Banned Nvidia chips found inside six Chinese AI models

Watch on YouTube · subscribe for the daily brief
Six of China's strongest AI models carry a hidden fingerprint, and it points to Nvidia chips the US banned from China. The finding comes from Will DePue, a former OpenAI researcher who worked on Sora and GPT-5, as reported by The Decoder. He calls it the first technical evidence that top Chinese labs can use Nvidia's Blackwell chips despite US export controls.
What he found
- The labs: DeepSeek, Moonshot AI, MiniMax and Ant Group. Six models released since April 2026 carry the Blackwell signature; earlier ones like DeepSeek-V3 and Kimi K2.5 show Hopper.
- How it works: before training, PyTorch fills a model's weights with random numbers, and the GPU splits that job across its compute cores. Hopper H100 and H800 use 132 cores, Blackwell B200 and B300 use 148, and the GB200 and GB300 racks 152, so each leaves a different pattern. Training changes most values, but a faint statistical echo survives.
- Which model, which chip: DeepSeek-V4-Flash and MiniMax-M3 map to B200 or B300; DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant Group's Ling-3.1-flash to GB200 or GB300 racks.
- Checks: he reproduced DeepSeek-V4-Flash and Kimi K3 exactly on real Blackwell chips, his method named the hardware of all 24 control models correctly, and the code is on GitHub.
- Scale: about 12,000 Blackwell GPUs across five training runs, more than half a billion dollars of hardware. Kimi K3 alone took about 4,500 GPUs for about 70 days.
The caveat: the fingerprint shows which hardware started the training, not where the chips are or who owns them. Chinese companies can rent data centers outside China, so this alone does not prove smuggling or a broken law. Alibaba's Qwen 3.8 and models from Zhipu, Tencent, Xiaomi and StepFun still show only Hopper.
For builders: if you run these open models, nothing in your stack changes. The pressure lands on export policy, and on Nvidia.
Nvidia
- No leader yet: the board told CNBC it is still selecting a search firm and has not formally engaged any candidates. The Financial Times reports it approached Arm CEO Rene Haas, who spent seven years at Nvidia.
- Stock: down another 2% on Friday, below $4 trillion in market value for the first time since July 2025; down 12.75% for the week, its sixth straight weekly drop.
- Startups: Nvidia joined 105 funding rounds this year, often as the largest investor. One investor told TechCrunch that interim CEO Colette Kress is "much more conservative".
Money and risk
A Federal Reserve survey released Friday names AI the top risk to the US financial system: 87% of respondents cited it, up from 50% in the spring, far ahead of a risk-asset correction at 52%. The three worries: stock valuations, cyberattacks by states or by frontier AI models themselves, and complex AI financing.
Security
- PostgreSQL: fixes for 52 security flaws. The worst, CVE-2026-6460 (CVSS 8.8), lets a database user who can create objects run code on the server; its root goes back to 2004. Update to 18.7, 17.12, 16.16, 15.20 or 14.25 and rebuild C extensions. No attacks reported. PostgreSQL 14 is now end of life.
- Check Point: attackers broke into unpatched Check Point gateways through the PHP2Shell flaws and a new Gaia OS privilege flaw, planted the TSGA backdoor, then moved to VMware ESXi hosts and ran a wiper.
Agents
- FAA: the SMART system, $875 million, now advises all 22 US air traffic centers on routes and arrival order. Humans check every suggestion. The controllers' union, 15,000+ members, says it was not consulted on design or testing.
- Meta Muse: can now phone US businesses for paying users; plans start at $20 a month. Many businesses hang up once they hear it is not a human.
- Amazon: plans a World Product Graph, Reuters reports: data on every product sold online, even ones Amazon does not sell. An internal document says many answers from its shopping assistant are "factually wrong". Amazon says such documents are brainstorming.
Safety
Paul Christiano, an OpenAI Foundation board member, told a Toronto audience that for most people in the room, the most likely cause of death is an AI accident. At the fastest pace of progress he sees "something like a coin flip" of an AI takeover, with a meaningful risk in the next six to twelve months. He admits no model can put a number on it, and says the industry may need to slow down or stop.
Money
- Anthropic IPO: about six times oversubscribed, Bloomberg reports, aiming to raise up to $64.4 billion, which would be the largest IPO ever. Pricing is set for November 18, ticker ANT on Nasdaq. Banks are turning away buyers from mainland China and Hong Kong, some citing ITAR.
- Mirendil: founded last year by two former Google DeepMind researchers, raised $1 billion in a Series A led by Kleiner Perkins and Clearlake. Its agents automate model training; it says they wrote more than 99% of the code behind its own base model.
Open source
llama.cpp reached v1.0.0. From now on, API and ABI breaks in llama.cpp and ggml come with a major version bump, so apps built on it can upgrade safely.
Main story today: ChatGPT Wallet, and how OpenAI's agents now pay for you. Our full story is on the channel.
The through-line
Traces. Model weights that remember the chip that made them, a bug that waited twenty-two years, a survey where AI now tops every other risk.
One thing to do today: if you run PostgreSQL, check your version and plan the update this week.
Sources covered
1 story · 76 sources (42 news, 28 company, 6 voices)
- Production MLOps at GS Caltex with Amazon SageMaker Canvas, part 1🏢AWS AI/ML↗🎙️Simon Willison↗🎙️Simon Willison↗🏢AWS AI/ML↗🏢Vercel AI↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Snowflake↗🏢Google Research↗📰The Hackers News↗📰SiliconANGLE↗🏢AWS AI/ML↗📰The Verge↗🏢Vercel AI↗🎙️Dwarkesh Patel↗🏢Hugging Face↗📰TechCrunch↗📰CNBC↗📰Engadget↗📰TechCrunch↗🏢Hugging Face↗🏢Meta AI↗🏢Google Cloud AI↗📰The New Stack↗📰SiliconANGLE↗📰Ars Technica↗🏢LangChain↗🎙️Damien Benveniste↗📰The Verge↗🎙️Barr Moses↗🏢Databricks↗📰The Decoder↗🏢OpenAI↗📰Engadget↗📰RIA Novosti↗📰TechCrunch↗🏢Hugging Face↗📰The Decoder↗📰TechCrunch↗📰The New Stack↗📰The Decoder↗🏢NVIDIA↗🏢ClickHouse↗📰The New Stack↗🏢MongoDB↗📰Ars Technica↗📰Ars Technica↗🎙️Ben Lorica↗📰TechCrunch↗📰The Decoder↗📰TechCrunch↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰The Decoder↗📰BleepingComputer↗📰SiliconANGLE↗📰The New Stack↗📰The Verge↗🏢ClickHouse↗📰SiliconANGLE↗📰The Verge↗📰InfoQ↗🏢OpenAI↗📰TechCrunch↗📰TechCrunch↗📰The Verge↗📰InfoQ↗📰The Hackers News↗🏢Starburst↗🏢Modal↗🏢Anthropic Research↗🏢Vercel AI↗🏢Vercel AI↗🏢Vercel AI↗