I have been writing about AI for a while now, and I do not think I have ever seen a week like this one. The most important man in the hardware business walked out the door, four nonprofits asked a judge to put a ceiling on frontier AI, and a forensic report says China's best models were built on chips that were never supposed to get there. Turns out those last two stories are tangled up with the first.
01 of 03
Nvidia Loses Jensen Huang, and Washington Smells Blood
Nvidia's board announced Thursday, November 12, that Jensen Huang resigned as president, CEO, and board member, effective immediately. Per the company's 8-K, the board accepted it the day before. CFO Colette Kress, who has run Nvidia's finances since 2013, is interim CEO and keeps the CFO job too. Longtime director Harvey C. Jones, former Synopsys CEO, becomes Nvidia's first board chair. The board says it will look at internal and external candidates for the permanent job.
The official statement does not say why. The reporting does. According to Tom's Hardware and TechCrunch, Huang admitted to an affair with an executive at Megaspeed, a company U.S. officials reportedly suspect of helping route advanced Nvidia chips into China, and that executive is reportedly suspected by U.S. intelligence of being a Chinese intelligence asset. None of that is proven in court, and Nvidia has not confirmed it.
The market did not wait for a verdict. TechCrunch reported Nvidia stock down nearly 12% in early trading Thursday, and as much as 19% since the prior Friday, which is more than $855 billion in market value. CoreWeave fell about 9%, Nebius about 16%.
Congress is now on it from both sides. House Republicans Brian Mast, John Moolenaar, and Rick Crawford, who chair Foreign Affairs, the China Select Committee, and Intelligence, asked Huang to appear, and Mast says he agreed to closed-door testimony before Thanksgiving. Separately, Rep. Richard Neal and Sen. Ron Wyden sent Commerce Secretary Howard Lutnick a letter demanding he preserve records. Their allegation: the H200-to-China deal, where chips came through the U.S. first and picked up a tariff, amounted to "an unconstitutional export tax," and it got approved after Nvidia gave $26 million toward Trump projects, including the White House ballroom.
Here is where I land. I do not care about anybody's marriage. I care that the single most important supplier in AI apparently had its CEO tangled up with a company accused of smuggling its own product to China, and, if the Democrats' letter is right, Commerce kept letting that company buy chips anyway. That is not a gossip story. That is an export control system that looks like it ran on vibes. Kress is a steady hand, but the real question is whether Nvidia's China business survives the next six months of subpoenas.
02 of 03
Four Nonprofits Sue Five AI Labs to Cap Frontier Development
On November 12, four nonprofits filed a 63-page complaint in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI, and SpaceX's AI unit. The plaintiffs are the Irreplaceable Education Fund, Public Citizen, the Heat Initiative, and the Alliance for Secure AI. Their lawyer is Matthew Platkin, the former New Jersey attorney general.
The legal hooks are California's Unfair Competition Law and public nuisance. The argument is that racing toward more capable systems, while the companies' own leaders publicly warn those systems could go catastrophically wrong, is an unfair business practice and a danger to the public. They point to warnings from Anthropic CEO Dario Amodei and Elon Musk as proof the companies know the risks.
What they want is narrower than the headlines say. As described in coverage of the complaint, each company would be capped at the capability of its most advanced system as of November 12, and barred from having AI systems direct research into more capable models. ChatGPT, Claude, and Gemini would not necessarily have to shut down. The Alliance for Secure AI is specifically calling for a ban on "recursive self-improvement." Meta declined to comment, and the other four did not respond to reporters. No hearing date has been set.
What I want to know is how a state court judge is supposed to define "most advanced system as of November 12" with a straight face. I get the impulse. When CEOs keep telling the public their product might end civilization, somebody is eventually going to take them at their word and file a lawsuit. But a court order is a blunt tool for a technical problem, and I would bet on this getting tossed or stuck in procedure long before it freezes anything. The bigger effect is discovery. If this survives a motion to dismiss, those internal safety documents start seeing daylight, and that is what the labs should actually be nervous about.
Source: Public Citizen's announcement
03 of 03
Researcher Finds Blackwell Fingerprints in Six Chinese AI Models
Will DePue, a former OpenAI researcher who now runs the nonprofit Unemployment Labs, published a forensic report on November 12 claiming six top Chinese models carry the fingerprint of export-banned Nvidia Blackwell GPUs. The six: DeepSeek-V4-Flash and DeepSeek-V4-Pro from DeepSeek, Kimi K3 and Kimi K3.1-Pro from Moonshot AI, MiniMax-M3 from MiniMax, and Ling-3.1-flash from Ant Group.
The trick is clever. When PyTorch sets a model's starting weights on an Nvidia GPU, the random number layout depends on the chip's count of streaming multiprocessors. B200 and B300 chips have 148, GB200 and GB300 have 152, Hopper chips have 132. Some of those starting weights never change during training, so you can read them off the published model and work backward to the hardware. DePue validated the method against 24 cases where the hardware was publicly disclosed, and reproduced the actual values on rented H100, B300, and GB200 machines.
Not everybody got flagged. Alibaba's Qwen, Zhipu's GLM, Tencent, Xiaomi, and StepFun models read as Hopper. DePue's rough estimate is about 12,000 Blackwell GPUs across five of the final training runs, though he says plainly he is "not confident" in those counts. He is also careful to say this shows what hardware initialized the models, not who owned the chips, where they sat, or that any law was broken.
To me this is the most important story of the three, even if it got the least airtime. Export controls have been argued about in the abstract for three years. Now somebody has a repeatable way to check a model's weights for the hardware it came from. The report does not say how those Blackwells got to China, but with the Megaspeed mess at Nvidia, a lot of people in Washington are about to start asking. If Commerce is not already copying this method, it should be.
Source: Unemployment Labs report