SemiAnalysis Claims Over 1.5 GW of Banned Nvidia Chips Reached China

More than 1.5 gigawatts of export-restricted Nvidia chips have "ultimately ended up in China," research firm SemiAnalysis said in a Nov. 12 newsletter post. The figure comes from its Accelerator Model, which tracks AI accelerators from production line to final destination. It excludes China-specific parts Nvidia designed to comply with U.S. export rules, such as the H20 and H800.

SemiAnalysis gave no unit count and no split across chip generations. Tom's Hardware, which reported the estimate on Nov. 13, did the arithmetic from rated power. The total equals roughly 1.5 million B200-class GPUs at 1,000 watts each, or about 2.1 million H200-class GPUs at 700 watts each. A real mix would land between those numbers.

For scale, xAI's first Colossus cluster held roughly 220,000 to 230,000 Hopper GPUs, according to Tom's Hardware. The 1.5 GW figure therefore implies several times the compute of that flagship installation reached China outside legal channels. That comparison is ours by way of the reported cluster size, not a SemiAnalysis claim.

What does SemiAnalysis actually claim?

The post, written by founder Dylan Patel under the title "Nvidia and the World Without Jensen," states: "Our Accelerator Model has tracked over 1.5GW of Nvidia GPUs that shouldn't be in China."

It goes further. The post says 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." SemiAnalysis cited no Commerce document and did not explain how it learned of the department's view. The claim is unverifiable on current public record.

The firm also stakes out a position on motive. It says Jensen Huang's resignation, four days after Nvidia's board began examining his conduct, "is directly tied to the vast smuggling of GPUs." It adds that the U.S. intelligence community was "a large part of the push" to remove him, calling him the "biggest defender" against faster China restrictions. SemiAnalysis labels this its "best conspiracy theory take" and presents no documents showing intelligence agencies influenced the board or that Huang participated in unauthorized exports. Nvidia has said Huang resigned as CEO and director, with CFO Colette Kress serving as interim chief executive.

The firm's policy stance is unusual among China-chip commentators. It argues that letting Chinese labs rent GPUs abroad serves U.S. interests better than banning the practice. It opposes the Remote Access Security Act, which passed the House 369-22 in January and now awaits a Senate vote.

Which prosecutions support the estimate?

SemiAnalysis anchors its number in two criminal cases. Both remain in court.

• Manhattan indictment. Supermicro co-founder Wally Liaw, Taiwan sales head Steven Chang and contractor Willy Sun were indicted in Manhattan federal court on March 17, 2026. Prosecutors say about $2.5 billion of servers went to a Southeast Asian company, and $510 million of them were then diverted to China. Dummy servers were staged to pass audits by Supermicro and the Commerce Department. Liaw has pleaded not guilty and faces trial next March. Sun pleaded guilty on Oct. 8.
• Taiwan case. Taiwanese prosecutors indicted nine people linked to Nvidia, Supermicro, Albatron, Chief Telecom and Flying Tiger over 130 Nvidia B300 servers, 74 of which reached China. Defendants allegedly used shell companies, forged documents and declared Japan as the destination. The Taiwan coast guard found the servers by chance during a drug-smuggling investigation in May. SemiAnalysis puts the case at about $65 million.

Neither case approaches 1.5 GW on its own. SemiAnalysis did not list the other shipments behind its model, so the bridge from two prosecutions to the headline figure rests on modeling the firm has not published in detail.

How does this fit the week's reporting?

The estimate lands amid a run of stories on restricted Nvidia silicon. Bloomberg reported on Nov. 13 that banned Nvidia servers ended up in a Shanghai data center linked to former Megaspeed executive Alice Huang. Reuters identified more than $3 billion of Chinese AI data center projects and deals built on restricted H100, H200 and H800 chips.

Nvidia's public position predates the new estimate. At Computex in May 2025, Huang said Nvidia had "studied all their deployments of our GPUs and counted every one" and found "no evidence of any AI chip diversion," noting that a nearly two-ton Grace Blackwell system would not fit in a backpack. Neither the SemiAnalysis post nor the Tom's Hardware report included an Nvidia response to the 1.5 GW figure.

The gap is now explicit. One side counts every deployment and finds nothing; the other models 1.5 GW of displaced silicon and sees Commerce agreement. Until Commerce speaks publicly or the model's shipment-level detail appears, the number stays an analyst estimate wearing a very large unit prefix.