Forensic read of model weights finds six Chinese frontier models started on export-controlled Blackwell GPUs
- Six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry weight fingerprints matching Blackwell B200/B300 (148 SMs) or GB200/GB300 (152 SMs), chips that need a US export license for China and have no reported license.
- The method checks out against known answers: for 24 models whose developers disclosed the training hardware, the fingerprint agreed every time, and NVIDIA's own Nemotron-3 Ultra reads as 152 SMs, matching Kimi K3.
- Matches run from 49 to 1,385 sigma, against roughly 8 sigma expected as the best chance match across the ~6 trillion seed-and-layout combinations searched; DeepSeek-V4-Flash and Kimi K3 were also confirmed on physical Blackwell chips.
- Not every Chinese lab has moved: the newest readable models from Alibaba (Qwen), Zhipu (GLM), Tencent, Xiaomi and StepFun still read as Hopper, and Meituan's LongCat-2.0 shows no NVIDIA fingerprint at all.
- The report shows hardware access only, not where the chips sat or whether controls were violated; training on rented GPUs outside China, for example in Singapore or Malaysia, is legal under current rules.
Hacker News 의견들
Export controls never really work. You find western chips in Iranian drones, resellers selling to resellers, the whole grey chain. It's pretty much uncontrollable.
They're not meant to be airtight. They raise the cost of doing business with someone you don't like, and the usual effect is setting a program back by months, not stopping the commodity from reaching the market.
Cold War precedent: the US ran shell companies to get titanium out of Russia for spy planes. If someone is willing to pay, smugglers find a way.
What if the government tolerated the exports on purpose? Marked-up grey market chips drain the buyer's money while killing any urgency to build their own silicon.
That logic is backwards. When one critical layer of your stack sits under an adversary's control, sovereignty becomes an existential priority, not a lower one. They can turn the tap off tomorrow.
So export controls are really export friction.
The post itself says the chips could be in Singapore or Malaysia. Chinese firms renting GPUs in data centers outside China is completely legal under current rules.
Right, and nobody is alleging the labs lied. In every case where they did disclose the training chip, this research found the disclosure accurate.
Worth noting the author is a SWE-turned-influencer China hawk and there's no code linked anywhere. Everyone here is assuming it's true without checking the contents.
I think you're confusing Will DePue with David Perdue. Similar last names, different people.
Honestly this convinces me export controls don't work, so why keep doing them.
Cute trick, but DeepSeek already published an inference engine with kernels only for SM100 and Ascend 950, benchmarked on Blackwell. Nobody found that remarkable two weeks ago, and now the same fact is a super secret requiring RNG init statistics?
I want to check the PyTorch code before believing this. Philox-based parallel PRNG is a well-studied area, and I'm surprised it can fingerprint hardware this reliably without leaning on some PyTorch quirk.
I guarantee this isn't an accident. Whatever gets shipped over there is carrying its own special payload.
Weird how many articles with the same tone dropped in the past 24 hours. Hard not to read that as coordination from someone who wants to hurt Nvidia.