Nvidia CEO Jensen Huang might hang on as CEO. But his influence on U.S. AI policy is over. | Fortune

• Nvidia was built around one man. Now a spy scandal threatens that model.
• Before the scandal broke, Nvidia CEO Jensen Huang spoke to Fortune about China policy.
• Anthropic looks to price its IPO at a $1.4 trillion valuation.
• Anthropic debuts new Fable model at a price 20% below OpenAI.
• OpenAI unveils a new cyber model.
• Anthropic inches closer to RSI, worrying AI safety experts.
• AI infrastructure provider Fireworks hacked.
• Google DeepMind achieves a breakthrough in formal verification.
• Podcaster Dwarkesh Patel has an interesting new AI benchmark.

Before we get to the news, can I just say, you are not going to want to miss this year’s edition of Fortune Brainstorm AI San Francisco.

We have an exceptional line-up of speakers for you: OpenAI president Greg Brockman and Microsoft AI CEO Mustafa Suleyman will both be onstage. Once the closest of partners, their two companies are increasingly, competitors, so I’m eager to hear how differently they see the road ahead. AI pioneers Fei-Fei Li, now of World Labs, as well as Andrew Ng, now of LearnVector, will be discussing their view of AI’s future too. Palo Alto Networks CEO Nikesh Arora will talk about AI and cybersecurity. Greylock partner Reid Hoffman, CapitalG’s Jill Chase, and Kleiner Perkins’ Mamoon Hamid and Ilya Fushman will weigh in on where the money is going. Recursive Superintelligence’s Richard Socher will talk about RSI and automating AI research.

Plus, we’ve got Nick Fox from Google. We’ve got Fireworks CEO Lin Qiao. We’ve got Cerebras’ Andrew Feldman, Vanta’s Christina Cacioppo, Unconventional AI’s Naveen Rao, Brex’s Pedro Franceschi, Maven Clinic’s Kate Ryder, Box’s Aaron Levie, Atlassian’s Mike Cannon-Brookes, New Work Foundation’s Clara Shih, Twilio’s Khozema Shipchandler, Rubrik’s Bipul Sinha, Harvey’s Winston Weinberg, Dynamic Creatures’ Marc Theermann, SendCutSend’s Jim Belosic, Apptronik’s Jeff Cardenas, Accenture Song’s Ndidiamaka Oteh, and Senra Systems’ Jordan Black. It’s going to be epic. So please join me and many of my Fortune colleagues at the St. Regis in San Francisco on December 7th and 8th. You can apply to attend here.

The biggest news this week in AI has been the revelation of Nvidia CEO Jensen Huang’s affair with a woman that U.S. intelligence reportedly believes is a Chinese spy. Whether Jensen manages to stay in his job is an open question. A lot of that will depend on exactly what Jensen knew and exactly what information he disclosed to Alice Huang, the woman at the center of the scandal. As more information comes out, it is already not looking too good for the embattled Nvidia CEO. In addition to Nvidia’s board announcing an independent investigation, House Republicans have already called the CEO to testify at a closed door hearing next week, and U.S. President Donald Trump, who initially defended Jensen, has said the FBI is now investigating the case too.

There seems to be mounting evidence that Jensen knew or should have at least suspected that Alice Huang’s company, Megaspeed, was involved in supplying advanced Nvidia chips to Chinese companies, either allowing them to access those chips in data centers it set up in Malaysia (which would have been legal under U.S. law prior to 2025 but not since then), or by smuggling Nvidia GPUs directly into China (which would have violated U.S. export laws). If Jensen was complicit in smuggling advanced Nvidia chips into China, that would probably be enough to force his resignation.

The New York Times also reported that Nvidia’s general counsel intervened in 2025 to get a Commerce Department investigation into Megaspeed moved to the Justice Department, where it seems to have languished. If more emerges showing that Jensen used his clout to shield Megaspeed from federal investigators, that too could make it difficult for him to stay on as CEO. And, if it turns out that he passed Alice Huang confidential information about discussions he had with President Trump or about Nvidia’s chip designs, then he is probably toast too.

If Alice Huang was, as some reporting suggests, primarily trying to influence Jensen’s position on export controls, then the picture may not be as bad for him. The Nvidia CEO’s stance on the export of chips has been remarkably consistent going back to a period before he is believed to have met Alice Huang. Plus, he didn’t exactly need convincing to oppose export regulation, given that Nvidia once made a lot of money selling its GPUs in China.

Jensen Huang’s position on exports to China hasn’t changed

When my colleague Beatrice Nolan and I sat down with Jensen in London a few weeks ago, before news of his affair broke, he told us the U.S. should want Nvidia to compete in the Chinese market. “I think the opportunity in China is large and growing, just like the AI market across the world is large and growing,” he said. “We could be of great service to the local market. We can compete and win as much of the AI market as possible for the United States, while also providing compute resources, which is really important to their economy.” Bea has more on what Jensen told us about China in a story she published earlier this week which you can read here.

If Jensen does end up stepping down, that could have ramifications for the entire AI boom, although it is likely that any successor would continue to try to provide financial support to the same cluster of AI labs, startups, and neoclouds that Nvidia has funded. That said, Jensen was a singular force. He ran Nvidia as if it were a giant startup, without a clear successor, or even clear candidates, and with a remarkably flat management structure. Jensen famously had between thirty—and at one point 60—direct reports. He had, by many accounts, a direct role in almost of all the business’s key decisions. If he goes, the company could stumble and its core business could suffer.

Even if Huang stays as CEO, U.S. AI regulation is now more likely 

Even if Jensen manages to hang on as Nvidia’s CEO, his influence on U.S. policy is likely to be greatly diminished. And that may have several key implications.

The most obvious is that the China hawks will have removed a key opponent of more restrictive export rules on selling chips to China. In fact, this may have been the main reason that someone has leaked the U.S. government’s conclusions about Jensen and Alice Huang at this time.

The less obvious policy implication is that Jensen has become a leading advocate for open source AI and was opposed to any government regulations that might impose licensing standards on these models that could effectively limit or eliminate their distribution in the U.S. With Jensen out of the way, those who believe open source AI models pose risks that cannot easily be mitigated will have a better chance at getting the federal government to enact licensing rules on all AI models.

We will have to see how this story develops over the coming days, weeks, and months. But one thing is clear: the AI boom’s biggest booster has been humbled.

With that, here’s more AI news.

Jeremy Kahn
jeremy.kahn@fortune.com
@jeremyakahn

Before we get to the news, just a reminder to check out our new vodcast, Fortune AI Weekly. This week, Beatrice Nolan and I talk to Ramp’s lead economist Ara Kharazian to find out what’s really going on with AI spending, ROI, and jobs. We also discuss the Democratic sweep in the midterm elections and its impact on AI. And we talk about Google’s controversial policy on schools using, or choosing not to use, its Gemini AI models. You can check out the vod here on YouTube.

FORTUNE ON AI

Anthropic sets IPO range valuing it at up to $1.4 trillion amid AI stock sell-off—by Beatrice Nolan

Nvidia built itself around one man. A scandal is testing that bet—by Sebastian Herrera

Anthropic debuts Claude Fable, Mythos 5.2 models, priced 20% below OpenAI’s GPT-6 Astra—by Emily Forlini

Before news broke of his affair with a suspected Chinese spy, Nvidia’s Huang told Fortune his company “could be of great service” to China—by Beatrice Nolan

Claude can now join video calls as an active, visible teammate in Anthropic’s latest push to reinvent white collar jobs beyond coding—by Emily Forlini

OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks—by Wen Shao

Exclusive: Publisher Spines raises $64 million in new funding to accelerate effort to automate book publishing with AI—by Jeremy Kahn

Exclusive: In the race for robot training data, ex-Meta researchers’ startup turns to gig workers with cameras—by Wen Shao

Exclusive: Alterra Labs announces $12.5 million seed round to build training environments for medical AI—by Wen Shao

AI IN THE NEWS

METR says Anthropic is closing in on RSI, raising AI safety alarms. Anthropic’s AI models are improving nearly twice as fast as they were previously, largely because AI is increasingly helping researchers develop more powerful models, according to a new study by independent AI evaluation nonprofit METR. The researchers found that Anthropic’s models have been advancing on benchmark tasks at roughly 1.9 times their previous rate since October 2025, estimating that AI’s contributions account for much of the acceleration rather than increased computing power. The findings provide some of the strongest evidence yet of AI accelerating its own development, potentially paving the way for a milestone known as “recursive self-improvement” (RSI) in which AI systems autonomously create increasingly capable successors. However, METR cautioned that AI is not yet capable of making major autonomous contributions to the most critical stages of model training, even as the accelerating pace of progress raises concerns about whether safety evaluations can keep up. The findings have alarmed AI safety researchers, who worry that AI-driven improvements could create a feedback loop of accelerating capabilities that outpaces efforts to keep increasingly powerful systems under control. But the report also raises questions about transparency: METR said Anthropic prevented it from publishing most of the evidence underpinning its conclusions, arguing that disclosure could help competitors accelerate their own AI development. Read more from the Verge here.

Anthropic strikes $3.4 billion deal for on-site gas generators for data centers. Anthropic has struck a $3.4 billion deal to acquire commitments for 1.7 gigawatts of gas-fired generators from Texas-based OnePWR Solutions, marking the largest known single purchase of power-generation equipment by an AI lab, the Financial Times reported. The deal gives Anthropic access to Caterpillar generators ordered two years ago, allowing it to bypass lengthy manufacturing delays and take greater control of the electricity supply needed for its rapidly expanding data centers. The purchase underscores how AI companies are increasingly moving directly into the energy business to secure scarce power infrastructure, as Anthropic prepares for its IPO and seeks to finance nearly $900 billion in computing commitments through 2045. But the purchase also raises environmental concerns, as the gas generators produce carbon dioxide and other pollutants, and their use is likely to increase grassroots opposition to these particular data center projects and data center construction more generally.

Microsoft winds down Oracle computing deal. Microsoft is winding down a deal to rent AI computing capacity from Oracle, declining to renew a contract for servers equipped with older Nvidia chips that no longer meet its technical requirements, Bloomberg reported. The move is part of a broader reassessment of Microsoft’s more than $80 billion in commitments to rent computing power from third-party cloud providers, including CoreWeave, Nscale, and Nebius, amid concerns about high costs and data center construction delays. The decision highlights growing uncertainty around the economics of the AI infrastructure boom, particularly for Oracle, whose aggressive data center investments have driven its free cash flow into negative territory and contributed to investor concerns.

Fireworks hacked. Fireworks AI, the fast-growing startup that helps developers run and train open-source AI models, suffered a major cyberattack that exposed customer billing information, model weights, and datasets. The breach, which Fireworks said has links to a cybercriminal group involved in stealing and reselling AI access tokens, has prompted some major customers to consider switching providers and could complicate the startup’s plans to raise funding at a valuation of $30 billion or more. This is the second breach Fireworks has suffered in the past month. You can read more here from The Information.

EYE ON AI RESEARCH

Google DeepMind shows AI agents can produce formal verifications of software. Formal verification is a way of mathematically proving software will function according to certain specifications. The process for this has existed for decades, but as software has become more complex, people have been able to afford, in terms of both time and cost, to formally verify only a tiny fraction of software. Generally speaking, only software that runs extremely critical systems where the consequences of an error can be fatal–such as commercial airline autopilots and some industrial control systems for oil refineries or drilling rigs, or some extremely critical banking software—has been formally verified. A lot of other, still very important software, has not been subjected to such rigorous checks because it is simply too complex, time-consuming, and expensive a process.

Now Google DeepMind has shown that a system built on various powerful AI models, including its most frontier AI model, Gemini Carbon, can complete a successful formal verification for a complex and important software algorithm. In this case, Google DeepMind used a system called AlphaProof Code to formally verify the reference implementation of the encryption protocol for Signal, the highly-secure messaging app. The same reference software also underpins encryption in WhatsApp, Messenger and Google Messages. The AI agents Google DeepMind used produced about 6.5 million lines of proof code in the Lean programming language, a way of expressing mathematics in code. The way Lean works is that is that if the mathematics is wrong, the code won’t properly compile. If it does, the mathematical steps are valid – which in this case also means that software adheres to whatever specifications are being tested and verified.

In this case, the proofs showed that Signal’s post-quantum handshake and its “Triple Ratchet” deliver forward secrecy, authentication and post-compromise security. Every proof is machine-checked against roughly 26,000 lines of specifications that humans reviewed and approved. To make sure the specifications would catch real flaws, the researchers reinserted two bugs in the Signal code that had previously been fixed, and the system proved the resulting code was insecure.

Formal verification has long been seen as the gold standard for secure software, but it has been too expensive and time-consuming to use widely: one formal verification of a complex code by humans took twenty years, another took six. 

“Six months ago, if you told me that somebody could do this, I would say that you were kidding me,” Swarat Chaudhuri, a Google DeepMind research scientist and University of Texas at Austin computer science professor, told Fortune.

Chaudhuri said the work matters more now that attackers are using powerful AI models to find and exploit software flaws. Proofs offer a way out of the “cat and mouse game” of finding bugs and patching them, because they rule out whole classes of errors at once. 

In Google DeepMind’s new AlphaProof Code system, humans and AI agents work together to write the specification, meaning the formal statement of what the code must guarantee. Teams of 40 to 80 autonomous prover agents then write the proofs over one to three days, and the proofs count only if Lean’s kernel checks them.

Chaudhuri said that the hard part of verification is moving from writing proofs to deciding what to prove, which still needs human expertise; the Signal effort, for instance, included cryptographers who helped write the specification that AlphaProof Code then checked. 

For now the project is purely research and there is no product plan, but the specifications and proofs are public, and Google DeepMind has been coordinating with the Signal Foundation. Chaudhuri compared the moment to deep learning around 2012: an old idea that “suddenly all of the ingredients are there” to make work. You can read more about the work on Google’s blog here.

AI CALENDAR

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Nov. 18-19: Slush, Helsinki, Finland.

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

What you can learn by having frontier models play ‘Blood on the Clocktower.’ Podcaster Dwarkesh Patel has created an unusual new benchmark for frontier AI models: getting them to play Blood on the Clocktower, a social deduction game similar to Mafia or Werewolf in which players must lie, bluff, form alliances, and uncover conspiracies. In a fascinating YouTube episode explaining the experiment, Patel describes how he built a computerized version of the game that allowed competing AI models to play hundreds of matches against one another, testing their ability to reason strategically, coordinate with teammates, and sustain elaborate deceptions over time. Patel said he spent approximately $20,000 in AI API credits developing the system and running the experiments, with Claude doing much of the coding and autonomously refining the game environment through more than 20 rounds of testing.

The results suggest that frontier models have become remarkably sophisticated at deception and collective reasoning, with OpenAI’s GPT-6 Astra emerging as the strongest player among the models tested. Patel found that OpenAI’s latest models were particularly good at detecting deception because they were more skeptical of individual pieces of evidence, whereas Anthropic’s models tended to accept information too readily. But the experiment also exposed continuing weaknesses in AI models’ long-term strategic planning and reasoning. Overall, Patel concluded that the AI-assisted process of building and refining the benchmark itself—which included coding the harness the AI models used—demonstrated the capabilities of frontier AI models perhaps better than their actual game play did. He said this demonstrated how AI models could be used to buidl simulated environments that could then be used to train future AI systems—an example of the steps toward recursive self-improvement, or RSI, that frontier labs such as Anthropic, OpenAI, and Google DeepMind have begun talking about.