NVIDIA reportedly agreed to buy Hugging Face for $12.9B — the open-model hub just got a chip vendor
The Information reported Wednesday that NVIDIA has agreed to acquire Hugging Face for $12.9 billion, and Reuters corroborated the figure. Neither company has confirmed, so this is a reported deal that could still fall apart, but the number is one of the biggest in NVIDIA's history. The multiple is the part that jumps out: Hugging Face's annualized revenue is around $150 million (up from roughly $100 million two months ago), which puts the price at about 86 times revenue. The company hosts more than two million public models plus datasets, Spaces and the Inference API, which is why it gets called the GitHub of AI. NVIDIA was in Hugging Face's 2023 round, a $235 million raise at a $4.5 billion valuation alongside Salesforce and Google. Last year the startup rejected a $500 million NVIDIA investment that would have valued it at $7 billion, citing the risk of a single dominant investor. CEO Clément Delangue has said the company is "close to profitability."
Why NVIDIA would pay that price: every open-weight launch, GLM-5.3-Flash, Kimi K3, Qwen3.8, ends with a Hugging Face link, and NVIDIA's own Nemotron family lives there too. Owning the hub means owning default inference routes, featured models and the developer funnel, at a moment when OpenAI and Anthropic are building custom silicon (OpenAI's Jalapeño chip posted its first numbers this week). Jensen Huang's line on the earnings call was "nearly all open models run on NVIDIA"; buying the venue where those models are hosted removes the ambiguity about that claim. The deal also sits inside a spending spree: NVIDIA reported $96.2 billion revenue for Q2 FY27, up 106% year over year, guided to $108 billion next quarter, forecast 70% growth next fiscal year, and said it has $18 billion committed to equity investments through fiscal 2027. It is also reportedly in talks to invest in Perplexity at a $30 billion-plus valuation. The risk is the one the community keeps raising: Hugging Face's value is vendor neutrality, and a chip company owning the neutral hub changes the incentive structure. Walling off downloads would fragment the ecosystem overnight, because Meta, Alibaba and Mistral would self-host or mirror. Watch whether the report hardens into a signed agreement, and whether developers treat NVIDIA ownership as a feature or a threat.
— The Information (via Reuters) · CNBC TV18 · 247wallst
🔗 Reuters via The Star: NVIDIA agrees to buy Hugging Face for $12.9B · CNBC TV18 on the deal context · 247wallst on the 86x revenue multiple
Anthropic opened a research preview of the Model Hardware Standard — MCP now reaches into the physical world
Anthropic on August 27 opened a research preview of the Model Hardware Standard (MHS), a spec built with HHMI Janelia Research Campus that lets AI agents operate physical lab and manufacturing equipment: microscopes, liquid handlers, robotic arms. The pain it attacks is integration. Most devices speak their own vendor protocol, so wiring a lab bench together takes specialists weeks or months. MHS defines a standardized driver with two primitives, "read" (get temperature) and "write" (set temperature), and publishes each device's capabilities and safety limits in a discoverable format. Users can annotate machines in natural language, the weight of a robot arm or its operating limits, and the driver generates a reference file the agent reads before touching anything. Control runs through MCP, a command-line interface, or code files, and it is explicitly model-agnostic: any agent harness can use it, not just Claude.
The early results are concrete. QuEra Computing applied MHS to the laser subsystem of a neutral-atom quantum computer and reported 99.3% laser relock recovery without human intervention. HHMI Janelia compressed a microscopy imaging workflow that took weeks into a day. Carnegie Mellon ran serial-dilution dose-response experiments about three times faster. Genentech coordinated a liquid handler, robot arm and plate reader through a BCA protein assay, and also had to teach Claude that foaming in the samples was a physical problem, not a software failure, which is a useful reminder that the model still reasons mostly from text and images. Partners adding support read like a who's who of lab automation: AWS (via Strands Robots), Tecan, Universal Robots, Automata, QIAGEN, Danaher, Doosan Robotics, and Hugging Face's LeRobot. The pattern I find most interesting is Anthropic's two-stage workflow: Claude explores, adjusts a laser, watches the result through a camera, learns the sequence, then packages it into a deterministic code file that runs without model reasoning at each step. That explore-then-compile loop is showing up across agentic hardware, and it is probably the right division of labor. MHS goes open source after the preview; the safety-evaluation work with launch partners happens first.
— Anthropic (official) · FinWire · ChatAI
🔗 Anthropic: Previewing the Model Hardware Standard · FinWire on the partner list · ChatAI on the MHS pilots
Salesforce shipped Slack Code — coding agents just moved into the team group chat
Salesforce announced Slack Code and dedicated code channels on August 24, with rollout starting August 20. Tag an agent in any Slack conversation and it provisions a project channel with four tabs: the conversation, the agent's plan, line-by-line code diffs, and a live preview of the running output. When the task finishes, the channel auto-archives but keeps a searchable audit log. The launch partners are the big five of coding agents: Anthropic's Claude Code, Cognition's Devin, GitHub Copilot, OpenAI's ChatGPT, and Vercel's agent. Slack Code runs on every Slack plan; the agent subscriptions are sold separately by each provider. The governance story is what enterprises will care about: agents inherit the invoking user's access control lists, so IT provisions no new service identities, and a human sign-off is required before anything pushes to production. Anyone in the channel can pause, redirect, or stop an agent mid-task.
The deeper argument is about where coding agents live. Until now, agent work happened in a solo terminal: one engineer, one agent, and the rest of the team sees the merged commit but never the process. Slack Code inverts that. Non-engineers can trigger an agent, review a plan, and steer a fix without opening a terminal. Slack's Rob Seaman puts it plainly: "AI only creates value when it's part of how a team actually works, not something people go do alone in another tab." The honest tradeoff is overhead: research has shown agents collaborating succeed at lower rates than agents working solo, and more people redirecting mid-task adds friction. This is not optimized for individual speed; it is optimized for visibility, review and audit, which is what regulated industries and larger teams need. It is also Salesforce's strategic answer to the question of what Slack becomes in an AI world: the coordination layer, not a data source. Cognition's president notes its internal merged pull requests rose 10x while headcount grew only about 40%, which is the growth story Salesforce is betting on.
— Salesforce (official) · WithO2 · Newshunt
🔗 Salesforce press release: Slack Code · WithO2 on ACL inheritance and adoption data · Newshunt on the multiplayer shift
Anthropic's agent stack hit general availability — computer use, browser use, Skills, and Files shipped together
On August 20 Anthropic moved four pieces of its agent platform out of beta on the same day: computer use, a new browser use tool, the Skills API, and the Files API. No flashy model launch, just the plumbing that decides whether a team can ship a production agent or stays stuck in a demo. The biggest change is mechanical: computer use now takes several actions per model call instead of one, so long desktop tasks finish in fewer round trips. Early-access customers reported 20-40% fewer round trips and roughly 30% lower cost per completed workflow. The new browser use tool reads page structure and targets fields and buttons by element reference instead of guessing pixel coordinates, which makes it more reliable on clean web applications. Computer use is now eligible for HIPAA-regulated workloads under Anthropic's Business Associate Agreement.
The Skills API turns team procedures into versioned assets: a folder of instructions, scripts and templates that Claude loads only when a task calls for it, running in Anthropic's code-execution sandbox. The Files API adds upload-once, reference-by-ID storage with 5x higher rate limits, about 500 requests per minute, 1TB per organization, and configurable expiration between one hour and 90 days (set once at upload, so plan expiry before you upload). The evidence Anthropic leads with is a customer case: a claims-automation team cut its longest workflow from 32 minutes to 13, cost per task fell about 30%, and completion hit 100% with no prompt changes. Skills and Files are live on Microsoft's Foundry; the updated computer use and browser use tools are headed to Google Vertex AI. My read: the multi-action turn is the quiet architectural shift here. Round-trip reduction changes the economics of long-horizon tasks more than any benchmark score, and the HIPAA eligibility is the signal that Anthropic expects regulated enterprises to run this at scale, not just experiment.
— Anthropic (official) · Brocker · Byteiota
🔗 Anthropic: Build production agents with computer use, the Skills API, and the Files API · Brocker on the GA details · Byteiota on multi-action turns and costs
Google DeepMind shipped Gemini Omni 1.1 Flash — 40-second scene extension and 4K finishing for generative video
Google DeepMind released Gemini Omni 1.1 Flash on August 27, a production-ready update to its generative video model. The headline is scene extension: the model can now analyze up to 10 seconds of prior context, earlier versions only looked at the final second, which is why AI videos drifted into nonsense after a few beats, and continue the clip in 10-second increments up to 40 seconds total. There is also first-and-last-frame conditioning: you hand the model a starting frame and an ending frame and it generates the continuous motion between them, aimed at camera orbits, dolly zooms and seamless loops. And you can pass up to three seconds of reference video so characters stay consistent across shots.
The economics are the part worth paying attention to. 360p draft clips generate up to 60% faster and cost about a third of 720p, and final output upscales to 1080p or 4K. The per-second pricing ladder is explicit: $0.03 for 360p, $0.10 for 720p, $0.15 for 1080p, $0.30 for 4K. That is the cheap-draft, expensive-final pattern film pipelines have used for decades, and Google is the first of the big video labs to publish it as a clean ladder. Availability: the Gemini API in AI Studio, the Gemini Enterprise Agent Platform, Google Flow for AI Plus/Pro/Ultra subscribers, and scene extension in the Gemini app. Adobe Firefly, Figma Weave, GMI Cloud and Runway are already integrating it. The competitive field is Sora, Runway, Kling and Meta's Movie Gen; Google's bet is that control features, not one-shot wow clips, are what pull professional work onto the platform. The 40-second ceiling and the 3-second reference limit make this a system for directing bounded shots, which is honest about what the model can do today.
— Google DeepMind (official) · AI Chat Daily · Superpower Daily
🔗 Google DeepMind: Gemini Omni 1.1 Flash · AI Chat Daily on the shipping details · Superpower Daily on the pricing ladder
A multi-agent framework just swept the GPU kernel leaderboard — agents are getting good at optimizing silicon
A paper submitted to arXiv (cs.MA) this week presents KernelArc, a multi-agent framework for autonomous GPU kernel optimization. Strategy-specialized agents run in parallel and coordinate through three mechanisms: conclusions-only shared memory, a deterministic benchmark guard, and read-only cross-agent state with plateau-triggered drafting. Evaluated on NVIDIA H100 and B200 across category-representative SOL-ExecBench workloads, KernelArc produced custom BF16 GEMM kernels, static cuBLASLt Expert-API configuration tables, fused mixture-of-experts backward, shape-gated decoder-layer fusion, native NVFP4 grouped-query attention, and paged prefill attention. In the public SOL-ExecBench leaderboard snapshot from August 20, 2026, it ranked first on every representative L1, L2, Quantization and FlashInfer task it was evaluated on.
This sits inside a trend that yesterday's Jalapeño story also touched. OpenAI said its AI-generated kernels ran 1.5-1.8x faster than human-expert versions on select GPT-OSS modules. NVIDIA's own AVO system ran autonomously for seven days on B200 and produced an attention kernel up to 3.5% faster than cuDNN and 10.5% faster than FlashAttention-4. There is a healthy counterpoint: Together AI's research found frontier models still fail at writing fast multi-GPU kernels, solving only 28 of 87 problems zero-shot with Fast@1K plateauing around 31%, because multi-GPU communication, not single-GPU compute, is the hard part (tensor-core throughput grew 7.2x from A100 to B200, but intra-node bandwidth only 3x). KernelArc's thesis is that shared multi-agent search broadens exploration within a fixed candidate budget, and that the value of coordination features depends on the kernel and the optimization stage. My read: single-GPU kernel optimization is becoming the first domain where autonomous agentic R&D produces shipping-quality results, and the benchmark guard is the key design choice, because agents can only keep results that measure faster, which keeps the whole loop honest.
— arXiv (cs.MA) · Import AI · Together AI research
🔗 arXiv cs.MA listing: KernelArc · Import AI 470 on kernel-writing agents · StartupHub on LLM multi-GPU kernel limits
Meta's XR Operator lets coding agents test VR apps in a simulator — the missing visual-verification step
Meta shipped Meta XR Operator as an experimental component of Meta XR SDK v205. It is an OpenXR API layer that runs a local MCP server inside your app's process, so any MCP-compatible coding agent, Claude Code or Codex, can see, navigate and interact with a running Quest app in Meta XR Simulator. The agent can capture screenshots, read room geometry such as walls, floors and furniture, change head position and controller state, buttons, triggers, grips and thumbsticks, walk the Unity scene graph, read UI canvases, and register custom tools from your own C# code. The workflow is the full loop: build the scene, launch the app, screenshot, spot the issue, fix it, verify the fix visually. Natural-language testing is included: describe a scenario in a sentence, the agent executes it and returns pass/fail evidence with screenshots. Meta tested it with Beat Games, the Beat Saber studio, and an early agent-driven demo wrote a whole Tic-Tac-Toe game, aimed the controller at cells, pressed A, and played through to a win state; Beat Games reportedly caught a UI overlap defect in Beat Saber's menu that earlier automation had missed.
The caveats are honestly scoped and worth listing: no audio perception, no animation or motion evaluation, no per-finger hand tracking, and it is too slow for real-time interaction, so it works best on static, deterministic scenarios, menu flows, UI navigation, visual regression checks. Unity via Meta XR Core SDK is the supported path; Godot and Unreal users get a standalone package and do their own integration. The significance is the direction: AI coding agents have been good at writing and running tests, but confirming that something looks and behaves right required a human with a headset. XR Operator hands that job to the agent, and it is part of Meta's broader bet on building better developer tools, with more coming at Meta Connect on September 22, rather than paying studios for exclusive content. It is experimental, tool names and APIs will change, and you should not build a production process around it yet. But the pattern, an agent that tests its own work inside a simulator, is exactly where coding agents are heading.
— Meta XR SDK v205 (official) · Road to VR · VR.org
🔗 VR.org on XR Operator's MCP mechanics · Road to VR on the early tests · 映维网 on the build-test-verify loop
Top comments (0)