SemiAnalysis: 1.5GW of banned Nvidia GPUs tracked into China

SemiAnalysis's Accelerator Model has tracked more than 1.5 gigawatts of Nvidia AI accelerators that "shouldn't be in China" ending up there anyway, reports Tom's Hardware. The firm says that budget maps to roughly 1.5 million B200-class GPUs at 1,000 watts a card, or about 2.142 million H200-class units at 700 watts. SemiAnalysis does not disclose the product-generation split, so the figure is a range, not a count.

"Our Accelerator Model has tracked over 1.5GW of Nvidia GPUs that shouldn't be in China," the firm writes, adding that "the Department of Commerce now also finally believes that there is more than 1.5GW of NVIDIA chips in China that shouldn't be there." The estimate excludes China-specific parts like the H20 and H800 that were designed for export compliance.

The report also ties Jensen Huang's abrupt departure from Nvidia to the smuggling file. SemiAnalysis argues the resignation is "directly tied to the vast smuggling of GPUs" and writes that "We believe the US intelligence community is a large part of the push to get Jensen out of Nvidia." The firm labels this its "best conspiracy theory take" and presents no supporting documentation, resting its case on a second New York Times story it says cites the intelligence community as its basis five times.

Open prosecutions give the GW number some texture. Supermicro was indicted in March 2026 over roughly $2.5 billion in servers sold to Southeast Asia, with about $510 million allegedly diverted onward to China and Alibaba; the named defendants are Wally Liaw, Steven Chang and Willy Sun. A separate Taiwan case covers 130 Nvidia B300-based servers arranged for unauthorized export, of which 74 reached China. Neither case on its own approaches 1.5GW.

This lands inside a cluster of Huang-exit stories our tracker has been logging this week, including WaPo's report on Rubio rewriting a CIA memo and the FT's finding that Huang never reached Beijing's chip policymakers in two decades.