Threat assessment: high. Export controls were supposed to keep China’s frontier labs off America’s best chips. New research suggests they didn’t work. According to TechCrunch AI, the leading open-weight models from Chinese companies, including Moonshot’s Kimi K3 releases and DeepSeek’s V4 models, show signs of training on Nvidia Blackwell GPUs. The U.S. bans those chips from export to China and to Chinese-owned firms.
The findings land in a brutal week for Nvidia. TechCrunch AI reports that CEO Jensen Huang left the company after admitting to an affair with Alice Huang, an executive at Megaspeed. Megaspeed has been linked to smuggling Nvidia chips into China, and Alice Huang has been alleged to be an agent of China’s intelligence service.
🎯 What the research found
Will DePue, a former OpenAI researcher who now leads the nonprofit Unemployment Labs, analyzed the openly released weights of models from four Chinese companies. Weights are the learned parameters that make up a model. They still carry traces of the hardware setup used to train them.
Key points:
- Targets: DeepSeek, Moonshot AI, MiniMax and Ant Group.
- Finding: The most closely watched models, Kimi K3 and DeepSeek V4, carry fingerprints of Blackwell training.
- Corroboration: Earlier this year, White House technology adviser Michael Kratsios said Moonshot had used export-banned Nvidia chips, citing U.S. intelligence sources.
- Scale: Epoch AI estimates that about two million advanced chips entered China in 2024 and 2025 through schemes built to hide where they were going.
There’s a caveat. None of this proves the models were trained entirely on Nvidia hardware, or that anyone broke export law. The chips could have been sitting in an offshore data center, which is legal under current rules.
🌏 The Malaysia loophole
This is the real story. Chinese firms have found it easy to rent compute abroad, and that demand has driven a data center boom in Malaysia. Epoch’s data puts imports of Nvidia AI accelerators into Malaysia at 275,000 units across 2024 and 2025.
The rules focus on who owns the chips and where they sit, not on who uses them. China hawks have pushed to change that for years. “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.
⚠️ Nvidia’s exposure
Megaspeed has been at the center of that offshore trade for years through its subsidiary Speedmatrix. U.S. and Singaporean authorities have both investigated it. Bloomberg also reported Friday that a company linked to Alice Huang helped smuggle banned chips to a data center in Shanghai.
What stands out here is the liability question. U.S. companies currently have limited duty to track what customers do with controlled chips. But Nvidia and Jensen Huang both vouched for Megaspeed’s data centers, saying Nvidia staff had visited them and confirmed they were legitimate. If investigators confirm smuggling and show that Huang knew about it, Nvidia’s problem gets much bigger. That risk likely played a part in the decision to part ways with him.
Washington isn’t escaping scrutiny either. Commerce Secretary Howard Lutnick, whose agencies license chip exports, could reportedly lose his job over claims that his investigators missed suspicious buyers and got too close to Nvidia officials.
📦 The H200 wildcard
The Trump administration lets Nvidia sell the less powerful H200 to China, as long as shipments take a roundabout route that lets the U.S. collect a 25% tariff. Yet roughly 600,000 H200s are reportedly stuck in Hong Kong, still waiting for Chinese import clearance.
Nobody agrees on why. Theories include:
- Beijing pushing its domestic chip industry (Stanford’s Graham Webster doesn’t buy this one, since demand is so large)
- Leverage ahead of a possible trade deal
- Fears of surveillance tech inside the chips
Short seller Culper flagged this risk back in May. It argued Nvidia has a “China problem” because so much of its revenue depends on Chinese demand, whether through remote data centers or smuggling.
🧭 What to watch
- Tighter rules: Expect pressure for end-user controls that cover remote compute access, not just shipments.
- Nvidia leadership: A new CEO will inherit open questions about export compliance.
- Open-weight scrutiny: DePue’s method could become a standard way to audit where models were trained.
- Commerce shake-up: A change at the top could reset how strictly licenses get enforced.
The big takeaway for practitioners is that the best Chinese open models may owe their edge to the very hardware Washington tried to block. Policy will likely follow the evidence. You can find the full report at TechCrunch AI.