Banned Blackwell chips trained six Chinese frontier AI models

Images

×

Image description the-decoder.com

Image: Will DePue / Unemployment Labs Will DePue / Unemployment Labs

Tim Fernholz techcrunch.com

China warns US over chip curbs amid Nvidia row euronews.com

‹ ›

Six of China's most powerful AI models carry the fingerprint of Nvidia's export-banned Blackwell chips, according to a former OpenAI researcher who has published what he describes as the first technical evidence that the country's leading labs trained their newest systems on hardware Washington forbids them to have. Will DePue, who worked on Sora, GPT-4o and GPT-5 at OpenAI before founding the research group Unemployment Labs, says senior researchers at OpenAI and Anthropic have independently checked and confirmed his findings.

The evidence was in the files the labs released themselves. Before training begins, the PyTorch framework fills a model's empty weight slots with random numbers generated from a seed, and the GPU divides that work across its compute cores. Different chips have different core counts — Nvidia's older Hopper generation carries 132, Blackwell 148 or 152 — and initialisation leaves a characteristic pattern in slots that training never touches. DePue says he reproduced the results on real Blackwell chips and published his verification code on GitHub; the method correctly identified all 24 control models with known hardware, including Meta's Llama 3.1 trained on H100s.

The six models, released since April, come from DeepSeek, Moonshot AI, MiniMax and Ant Group. Four of them, including DeepSeek-V4-Pro and Kimi K3, match the 152-core pattern of Nvidia's liquid-cooled GB200 and GB300 flagship racks. DePue estimates the training runs he examined consumed roughly 12,000 Blackwell GPUs, as much computing power as 30,000 older H100s. At about $70,000 per chip, the largest clusters cost $200 million to $300 million each, more than half a billion dollars in total; Kimi K3 alone, he calculates, needed some 4,500 Blackwell processors running for about 70 days.

The finding lands on an affair already in motion. Jensen Huang left Nvidia this week after admitting an affair with Alice Huang — no relation — an executive at Singapore-based Megaspeed whom a US intelligence memo circulated this spring reportedly identified as an asset of China's Ministry of State Security. Bloomberg reported on Friday that a company she set up, Shanghai Longteng Intelligent Computing, ended up with servers holding H800 chips, banned from export to China since 2023, after they were bought from Dell by an official Nvidia partner in Singapore. Malaysia is probing a Megaspeed subsidiary over possible diversion. Epoch AI estimates that around two million advanced chips entered China in 2024 and 2025 through schemes designed to hide their final destination, and that 275,000 Nvidia accelerators were imported into Malaysia alone, feeding an offshore data-centre boom.

'If we care about the U.S. tech edge, it doesn't make any sense to me that you would not allow Chinese companies to access chips in mainland China or Hong Kong and Macau, but then say, actually, if you just go over to Malaysia, you can do whatever you want,' Sam Bresnick of Georgetown's Center for Security and Emerging Technology told TechCrunch. Beijing calls the spying allegations a pretext for a 'technological blockade' and warns Washington against tighter curbs; Congress, in House Speaker Mike Johnson's words, is developing a consensus for limiting exports. The Trump administration still permits the less powerful H200, provided shipments take a circuitous route that lets the US claim a 25 percent tariff.

White House technology adviser Michael Kratsios said earlier this year, citing US intelligence, that Moonshot had used export-banned chips — a suspicion without published proof until now. The models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun still show only Hopper signatures: the previous generation of export-restricted silicon, doing the same work.