NVIDIA buys Hugging Face for $12.93B, and the price is the least interesting number

Good morning 👋 NVIDIA just bought the platform every AI lab, including its own rivals, uses to publish open weights. The price is almost the least interesting part.

In today's issue:

• 🔭 NVIDIA buys Hugging Face for $12.93B, and the price is the least interesting number
• 🧠 DeepMind's weather model goes live off satellite feeds, Nano Banana lands inside Docs and Slides
• 🔬 SWE-Gate finds the gap between green tests and a mergeable PR
• 🚀 Resect AI raises $25M from private equity to catch hallucinations before they ship
• 🛠️ Benchmark local inference before NVIDIA reprices the stack
• 🔗 OpenAI backs frontline cyber defense with $1B, Hugging Face gives agents memory that outlasts the session

🔭 THE ONE THING

🤝 NVIDIA buys Hugging Face for $12.93B, and the price is the least interesting number

NVIDIA is acquiring Hugging Face for $12.93B ($11.9B in cash to stockholders, up to $1B more in employee retention equity), confirmed by NVIDIA's own announcement and an SEC 8-K dated Sept. 2. Clément Delangue told CNBC he approached NVIDIA himself this summer. Hugging Face is the shelf every lab, including NVIDIA's rivals, uses to publish and pull open weights. Fold that into the company selling the GPUs those weights run on, and open-model distribution just picked an owner. Worth watching whether the Hub stays neutral for AMD and TPU users, or quietly starts favoring NVIDIA's own stack.

🧠 MODELS & RELEASES

• 🌦️ Google DeepMind shipped WeatherNext 3, swapping stale six-hour numerical runs for live geostationary satellite feeds and sharpening the grid from 25km down to 5km with hourly updates. Medium-range precipitation error drops as much as 60%, and it's not a paper demo, it's already live in Search, Gemini, Maps and Earth Engine. source
• 💻 NVIDIA and Microsoft used IFA to make local inference boring: Hermes Agent, OpenClaw and Perplexity's portable app now auto-detect your GPU and configure models on-device, with llama.cpp picking up to 1.9x more throughput on a 5090. The hardware for it, RTX Spark, ships in October: a petaflop of Blackwell and up to 128GB of unified memory crammed into a laptop or a compact always-on desktop. source
• 🎨 Google built Nano Banana into Workspace as Pics: generate and retouch images, down to editing text inside them, without leaving Docs or Slides. No extra charge if you're already on AI Pro, Ultra, or most business Workspace tiers. Live now at pics.new. source

🔬 RESEARCH HIGHLIGHTS

• Chain-of-thought legibility isn't interpretability. A new paper checks which steps in a model's reasoning trace actually move the outcome, measured through Monte Carlo rollouts, against which steps an LLM judge flags as important. The judges beat random guessing. They also fall well short of the ceiling, worst on traces that land on the correct answer. So the readable trace and the functional one diverge more than anyone building a process reward model on top of "the reasoning looks sound" would want to hear. link

• SWE-Gate finds the gap between green tests and a mergeable PR. Across 303 real repair tasks pulled from 75 Python repos, agent patches racked up 644 passes on functional tests. 221 of those, roughly a third, still failed constraints lifted straight from actual review comments: style, docs, whether the fix fit the codebase's architecture. Coding-agent benchmarks that stop at "tests pass" have been grading on a curve nobody told you about. link

🚀 AI STARTUPS

• Resect AI came out of stealth with $25M, and the money isn't the usual VC crowd, it's private equity, unnamed firms at that. The pitch: an "accountability layer" that catches and patches LLM hallucinations in-model rather than after the fact, aimed at finance, healthcare, and publishing shops that can't ship a wrong answer. CEO Kevin Owens and Chief AI Officer Tim Walton are running it out of Portland. No backer names, no valuation. source

🛠️ TRY THIS

Benchmark local inference with Hugging Face's new WebGPU kernels before NVIDIA reprices the stack

1. Install @huggingface/kernels from npm alongside whatever browser inference setup you're already running. It ships 207 WebGPU ops pulled straight from the Hub.

2. Swap one hot operation, matmul or softmax, from onnxruntime-web over to getKernel() and rerun your existing test suite against it.

3. Time both paths on the same device. HF claims a 2.57x geomean speedup across 809 cases, with one Einsum op running 10,000x faster, so your actual op matters more than their average does.

4. If the win holds, that's real money: browser inference gets cheaper right as NVIDIA takes over pricing on the compute you'd otherwise lean on.

Prompt: Write a benchmark script that loads an ONNX model with onnxruntime-web, times 100 inference passes on my slowest operation, then repeats the same passes using @huggingface/kernels' getKernel() for that op. Print median latency for both and the speedup ratio.

🔗 QUICK LINKS

• OpenAI put $1B behind Daybreak for Frontline Defenders, opening frontier cyber AI, training, and support to essential-services operators who'd otherwise be priced out. source
• Hugging Face shipped Funes, a memory layer that lets Claude Code, Codex, and other agents stop starting from zero every session. source
• Kier Group is running Copilot on UK construction sites for safety checks and paperwork. Unglamorous, but it's the kind of rollout that sticks. source
• IBM Research wired its time-series models into Confluent streams, so forecasts arrive with the data instead of a batch job later. source

Watch whether the Hub stays neutral. That's the tell for how this actually plays out.

Pradeep Perugu