A forensic investigation shared by a former OpenAI engineer reveals that Chinese AI companies have been training models using Nvidia's most advanced Blackwell architecture. Digital fingerprints unique to this chip were found in models belonging to DeepSeek, Moonshot AI, MiniMax, and Ant Group. This chip falls under the hardware category the US has banned from being sold to China.

What happened?

The report examined specific computational patterns and performance signatures in model outputs, and claims they show Blackwell-generation GPUs were used in the training infrastructure. Four major Chinese labs are named — suggesting this isn't an isolated case, but a widespread practice.

Why does it matter?

Since 2022, the US has been progressively tightening restrictions on the sale of Nvidia's most powerful AI chips to China. First the A100 and H100 were restricted, prompting Nvidia to release performance-reduced "China-specific" variants like the H800 and H20. As Washington banned those variants too, Nvidia designed new discounted versions — a chase that has been going on for years. Blackwell is the newest and most powerful link in that chain; in theory, it was never supposed to enter the Chinese market.

Yet the fact that this chip is showing up in the training infrastructure of China's top-tier labs shows that enforcement exists only on paper. The technical explanation is simple: GPUs are physical objects that can be moved through cross-border logistics networks. Shell companies in Southeast Asia, relabeling, and access via cloud rental are long-known loopholes. Independent analysts have repeatedly pointed to an unusual spike in shipments routed through Singapore in recent years. A chip isn't a fusion reactor — it's an electronic component manufactured in the millions and relatively easy to ship. Effectively halting hardware exports isn't all that different from stopping drug trafficking — if demand is high enough and margins are sufficient, the network rebuilds itself.

Who does this affect, and how?

For the US, this undermines a core assumption of its sanctions strategy: that controlling physical hardware can delay AI capability. On the Chinese side, the picture is more complicated. Huawei's domestic alternative, the Ascend series, is improving but still can't match Nvidia's CUDA software ecosystem or raw performance. The fact that labs are turning to risky, costly black-market supply is indirect evidence that domestic chip production still isn't sufficient.

PartyGainRisk
Chinese AI companiesAccess to the most powerful hardwareSanctions, supply disruption
NvidiaHidden demand, indirect revenueReputational damage, legal risk, US pressure
US policy-Enforcement appears ineffective

Nvidia itself is in a difficult position here: the company officially maintains that it complies with export rules and that responsibility lies with intermediary resellers. But the presence of smuggled chips in the market creates a gray area that indirectly feeds back into the company's revenue as well.

What's next?

Washington's response to this report is worth watching — will new sanctions follow, or will there be a pragmatic loosening given that the enforcement mechanism is already full of holes? Whether Nvidia's supply-chain tracking commitments have any real substance is also an important question. Meanwhile, how much of the performance gap Huawei's upcoming Ascend models manage to close will determine whether China's reliance on black-market supply actually decreases.

Export controls show that nation-state borders are becoming increasingly porous for digital products. Hardware is still a physical object — but the global logistics network is extraordinarily skilled at moving that object to places it isn't supposed to go. If this report is accurate, the US strategy of trying to preserve technological superiority in the AI race through hardware controls appears, at least so far, not to be delivering the intended result.