiconductor industries with tools for accelerating product design, validating simulations, and optimizing production.
The launch came with headline partnerships. Mistral announced it is working with Airbus across its commercial aircraft, helicopter, defense, and space divisions, implementing AI from initial design through to on-board capabilities. For BMW Group, Mistral is serving as a central partner for what the automaker calls its “Large Industry Model” initiative, focused on multimodal reasoning models for crash simulation and other complex engineering tasks. ASML, already Mistral’s largest shareholder, is also an early adopter.
Mensch framed the industrial push as addressing a fundamental gap in how AI is currently deployed. “AI is great today at automating tasks for knowledge workers and for people that are doing software engineering,” he told the summit audience. “But once you move to all the kind of engineers, well, they are underserved.”
The reason, he explained, is structural. Simulating the behavior of a wing or a factory process requires compute-intensive physics solvers that can take hours or weeks per design variant. Traditional simulation creates a bottleneck that makes AI-assisted iteration impractical.
Mistral’s answer is what it calls “physics AI” — data-driven models trained on solver outputs that can predict physical behavior in seconds rather than hours, running on a single GPU. As Mistral’s own blog post on the technology acknowledges, physics AI is “not a replacement for first-principles solvers in every regime” — it is a throughput accelerator for the majority of design-loop iterations, with traditional solvers reserved for verification and edge cases.
“We now have both the language intelligence and the physical intelligence models, and by combining them together we are building delegation loops that allow us to create better tools, that allow us to create better objects that actually have an impact on the physical world,” Mensch said.
The ASML partnership offered a concrete illustration. In a video testimonial shown at the summit, an ASML representative described how the company’s lithography machines run around the clock at customer fabrication plants, and field service engineers need to diagnose issues as rapidly as possible. By combining ASML’s internal engineering expertise with Mistral’s models, “we were able to develop a solution that’s 120 times faster with a similar accuracy as we have today,” the representative said. Another ASML speaker described AI agents acting as “an always-on code reviewer” to catch software defects before they reach customers.
Mistral’s full-stack ambitions extend all the way down to the physical layer. Launched in June 2025, Mistral Compute is a €4 billion ($4.66B USD) investment in data centers in France and Sweden, with a stated roadmap of 200 MW of capacity by 2027 and 1 GW by 2030.
Lacroix described the company’s existing 40 MW facility at Bruyères-le-Châtel, south of Paris, which was built in collaboration with Eclarion and has been training models since early 2026. “It’s been very interesting to see how we can transfer rigor, which is one of our company values, into down to the hardware layer,” he said, describing the process of “fixing compute trays and fixing fibers, allowing us to reach the very best speeds possible on that hardware for training.”
On Wednesday, Mistral announced a new 10 MW facility at Les Ulis in the Essonne department, also south of Paris, dedicated to inference operations and scheduled to open in Q3 2026. Lacroix also referenced a site in Borlänge, Sweden, planned for development through 2027, which will host NVIDIA’s next-generation Vera Rubin GPUs. “One of the benefits for us of owning the hardware layer is also that it lets us be at the very bleeding edge of what infrastructure provides,” he told the audience.
The infrastructure push is funded in part by the $830 million debt financing round announced in March 2026, which Clay’s funding tracker attributes to a consortium of seven banks: Bpifrance, BNP Paribas, Crédit Agricole CIB, HSBC, La Banque Postale, MUFG, and Natixis CIB. And this infrastructure ownership is not merely a hedge against GPU scarcity — it is central to Mistral’s pitch to security-conscious enterprise and government customers. The company’s February 2026 acquisition of serverless platform Koyeb has been integrated into Mistral Studio to support both hosted and on-premises deployments, giving customers a choice between running inference on Mistral’s hardware or their own.
“More and more, the compute world has been getting supply constrained,” Lacroix told the audience. “One of the reasons we’ve been doing all of this and developing all of this data center capacity is to secure compute capacity not only for ourselves but also for our customers.”
In a consumer-facing rebrand with significant enterprise implications, Mistral announced that Le Chat — its conversational AI assistant launched in February 2024 — is being renamed Vibe and reimagined as a unified agent platform for enterprise productivity and software development.
“We are transitioning Le Chat to the Vibe family,” Lacroix told the audience, explaining that the evolution was driven by the growing power of agentic models, particularly the new Mistral Medium 3.5. As the team used Vibe’s coding CLI internally with increasingly complex tasks, “we realized that this really didn’t need to be bound to the CLI, it didn’t need to be limited to code, and we could do a lot more with it,” he said.
Vibe encompasses two primary modes. Vibe for Work is a web and mobile agent that connects to enterprise tools — Google Workspace, Outlook, SharePoint, Slack, GitHub — to perform multi-step tasks such as summarizing emails, analyzing spreadsheets, drafting reports, and scheduling recurring workflows. Vibe for Code is a coding agent available through a web interface, a new VS Code extension, and the existing CLI, capable of building features, fixing bugs, refactoring code, and shipping pull requests. Critically, the same underlying agent powers both modes. “When you access it through our web app or through the CLI, you have access to the same connections, the same tools, the same understanding of who you are, what you do, and what you’re trying to achieve,” Lacroix said.
Pricing starts at free for basic use, $14.99 per month for Pro, $24.99 per user per month for Teams, and custom pricing for Enterprise deployments. Alongside Vibe, Mistral also launched Search Toolkit, an open-source framework for building production search pipelines already in use by shipping giant CMA CGM, which uses it alongside Voxtral to process audio from multiple data sources and return alerts within 15 seconds.
Chief Scientist Guillaume Lample used his portion of the keynote to describe a philosophical shift in Mistral’s model strategy: consolidation of capabilities into fewer, more versatile models rather than maintaining separate specialized products.
Mistral Medium 3.5, the company’s current flagship, absorbs capabilities that previously required distinct models. Pixtral (image processing), Magistrale (reasoning), and DevStral (coding) have all been deprecated as standalone products, with their capabilities folded natively into Medium 3.5. “Now all our models are natively multimodal,” Lample said. “We no longer have Magistrale. This model is deprecated, because all our models will natively be doing reasoning.”
The company is also working on Mistral Large 4, which Lample said would arrive “in a couple of months at most, during the summer,” with expanded capabilities in industrial applications such as fluid dynamics, computational chemistry, computer-aided design, and cybersecurity. On the smaller end of the spectrum, Lample highlighted Mr. Lossier, a 1-billion-parameter OCR model that can process thousands of pages per minute on a single GPU, and the Voxtral speech model family, which has expanded from automatic speech recognition to include text-to-speech with voice cloning. A “duplex” model for real-time conversational speech is planned for release within months.
Lample also made the case for open-weight models becoming more — not less — important in the agentic era. “Today we are building these agentic workflows, these models are running in the background, they are doing a lot of actions, a lot of tool calls, so they are extremely token-hungry, much more than before,” he said. “What we are seeing today is actually a comeback of this small model and the efficient model.” Upcoming models will be trained on more than 200 languages, a multilingual strength now powering a partnership with Amazon to improve non-English interactions on Alexa+.
Mistral’s positioning stands in sharp contrast to the strategies of its most prominent American rivals. While OpenAI and Anthropic have each attracted hundreds of millions of consumer users and derive significant revenue from subscription products, Mistral has leaned almost entirely into enterprise and government deployments. As TechCrunch reported in March when Mistral announced its Forge customization platform at Nvidia GTC, CEO Mensch has described the company as being “on track to surpass $1 billion in annual recurring revenue” — a figure driven largely by corporate clients.
The Forge platform, which lets enterprises train custom models on their own data rather than simply fine-tuning or applying retrieval-augmented generation to existing models, represents the foundation on which the company’s industry-specific solutions are built. As Mistral’s head of product, Elisa Salamanca, told TechCrunch, Forge “lets enterprises and governments customize AI models for their specific needs.” Early partners include Ericsson, the European Space Agency, Italian consulting company Reply, and Singapore’s DSO and HTX, alongside ASML.
Mistral has also built an expanding network of systems integration partnerships to drive enterprise adoption. In February 2026, Accenture and Mistral announced a multi-year strategic collaboration, with Accenture itself becoming a Mistral customer. Mauro Macchi, Accenture’s CEO for Europe, Middle East, and Africa, said at the time that the partnership brings together “sovereign models and the capability to scale technology across industries, geographies and business functions.”
The BNP Paribas relationship offers the most detailed public case study. In a video testimonial at the summit, a BNP Paribas representative described deploying Mistral’s models on-premises to satisfy strict security requirements, developing AI agents for KYC processes that reduced incomplete files from 80% to 10% and compressed processing time from weeks to days. The bank’s LLM platform at its Corporate and Institutional Banking division has now rolled out to 65,000 users. Mensch noted the significance: “We started to collaborate in 2023 where we were 15 people, so that was, I think, really a leap of faith at the time.”
The industrial vertical is also being extended to government clients. Mistral disclosed that it is working with France, Luxembourg, Singapore, Morocco, Greece, and Slovakia to build citizen-facing AI services — from deploying agents that help job-seekers through France Travail to building models that understand Moroccan Darija and Amazigh languages. “We think that AI needs to be specialized and understand structural nuances,” Mensch told the audience. “It needs to speak languages as good as it speaks English.”
For Mistral, Wednesday’s announcements amount to a declaration that the company intends to compete not by matching American AI giants on any single dimension, but by assembling capabilities none of them are willing or able to offer in combination: open-weight models, owned infrastructure, on-premises deployment, physics simulation, and deep vertical customization — all under a single roof.
The strategy demands execution on multiple fronts simultaneously, each requiring enormous capital and specialized talent. The competition is formidable and accelerating. OpenAI has been rapidly expanding its enterprise offerings. Anthropic, backed by billions from Amazon, is building its own corporate AI practice. Google, Microsoft, and Amazon all offer AI platforms deeply integrated with cloud infrastructure that most enterprises already use.
But Mistral is wagering that the world’s most consequential AI deployments — the ones governing how aircraft get designed, how banks process compliance, how governments interact with citizens — will ultimately go to providers that offer sovereignty over data, models, and compute. “AI is too strategic to be left in the hands of a few,” Mensch said, echoing the conviction he described from Mistral’s founding three years ago.
Three years in, the company that started as a Paris research lab with a handful of employees now trains models in its own data centers, simulates physics for the manufacturers that build the world’s planes and cars, and is rewriting its assistant into an agent that can file your pull requests and summarize your inbox in the same conversation. Whether that sprawling ambition coheres into a durable business or stretches Mistral too thin is the €11.7 billion ($13.6B USD) question. The 1,000 people now working there are betting that in enterprise AI, owning the full stack is not a liability — it is the product.