Update, 17 November 2026, 05:40 UTC. SemiAnalysis reposts the segment: what comes after the 1.5 GW. SemiAnalysis posted the Fox Business segment of 12 November as a captioned video on its X account at 02:38 UTC on 17 November. In it, founder Dylan Patel repeats that “there’s more than a one and a half gigawatts of chips in China that are not supposed to be there according to US law”, then adds: “How they get there, that whole web, that’s what’s going to unwind. That’s what’s really dangerous”. Asked what comes of the expected hearings, Patel set the closed-model labs against the open-source ecosystem, said much of that ecosystem is “built on the back of Chinese models” trained on Nvidia chips, and argued that “open source looks a lot weaker” if Nvidia can no longer fund it.
These are SemiAnalysis’s views, quoted from our own transcript of the video; the segment gives no new figure beyond the 1.5 GW above, and neither SemiAnalysis nor Nvidia has published evidence of how the chips moved. Separately, the Justice Department is investigating Nvidia buyer Megaspeed; our piece on the DOJ probe of Megaspeed covers it.
SemiAnalysis wrote in its newsletter of 12 November that its Accelerator Model “tracks every AI chip’s production and destination”, and that “our data shows more than 1.5GW of Nvidia chips, which are restricted, still ultimately ended up in China”. The research firm says the figure excludes the chips Nvidia designed for the China market under US rules.
Epoch AI, a research group that published its own estimate in April, puts chips smuggled to China through 2025 at 290,000 to 1.6 million H100-equivalents, with a median of 660,000. At the H100’s maximum rated power of 700 watts, per Nvidia, that median is about 0.46 GW, and SemiAnalysis’s 1.5 GW is 3.2 times it, by our arithmetic.
Business Pill · Sovereign AI Explained
An evergreen Business Engineer explainer on sovereign AI, why states treat AI compute as strategic. It teaches the general idea only and does not discuss this story.
The key insight: As we read the two estimates, the distance between them is the finding: SemiAnalysis counts more than 1.5 GW of restricted Nvidia chips in China, while Epoch AI’s median of 660,000 H100-equivalents through 2025 comes to about 0.46 GW at 700 watts each, so the newer count is 3.2 times it, by our arithmetic.
What SemiAnalysis Says It Tracked
The newsletter’s own wording on the exclusion is “This figure doesn’t even include chips that Nvidia created to circumvent US rules”. Tom’s Hardware, which reported the estimate on 13 November, names the H20 and the H800 as examples of the excluded parts.
SemiAnalysis also writes 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 newsletter does not say how it knows the department’s view, and Commerce has not published that figure in the material read for this piece.
The firm’s founder gave the same figure on Fox Business on 12 November, the day the newsletter came out. As our filings piece on Nvidia’s China exposure recorded, the segment included this line: “We view there’s more than one and a half gigawatts of chips in China that are not supposed to be there according to US law.”
The newsletter gives no unit count, no split by chip generation and no period. Tom’s Hardware notes that “the report does not explain how it converts accelerator quantities into power figures”; its own conversion gives about 2.142 million chips if all were H200-class parts at 700 watts, or 1.5 million at 1,000 watts each for B200s.

Two Estimates, Two Methods
Epoch’s report says “between 290,000 and 1.6 million H100-equivalents (H100e) were smuggled to China through 2025”. It adds that the median “would be roughly a third of China’s total compute”.
Epoch built the number from two sides: a database of alleged diversions, adjusted for how much goes undetected, and a resale model of gray-market vendors and their volumes. It says “Cumulative allegations of diverted or missing chips total almost 300,000 H100e” by the end of 2025, and that “the allegations are largely not yet proven in a court of law”.
Converted the same way, at 700 watts per H100-equivalent, Epoch’s range runs from about 0.20 GW to 1.12 GW. SemiAnalysis’s figure sits above the top of that range, at 1.3 times the 1.12 GW upper bound, by our arithmetic.
The windows differ. Epoch stops at the end of 2025; SemiAnalysis states no end date for its count, and its newsletter came out in November 2026. Epoch notes that one alleged scheme “switched to aggressively shipping Blackwells once they became available (in 2026)”, which falls outside its own count.
Against China’s Datacenter Base
In October, as our piece on China’s datacenter fleet reported, the FT cited SemiAnalysis for 24 gigawatts of operational datacenter computing capacity in China. The 1.5 GW of restricted chips is 6.25% of that base, by our arithmetic.
Both estimates count chips moved into China. Epoch says its figures “do not cover cloud compute located outside of China but used by Chinese customers”, the channel that the US remote-access rule we covered today addresses.
For the money side, our count of China AI projects anchored on restricted chips reached $3.2 billion, and our piece on Megaspeed’s order records set out about $12 billion of Nvidia orders arranged for China, per the records reported. Every Nvidia piece is indexed on the Nvidia hub.

The Cases and the Bill
SemiAnalysis supports its number with two cases. In a March 2026 indictment in the Southern District of New York, prosecutors allege that $2.5 billion of servers went to a company in Southeast Asia, of which $510 million were smuggled into China. In Keelung, Taiwan, nine people have been indicted over 130 B300 servers, 74 of which reached China, according to the newsletter.
The newsletter says “Both of these cases are being litigated.” One defendant in the New York case has pleaded guilty to four counts, according to the newsletter; the other allegations are the prosecutors’ and have not been proven in court.
On policy, SemiAnalysis writes that the Remote Access Security Act “passed the House 369-22 on Jan. 12” and had yet to receive a Senate vote when the newsletter was written, adding: “We know that NVIDIA and Oracle have been lobbying against it, hard.” The firm argues for “banning export to China while allowing remote access by trusted clouds”.
The Structural Read
As we read the methods, the two numbers are built differently: Epoch models alleged diversions and gray-market resale through the end of 2025, while SemiAnalysis says its Accelerator Model tracks production and destination and states no period, so 2026 shipments of Blackwell parts, which Epoch’s window leaves out, may account for part of the distance.
As we read the policy side, both estimates count chips moved into China; renting compute from clouds abroad is the separate channel that the US remote-access rule and the Remote Access Security Act address, and SemiAnalysis argues for allowing that channel through trusted clouds.
SemiAnalysis, newsletter, 12 November 2026
“Our Accelerator Model has tracked over 1.5GW of Nvidia GPUs that shouldn’t be in China”
Three Implications
FOR ENFORCEMENT SemiAnalysis cites two cases: a $510 million diversion to China alleged in a New York indictment, and 74 B300 servers that Taiwanese prosecutors allege reached China. Both are still being litigated.
FOR THE SCALE Epoch says its median would be roughly a third of China’s total compute; the 1.5 GW is 6.25% of the 24 GW datacenter base SemiAnalysis put on China in October, by our arithmetic.
FOR THE RANGE Converted at 700 watts per chip, the SemiAnalysis count is 1.3 times the top of Epoch’s 90% range, by our arithmetic; the two windows differ.
The Business Engineer Lens
This story maps onto the Business Engineer framework AI’s Geopolitical Chokepoint.
The essay argues the AI supply chain is “creating geopolitical leverage that operates independently of who has the best models”.
As we read it, export controls on chips are one form of that leverage, and the two estimates are attempts to measure how much of it leaks. The essay itself is about the physical supply chain and does not discuss these estimates.
What Is Not Established
SemiAnalysis does not publish its model, its unit count, its power basis or the period its figure covers, and the Commerce view is as SemiAnalysis states it. Epoch’s figures are model estimates with a 90% confidence range, built largely on unproven allegations.
The 700-watt conversion is ours, and it treats every chip as an H100; a different mix would change the GW figures. The newsletter also links its estimate to the timing of Nvidia’s leadership change, which it calls “our best conspiracy theory take on the situation”; Tom’s Hardware notes that the report “provides no direct evidence” on that point, and this piece does not rely on it. Nvidia, SemiAnalysis and Epoch AI were not contacted for this piece.
The Bottom Line
SemiAnalysis says more than 1.5 GW of restricted Nvidia chips ended up in China. Epoch AI’s median of 660,000 H100-equivalents through 2025 is about 0.46 GW at 700 watts each, so the SemiAnalysis count is 3.2 times it, and 1.3 times Epoch’s upper bound, by our arithmetic.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The 1.5 GW estimate, the Commerce view and the case details are as stated in SemiAnalysis’s newsletter of 12 November 2026 (free portion, read in full); apart from one guilty plea in the New York case, the allegations are prosecutors’ and have not been proven in court. Epoch AI’s range and method come from its report of 29 April 2026, read in full. The 700-watt conversion, the GW figures for Epoch, 3.2 times, 1.3 times and 6.25% are our arithmetic.
This piece does not adopt SemiAnalysis’s view on the timing of Nvidia’s leadership change. Nvidia, SemiAnalysis and Epoch AI were not contacted.
Sources: SemiAnalysis: Nvidia and the World Without Jensen (newsletter, 12 Nov 2026) · Epoch AI: Diversion and resale: estimating compute smuggling to China (29 Apr 2026) · Tom’s Hardware (13 Nov 2026) · Nvidia: H100 specifications









