The Chinese Efficiency Story May Be a Smuggling Story
Machine learning researcher Nathan Lambert, who toured Moonshot and other Chinese labs this year, wrote after Kimi K3's release that he increasingly hears those labs have access to meaningful training compute by skirting export controls. If that is right, it complicates the most widely repeated claim in AI, that Chinese models are matching American ones on a fraction of the hardware.
Per Bloomberg, Mackenzie Hawkins reports that a year of prosecutions across five countries has exposed a shadow trade in restricted Nvidia processors. Trump officials believe DeepSeek installed Blackwell chips at its Inner Mongolia cluster, levied a similar accusation at Moonshot after Kimi K3 matched top American models on some benchmarks, and suspect Alibaba was an end customer in the case involving Super Micro's co-founder. Alibaba denies it. Neither lab has commented.
The scale is contested and both numbers belong in the same paragraph. SemiAnalysis, whose work Jensen Huang himself cites, says supply chain checks and comparisons of production against installation data show hundreds of thousands of chips have slipped through, worth billions and enough to fill a handful of hyperscale facilities. Nvidia says prosecutors allege diversion of less than half of one percent of its products, a drop in the bucket against the roughly two million processors Chinese firms will make this year. Bloomberg reported last year on plans for Chinese desert data centers housing 115,000 restricted processors, which Nvidia and officials called unrealistic. US officials now believe that much and more has been smuggled.
The due diligence questions are specific. Flying Tiger ordered 130 Super Micro servers with Blackwell chips for a Taiwanese data center, and the indictment says the space it had leased was clearly insufficient to hold them. Nvidia and Super Micro approved the deal after an in person visit, 74 units reached China through Japan and Indonesia, and Taiwan's Coast Guard seized the remaining 56. An Nvidia manager was arrested as the alleged central figure. In the larger case, Super Micro co-founder Wally Liaw is accused of routing $2.5 billion of hardware through OBON Corp in Bangkok, a company whose building leases the indictment calls fake and which would have needed over 100 megawatts of data centre space in a country running only a few hundred. Nvidia's answer is that a startup's growth plan in a friendly nation is an opportunity for America rather than a red flag to be feared.
Enforcement capacity is the quiet scandal. The Bureau of Industry and Security had roughly a dozen overseas enforcement staff until a budget increase this year. Its chief Jeffrey Kessler told lawmakers in July that there are cases the agency has not detected or lacks the resources to chase. BIS knew about the Flying Tiger network before the public did and did not block shipments, and its request that Super Micro halt sales to OBON was voluntary, with no warning to other vendors. One US official called it the worst whack-a-mole game they have ever played.
Policy is pulling the other way. Huang persuaded Trump in December to allow sales of up to a million previous generation Hopper accelerators, giving Chinese buyers hardware on par with Huawei's best, and the change was announced the same day as a smuggling bust. Liza Tobin, who ran China policy at the National Security Council, says the Nvidia chief executive has more influence over export control policy than members of the president's cabinet. Senators Warren and Banks suggested in March that Nvidia's leadership may have materially misled the government. Nvidia has not been accused of violating export controls and says it follows the law, which Daniel Remler, formerly of the State Department, calls the problem: a posture of following the law and no more.
What to watch is whether anyone revises their view of Chinese model efficiency. A great deal of capital allocation rests on the belief that frontier capability can be reached cheaply, because if it can, the $745 billion of hyperscaler capital spending this year is overbuilt. If the convergence on that benchmark chart was bought rather than engineered, the opposite conclusion follows, and the case for owning the full stack inside your own borders gets stronger rather than weaker.