Welcome to “Eye on AI”. I am Jeremy Kahn. This episode includes:

• NVIDIA was built around one person, and now a spy scandal threatens this model.
• Before the scandal came to light, NVIDIA CEO Jensen Huang gave an interview to Fortune regarding Chinese policies.
• Anthropic is considering pricing its IPO at a valuation of $1.4 trillion.
• Anthropic launches a new Fable model, which is 20% cheaper than OpenAI.
• OpenAI releases a new cybersecurity model.
• Anthropic is closer to RSI, which worries the security experts at AI.
• AI Infrastructure provider Fireworks suffers a hacker attack.
• Google DeepMind has made a breakthrough in formal verification.
• Podcast host Dwarkesh Patel has proposed an interesting new AI benchmark.

Before we dive into the news, I want to say that you definitely shouldn't miss this year's Fortune Brainstorm AI San Francisco.

We have prepared an outstanding lineup of speakers for everyone: President OpenAI, Greg Brockman, and Microsoft's AI CEO Mustafa Suleyman will all be on stage. Two companies that were once closest partners are now increasingly becoming competitors, so I'm very interested in hearing their different views on the future path ahead. AI pioneers, Fei - Fei Li (currently World Labs), as well as Andrew Ng (currently LearnVector), will also discuss their perspectives on the future of AI. Palo Alto Networks CEO Nikesh Arora will talk about AI and cybersecurity. Greylock partners Reid Hoffman, CapitalG's Jill Chase, as well as Kleiner Perkins's Mamoon Hamid and Ilya Fushman will discuss where the funds are flowing. Recursive Superintelligence's Richard Socher will talk about RSI and automation AI research.

In addition, there are also participants from Google such as Nick Fox, Fireworks CEO Lin Qiao, Cerebras with Andrew Feldman, Vanta with Christina Cacioppo, Unconventional AI with Naveen Rao, Brex with Pedro Franceschi, Maven Clinic with Kate Ryder, Box with Aaron Levie, Atlassian with Mike Cannon to Brookes, New Work Foundation with Clara Shih, Twilio with Khozema Shipchandler, Rubrik with Bipul Sinha, Harvey with Winston Weinberg, Dynamic Creatures with Marc Theermann, SendCutSend with Jim Belosic, Apptronik with Jeff Cardenas, Accenture Song with Ndidiamaka Oteh, and Senra Systems with Jordan Black. The event will be very exciting. Welcome to attend at the San Régis Hotel in San Francisco from December 7th to 8th, together with me and several colleagues from Fortune. You can apply to attend here.

The biggest news in the AI sector this week is the exposure of an extramarital affair between NVIDIA CEO Jensen Huang and a woman, who is reportedly considered by U.S. intelligence agencies to be a Chinese spy. Whether Jensen Huang will be able to retain his position remains uncertain at this time. It largely depends on how much Jensen Huang actually knows and what information he has disclosed to Alice Huang, who is at the center of the scandal. As more information comes to light, the situation for this embroiled NVIDIA CEO seems less than favorable. In addition to NVIDIA's board of directors announcing an independent investigation, Republicans in the House of Representatives have requested that he testify at a closed-door hearing next week. President Trump initially defended Jensen Huang, but now indicates that the FBI is also investigating this case.

It seems that there is an increasing amount of evidence indicating that Jensen Huang knew, or at the very least should have suspected, that Megaspeed, a company owned by Alice Huang, was involved in supplying advanced NVIDIA chips to Chinese companies. This may have been done by allowing these companies to access the chips through data centers established in Malaysia (which might have been legal under U.S. law before 2025, but no longer so thereafter), or by directly smuggling NVIDIA GPU to China (which would violate U.S. export laws). If Jensen Huang was indeed involved in smuggling advanced NVIDIA chips to China, this could very well be sufficient to force him to resign.

The New York Times also reported that in 2025, NVIDIA's chief legal counsel intervened to facilitate the transfer of the investigation into Megaspeed from the Department of Commerce to the Department of Justice, but this investigation seems to have been put on hold thereafter. If further revelations show that Jensen Huang used his influence to shield federal investigators regarding Megaspeed, then it will become difficult for him to continue serving in CEO. If it is proven that he disclosed confidential information discussed with President Trump, or confidential information regarding NVIDIA's chip design, to Alice Huang, then it is also very likely that he will not be able to escape responsibility.

If Alice Huang is, as some reports suggest, mainly trying to influence Jensen Huang's stance on export controls, then the situation might not be so bad for him. NVIDIA CEO's position on chip exports was quite consistent long before he was considered to know Alice Huang. Moreover, considering that NVIDIA has made substantial profits by selling GPU to China, he wouldn't have needed anyone to persuade him to oppose export restrictions in the first place.

Jensen Huang's stance on exporting to China has not changed.

A few weeks ago, before the scandal came to light, I met with Jensen Huang in London along with my colleague Beatrice Nolan. During that meeting, he told us that the United States should hope for NVIDIA to compete in the Chinese market. He said, "I believe the opportunities in China are both vast and growing, just like those in the global AI market. We can provide significant support for the local market. We can compete and strive to win as much of the US AI market share as possible, while also providing computing resources, which is very important for their economy." In a report published earlier this week, Bea further elaborated on Jensen Huang's views on China at that time, which you can read here.

If Jensen Huang ultimately resigns, it could have an impact on the entire AI craze. Although any successor is likely to continue providing financial support to the same group of AI laboratories, startups, and neocloud companies that NVIDIA does, Jensen Huang is a unique driving force. He has managed NVIDIA as if it were a giant startup, with no clear successor in sight, not even a definite candidate, and the management structure is extremely flat. It is said that Jensen Huang once had 30 direct subordinates, which at one point peaked at 60. He was directly involved in almost all key business decisions in many aspects. If he leaves, the company could be impacted, and its core business might also suffer.

Even if Jensen Huang continues to hold the position of CEO, US AI regulation is also likely to become stricter.

Even if Jensen Huang ultimately retains his position as NVIDIA CEO, his influence over U.S. policy could significantly weaken. This could have several key implications.

The most direct impact is that those in China who are more hawkish will lose an important opponent to their stance against stricter chip export regulations. In fact, this may well be the main reason why someone chose to reveal the U.S. government's conclusions regarding Jensen Huang and Alice Huang at this time.

Another less obvious policy impact is that Jensen Huang has become an important advocate for the open-source AI. He opposes any government regulation that may impose licensing standards on these models, as this could actually limit or even end their distribution in the United States. With Jensen Huang now excluded, those who believe that the open-source AI models pose risks that are not easily mitigatable will have a greater chance of pushing the federal government to establish licensing rules for all AI models.

We still need to continue observing how this story develops over the next few days, weeks, and months. But one thing is clear: the biggest proponents of the AI craze have been frustrated.

Next are more AI news.

Jeremy Kahn

jeremy.kahn @ fortune.com

@ jeremyakahn

Before we dive into the news, let me remind you again that you're welcome to watch our new video podcast, "Fortune AI Weekly". This week, I had a conversation with Beatrice Nolan and Ramp along with Ara Kharazian, the chief economist, to discuss the real situation regarding AI spending, return on investment, and employment. We also talked about the Democratic Party's overall victory in the mid-term elections and its impact on AI, as well as Google's controversial policy regarding whether schools should use or not use its Gemini AI model. You can watch this episode of the video podcast on YouTube.

FORTUNE ON AI

• Anthropic set a IPO range during the mass selling of AI shares, with a valuation as high as 1.4 trillion US dollars – Beatrice Nolan
• NVIDIA was built around one person. A scandal is now testing this bet – Sebastian Herrera
• Anthropic launches the Claude Fable and Mythos 5.2 models, with a pricing 20% lower than that of OpenAI's GPT-6 Astra —Emily Forlini
• Before the scandal came to light, NVIDIA's Jensen Huang told Fortune that his company "could provide great help to China" – Beatrice Nolan
• Claude Nowadays, it is possible to join video conferences and act as an active, visible team member, reflecting Anthropic's latest attempt to reshape white-collar work beyond traditional coding methods – Emily Forlini.
• OpenAI launches a new cybersecurity AI model to address the growing concerns about AI-driven cyberattacks – Wen Shao
• Exclusive: Publisher Spines Secures $64 Million in New Financing to Accelerate Automated Book Publishing with AI – Jeremy Kahn
• Exclusively: In the robotics training data competition, a startup founded by former Meta researchers has turned to employing gig workers equipped with cameras – Wen Shao
• Exclusive: Alterra Labs Announces Completion of $12.5 Million Seed Round Financing to Build Medical AI Training Environment – Wen Shao

AI News

METR claims that Anthropic is approaching RSI, triggering a AI security alert. According to a new study by the independent AI evaluating the non-profit organization METR, the AI model's improvement speed has nearly doubled compared to before, mainly due to AI increasingly helping researchers develop more powerful models. Researchers found that since October 2025, the progress of the Anthropic model on benchmark tasks has been about 1.9 times faster than before, and they estimate that this acceleration is largely due to the contributions of AI itself, rather than an increase in computing power. This study provides one of the strongest pieces of evidence to date indicating that AI is accelerating its own development, potentially paving the way for a milestone known as "recursive self-improvement" ( RSI ), where the AI system autonomously creates increasingly powerful successor systems. However, METR also reminds us that AI is still unable to make significant autonomous contributions at the most critical stages of model training, and although the acceleration in progress is concerning, it is uncertain whether security assessments can keep up. The research results have raised concerns among AI security researchers, who worry that the improvements driven by AI may form a feedback loop of accelerating capabilities, thereby surpassing human efforts to control increasingly powerful systems. But this report has also raised questions regarding transparency: METR states that Anthropic has prevented the publication of most of the evidence supporting these conclusions, on the grounds that disclosure could help competitors accelerate their own AI development. For more information, see The Verge.

Anthropic Completes $3.4 Billion Deal for On-Site Gas Generators for Data Centers. According to a report by The Financial Times, Anthropic has completed a $3.4 billion deal to secure a purchase commitment for 17 gigawatts of gas generators from OnePWR Solutions, which is headquartered in Texas. This is the largest single-order purchase of laboratory power generation equipment known to date for AI. The deal enables Anthropic to obtain the Caterpillar generators it ordered two years ago, thereby bypassing long manufacturing delays and gaining better control over the power supply needed for its rapidly expanding data centers. This move highlights that as Anthropic prepares to go public and seeks financing for its commitment to deliver nearly $900 billion in computing power by 2045, AI is increasingly entering the energy business directly to ensure access to scarce power infrastructure. However, this purchase has also raised environmental concerns, as gas generators emit carbon dioxide and other pollutants, and their use is likely to intensify local community opposition to these data center projects and broader data center construction.

Microsoft is gradually ending its computing power transactions with Oracle. Bloomberg reports that Microsoft is phasing out a deal to rent computing power from Oracle through AI, and will not renew a contract that uses servers equipped with older NVIDIA chips, as these chips no longer meet its technical requirements. This move is part of a broader re-evaluation by Microsoft of its commitment to lease computing power from third-party cloud service providers for over $80 billion, including CoreWeave, Nscale, and Nebius. The reason is concerns about high costs and delays in data center construction. This decision highlights that the economic viability of the AI infrastructure boom is facing increasing uncertainties, especially for Oracle, whose aggressive data center investments have pushed its free cash flow into negative territory, sparking concerns among investors.

Fireworks has been attacked by hackers. Fireworks AI is a rapidly growing startup that helps developers run and train open-source AI models. The company has suffered a major cyberattack, with customer billing information, model weights, and datasets being leaked. Fireworks indicates that this intrusion is related to a cybercrime gang involved in stealing and reselling AI access tokens. The incident has prompted some major customers to consider switching suppliers and may also complicate the startup's plans to raise funds at a valuation of $30 billion or more. This is the second intrusion Fireworks has experienced in the past month. For more information, see The Information.

AI Research Observation

Google DeepMind demonstrates that AI agents can generate formal verifications for software. Formal verification is a method of proving through mathematical means that software will operate in accordance with certain specifications. This process has existed for decades, but as software has become increasingly complex, only a very small portion of software can undergo formal verification in terms of time and cost. Generally speaking, only those software systems that run extremely critical functions and where the consequences of errors could be fatal are subject to formal verification—such as the autonomous flight control systems of commercial aircraft, the industrial control systems of certain refineries or drilling platforms, or some highly critical banking software. Many other equally important software applications do not undergo such rigorous checks due to the complexity, time-consuming nature, and high cost of the process.

Now, Google DeepMind has demonstrated a system composed of various powerful AI models, including its state-of-the-art AI model Gemini Carbon, which is capable of successfully completing the formal verification of a complex and important software algorithm. In this case, Google DeepMind used a system named AlphaProof Code to perform formal verification on the reference implementation of the encryption protocol for the highly secure instant messaging application Signal. The same reference software also supports the encryption in WhatsApp, Messenger, and Google Messages. Google DeepMind utilized AI to generate approximately 6.5 million lines of proof code written in the Lean programming language, which is a method of expressing mathematics through code. The working principle of Lean is as follows: if there are mathematical errors, the code cannot be compiled correctly; if it can compile successfully, then the mathematical steps are valid – which in this context also means that the software complies with the specifications being tested and verified.

In this case, these proofs demonstrate that the post-quantum handshake of Signal and its “Triple Ratchet” mechanism are capable of providing forward secrecy, authentication, and post-leakage security. Each proof was checked by machines against approximately 26,000 lines of specifications that had been manually reviewed and approved. To ensure that these specifications could detect real defects, researchers reinserted two vulnerabilities that had previously been fixed in the Signal code, and the system proved that the modified code was insecure.

Formal verification has long been regarded as the gold standard for secure software, but it is too expensive and time-consuming to be widely used: it once took 20 years to perform formal verification on a piece of complex code, and another time it took 6 years.

Google DeepMind, a research scientist and professor of computer science at the University of Texas at Austin, Swarat Chaudhuri, told Fortune: "If you had told me six months ago that someone could do this, I would have said you were joking."

Chaudhuri indicates that this task is particularly important at a time when attackers are leveraging powerful AI models to search for and exploit software vulnerabilities. This approach proves to offer a way out of the cat-and-mouse game of "finding vulnerabilities – fixing vulnerabilities," as it allows for the elimination of entire categories of errors in one go.

In the new AlphaProof Code system of Google DeepMind, humans and AI agents jointly write the specifications, which are formal statements regarding what must be ensured of the code. Subsequently, a team composed of 40 to 80 autonomous proof agents completes the verification within one to three days. Only when the kernel check of Lean is passed do these proofs become valid.

Chaudhuri says that the most difficult part of verification is not writing the proof, but deciding what to prove, which still requires human expertise; for example, the Signal project includes cryptography experts who helped write the specifications that AlphaProof Code then checks.

Currently, this project is purely for research purposes and there are no product plans. However, the specifications and proofs are all made public. Google DeepMind has also been coordinating and collaborating with Signal Foundation throughout this process. Chaudhuri compares this moment to the era of deep learning around 2012: an old idea "suddenly had all its elements in place," and thus it truly began to have an impact. For more information, please refer to the blog of Google.

AI Schedule

• November 16th to 17th: Fortune 500 Innovation Forum, Detroit. Applications are open for participation.
• November 18th to 19th: Slush, Helsinki, Finland.
• December 6th to 12th: NeurIPS Conference, Sydney, Australia.
• December 7th to 8th: Fortune Brainstorm AI, San Francisco. Applications are open for participation.

Brain Power Replenishment

What can be learned by having cutting-edge models play “Blood on the Clocktower”? The podcaster Dwarkesh Patel designed an unusual new benchmark for these models: to have them play “Blood on the Clocktower”, a social deduction game similar to “Mafia” or “Werewolf”. In it, players must lie, bluff, form alliances, and expose conspiracies. In an exciting episode of the show, Patel explained how he built a computer version of the game that allowed competing AI models to play hundreds of games against each other to test their strategic reasoning, ability to collaborate with teammates, and their capacity to maintain complex deceptions over the long term. Patel said that he spent about $20,000 on developing the system and running the experiments, with Claude doing most of the coding work, and they independently improved the game environment through more than 20 rounds of testing.

The results show that cutting-edge models have become quite complex in terms of deception and collective reasoning, among which the GPT-6 Astra of OpenAI performed the strongest in the test models. Patel found that the latest models are particularly adept at detecting deception, as they are more skeptical of individual pieces of evidence, whereas the models of Anthropic tend to be overly trusting of information. However, the experiments also revealed that AI models still have shortcomings in long-term strategic planning and reasoning. Overall, Patel believes that the process of AI assisting in building and refining this benchmark—including writing the framework used by the models—is a better reflection of the capabilities of these cutting-edge AI models than the actual model interactions themselves. He says this indicates that AI models can be used to create simulation environments, which can then be used to train future AI systems—this is one of the steps of recursive self-improvement (RSI) that cutting-edge laboratories such as Anthropic, OpenAI, and Google DeepMind have begun to discuss.