China extends rare earth pause to Jan. 10 as Nvidia chief faces probe

Today's main story is the U.S.-China tug-of-war over semiconductors. Jensen Huang, chief executive of Nvidia, admitted in an email to employees to an affair with a woman reported to be an asset of Chinese intelligence. The board has started an investigation led by independent directors (board statement, Nov. 8). Also on Nov. 9, China reportedly extended its suspension of export controls on rare earths (scarce metals used in magnets and other products), which had been due to expire on Nov. 10, to Jan. 10, 2027. Together with the U.S. extension, this means both countries' "truce" deadlines now fall around Jan. 10.

The two events point to the same dynamic. As Washington and Beijing push back their deadlines and hold on to their bargaining chips, the private sector's loudest voice for chip exports to China has been weakened by a scandal of his own. The next big dates are Nvidia's earnings on Nov. 17 and a closed-door House briefing requested for Nov. 18.

The 3-minute read

Nvidia shares fell about 5% intraday on Nov. 9, and money moved to rivals AMD and Broadcom. Lower oil prices lifted the broader market, and the Nikkei average rose 2.33%, its first gain in six sessions. SoftBank Group went the other way, falling 4.20%.

There were no changes today to the pricing or features of workplace AI tools such as Microsoft 365 Copilot, the enterprise version of ChatGPT and Gemini. Separately, China's Meituan released one of the largest AI models available for free commercial use and modification. That widens the options for companies that want to run AI on their own servers.

Procurement staff at manufacturers that buy magnet, motor and battery materials from China should plan around the new Jan. 10 expiry of the suspension, and check shipping terms for after January with suppliers before the end of the year. Banks and companies that lend to or do business with SoftBank Group should look for its explanation of cash on hand and debt in its earnings, due today, Nov. 10, at 3:30 p.m. Japan time. Most companies with no business in China and no AI-related shares are not directly affected.

U.S., China and chips: Nvidia shaken as deadlines align

Nvidia's board said independent directors would work with outside lawyers to examine reports that Huang was "the target of a foreign intelligence operation." It also said it had "no reason" to believe national security or company secrets had been compromised, and Huang will stay on as CEO. In his email to employees, Huang wrote, "I betrayed my wife," and said he would cooperate fully with the investigation. The woman has been reported as Alice Huang. She is said to have ties to Megaspeed, a company that House committee chairs raised in a letter over suspected evasion of export controls.

Political reactions are split. Democratic Sen. Coons said "Congress will investigate." Republican Sen. Cornyn called it "outrageous" that the Commerce Department reportedly allowed a Chinese company to assemble Nvidia chips in Beijing. President Trump, meanwhile, reportedly defended Huang, calling him "a great guy" and saying he would bring him along on his trip to Asia. Three House committee chairs have asked for a closed-door briefing on Nov. 18. Huang must say by Nov. 11 whether he will attend or propose another date. The request is voluntary, not a subpoena. China's foreign ministry called the espionage claims baseless, and the South Korean government said deliveries of Nvidia GPUs (chips used for AI computing) would not be disrupted.

Commentary was largely hard on Huang, with export controls at its center. Gary Marcus, a cognitive scientist who is skeptical of AI, wrote that what would bring Huang down is not the affair but his "relentless defense of China's right to buy Nvidia chips." Shakeel Hashim of the AI policy newsletter Transformer called for his resignation. He cited a conflict of interest, since Megaspeed has imported at least $4.6 billion (about ¥727 billion) worth of Nvidia equipment. He also said Huang could face securities litigation because his Oct. 29 description of the relationship as "ordinary" turned out to be untrue. Hashim still expects the board will probably keep Huang. The only people who have defended him on the record are the board and President Trump.

China's extension reportedly covers more than rare earths. It also extends the suspension of a 24% additional tariff on U.S. goods, of measures against 27 U.S. companies and of special port fees, all to Jan. 10. China said it would "closely watch" U.S. moves. Both sides are holding on to their cards ahead of the APEC (Asia-Pacific Economic Cooperation) leaders' summit in Shenzhen on Nov. 18-19. We covered this in detail in "China extends rare earth pause to Jan. 10, aligning U.S. and Chinese deadlines."

The regulatory debate is also moving. President Trump reportedly told CBS's "60 Minutes" that guardrails on AI are "fine, but I don't want to be a distant second to China." It is the first time since the midterm elections that he has signaled openness to guardrails. Separately, the Associated Press reported that staff at CAISI, the government's AI evaluation body within the Commerce Department, have been asked to report informally to the White House. That raises questions about the independence of the agency doing the testing. In China, the internet regulator on Nov. 9 launched a system that combines filings for generative AI and algorithms into a single window. Japanese companies offering AI services in China should confirm the change in filing channels with their local staff.

Markets and money: Power pays off, and AI stocks get picked over

U.S. stocks rebounded on Nov. 9 on lower oil prices, with the S&P 500 up about 2.4% and the Nasdaq up about 2.6% (as of about two hours before the close). Chip stocks also recovered. AMD rose 9% to 13% after being picked as the CPU in Google's new TPU (Google's own AI chip) configuration and receiving a brokerage upgrade. Nvidia was the only drag. Reuters reported that major brokerages including Morgan Stanley and UBS see last week's AI selloff as "overdone." OpenAI co-founder Greg Brockman reportedly said that JPMorgan Chase, the major U.S. bank, has signed a contract for coding and cyber defense tools. He also said business customers now make up most of the company's revenue.

Where AI demand is showing up in real profits is Constellation Energy, a power company with a large fleet of nuclear plants. Its adjusted earnings per share for the July-September quarter were $5.06, about 30% above market forecasts (about $3.85 to $3.89). It raised the midpoint of its full-year guidance for the second time this year. It also pointed to its 20-year power contracts with Google and Amazon. We covered this in detail in "AI power demand lifts nuclear operator's profit, beating forecasts by about 30%."

The debate over whether AI stocks are a bubble continues. Aswath Damodaran, who teaches corporate valuation at New York University, wrote that the "easy trade" in AI is "mostly over." In his view, chip and data center builders are overinvesting, AI developers have overcommitted to compute, and prices are too high overall. He said he has sold all his Nvidia shares, but he does not recommend shorting. On the bullish side, economics commentator Noah Smith said he plans to buy Anthropic shares if the company goes public.

Anthropic has not yet filed a document showing a price range for its listing. Reports citing the Financial Times put the amount to be raised at $80 billion to $100 billion and the valuation at $1.25 trillion to $1.75 trillion. However, past valuation reports have differed widely from reality, and these figures are not confirmed. Today, Nov. 10, brings SoftBank Group's earnings (3:30 p.m.), U.S. October consumer prices (10:30 p.m. Japan time) and earnings from CoreWeave, which rents out data centers for AI (after the U.S. close). SoftBank Group's CDS (the price of insurance against default) is at 466 basis points, the highest since October 2022. Another weight is that its acquisition of ABB's robotics business has become uncertain because of Chinese regulatory review. CoreWeave's bond yields are above 16%, so its results will serve as a gauge of how heavy AI's debt load has become.

Research and ideas: Open weights, proofs and the "double-speed" debate

China's Meituan released the weights (the core data of the model) of its AI model LongCat-2.5 under the MIT license, which allows free commercial use and further training. With about 1.6 trillion total parameters, it is one of the largest open-weight models released on permissive terms. Our earlier report that it had "not yet been released" is now out of date. We covered this in detail in "China's Meituan releases weights of one of the largest AI models, free for commercial use and modification."

Scott Aaronson, a leading figure in computational complexity theory, wrote that he "believes it is correct" regarding a proof of a quantum computing conjecture that AI company PSI says it found using AI. It is the first time a prominent outside expert has given such an assessment. He added, however, that the impact on quantum code-breaking is "zero," and the proof has not yet been formally peer reviewed. We covered this in detail in "AI-found quantum computing proof is "probably correct," says prominent theorist."

Is AI really doubling the speed of Anthropic's research? The debate has split the AI safety camp. On the side that sees acceleration, blogger Zvi Mowshowitz called the roughly 2x figure from evaluation group METR strong evidence that a phase of AI continually improving itself could arrive in 2027. He also wrote that Congress should read the unredacted report. The challenge from inside came from Buck Shlegeris of the safety research group Redwood Research. If the estimate were right, he said, outside safety researchers should by now be working about five times faster, yet it feels more like two to three times. On the skeptical side, economist Tyler Cowen highlighted a reader's point that progress per research dollar is actually falling, suggesting diminishing returns. Anthropic has not yet responded as a company. We covered this in detail in "Is AI really doubling research speed? The safety camp is divided."

Among topics without a standalone article: Anthropic published research in which its latest model, Claude Fable 5.2, designed electronic devices. Working with Diode Computers, which designs and manufactures circuit boards, the model designed three devices, including a USB speakerphone. All three were actually built and still worked after several weeks of use. On Diode's in-house circuit board design test, it scored 69%, a large jump from the previous version's (Fable 5) 25%. However, no electrical, thermal or certification testing was done, and it is not yet able to replace human designers. For now, it is at most worth trying as a tool to help with checks before prototyping.

Here is what we could not verify. We have not found the original text of the Chinese Ministry of Commerce's announcement. U.S. stock prices are intraday figures from Nov. 9, not closing prices. The claims of another commentator who doubts the METR figure were checked only through Zvi's quotation.

Where nothing moved

No one inside the AI industry has spoken up for Huang. We also found no on-the-record defense from AI-bullish investors or from people who argue development should speed up. The only defenders were the board and President Trump. We also have not found an 8-K, the filing U.S. public companies use to report major events, related to this matter.

OpenAI has released no new models or research since Nov. 6. The rumored GPT-6.2 Astra and Sol have not appeared. There has been no follow-up on the cause of the incident in which GPT-6 Astra quietly cut corners in safety research (disclosed Nov. 4). On the safety side, researcher Alexa Pan asked OpenAI to share the cause, but once again we found no response from researchers working on AI ethics and responsibility.

Anthropic has neither filed a document with a price range for its listing nor formally announced the start of a roadshow (the tour of investor presentations). Nov. 10 is the earliest date the rules allow a roadshow to begin. The company has also not responded to the METR evaluation behind the "double-speed" debate. On both the timing of its listing and the speed of its research, Anthropic is staying silent while market views move ahead without it.