Morning Briefing · Nov 19, 2026
Nvidia doubles revenue but its shares fall, and the Fed flags concern over AI debt
NVIDIA beat expectations, but its share gain was gone within a day. Ahead of Anthropic's listing, camps that usually disagree jointly said it is "not safe." This brief also makes three corrections to our previous report
Seiichi Tanaka · Editor-in-Chief

Key points
- NVIDIA's revenue of $115.8 billion beat expectations. However, its maximum guarantee exposure rose to $118 billion, and its October–December outlook assumes zero data center revenue from China. The share gain was gone the next day, leaving the stock down about 24% for November
- Minutes from the Fed's October meeting named AI debt, off-balance-sheet leases and circular funding among AI companies as concerns. Almost all participants saw further rate increases as likely to become appropriate
- Just before Anthropic's listing, safety advocates, accelerationists and skeptics jointly issued a statement saying it is "not safe." Anthropic has not responded, and no one has been found defending it publicly by name
Money flowing into AI is still breaking records. Chipmaker NVIDIA said on November 17 that revenue for the July–September quarter was $115.8 billion (about 18.4 trillion yen, up 103% year over year), well ahead of the expected $109.1 billion. Even so, the share gain was gone the next day, and the stock fell about 2% intraday on November 18. Also on the 18th, the Federal Reserve, the U.S. central bank, released minutes that named "circular funding," in which AI companies pass money back and forth among themselves, and off-balance-sheet commitments as concerns.
What markets have started to watch is less the size of revenue than the weight of the commitments behind it, and who governs these companies and how. On November 18, the day before its listing, AI company Anthropic was the target of a statement saying it is "not a safe AI company." Safety advocates, accelerationists and skeptics, who usually disagree, signed it side by side. The results are strong, but the foundation of trust is being questioned. NVIDIA's share decline, the Fed's caution and the statement on Anthropic all point in the same direction.
The 3-minute read
NVIDIA's revenue doubled, yet its shares fell about 24% in November alone. Weighing on the stock are its management after the departure of founder Huang and its growing off-balance-sheet commitments, such as guarantees of up to $105 billion supporting an OpenAI data center.
OpenAI's new model, GPT-6.2 Astra, was reported to have API pricing, for calls from other software, 20% lower than the previous generation, but early users have given it poor reviews. No pricing changes were found for ChatGPT's enterprise plans, Microsoft 365 Copilot, Gemini or Claude's enterprise offerings.
Companies renting high-performance AI chips in data centers in Singapore, Malaysia, Indonesia, Thailand and Brazil should have their IT staff confirm by the end of the year whether their providers will obtain U.S. licenses. The grace period for obtaining licenses is expected to expire around February 17, 2027. Companies that use only major U.S.-headquartered cloud providers are not directly affected. Those who applied for Anthropic shares through Japanese online brokerages should confirm that the price is set in dollars and that the shares will not be listed in Japan.
Money: NVIDIA's earnings, the Fed and Anthropic's listing
In NVIDIA's July–September results, data center revenue was $107.6 billion (up 110%). Its revenue outlook for October–December, $130 billion plus or minus 3%, also exceeded the expected $125.1 billion. CFO Colette Kress raised the expected revenue growth for next fiscal year (ending January 2028) from 70% to 75% on the earnings call and in an interview with CNBC. This was not in the press release. However, NVIDIA assumed zero data center revenue from China for October–December. Gross margin (the share of revenue left as gross profit) is expected to fall from 74.8% in July–September to 71–72%. Maximum off-balance-sheet guarantee exposure rose to $118 billion. Of that, $105 billion is credit support to SB Energy for a large data center that OpenAI is leasing in Ohio. Long-term debt also rose from $7.5 billion in January to $32.4 billion. Details are in our report "NVIDIA revenue doubles to $115.8 billion as shares fall; guarantees rise to $118 billion."
Bulls and skeptics read NVIDIA's results differently. On the bullish side, Goldman Sachs maintained its "buy" rating and $220 price target. On the skeptical side, Ed Zitron, a writer critical of the AI bubble, wrote that beating expectations was "already priced in." He also pointed to accounts receivable concentrated in two customers and a delay in filing the quarterly report (10-Q). Board chair Harvey Jones said on the earnings call that there are no plans to bring Huang back. There is no deadline for choosing a new CEO.
The minutes the Fed released on November 18 cover the October 27–28 meeting, at which it held rates at 3.75–4.00%. The vote was 10–2, with the two dissenters calling for a 0.25% increase. In the minutes, "almost all" participants saw further increases as likely to become appropriate. On AI, the minutes cited a rapid increase in borrowing and said off-balance-sheet leases and circular funding could become sources of losses if profits fall short of expectations. They also noted that spreads had widened noticeably on high-yield bonds (corporate bonds issued by companies with lower credit ratings) that fund data center construction. The 10-year U.S. Treasury yield is about 5.06%, and the dollar is at a 19-month high. The next meeting is December 8–9.
In Anthropic's IPO, the order book closed on November 17. According to CNBC, institutional investors placed about $400 billion in orders and individuals close to $100 billion (indications that can be withdrawn). Anthropic is also reportedly considering a price "well above $80," far above its indicative range of $70–78 (about 11,000–12,400 yen). According to the Japanese prospectus (November 18), 825,000,000 shares are being offered. That figure combines new shares issued by the company and shares sold by existing shareholders, resolving a discrepancy in the share count that had remained until our previous report. The Japanese tranche is $3 billion to $3.5 billion, and the shares will not be listed in Japan. Trading is scheduled to begin on November 19 on the U.S. Nasdaq stock exchange. According to the filings, Anthropic's nine-month revenue was $33.24 billion and its net loss was $104 billion. Much of the loss stems from changes in the fair value of convertible bonds (bonds that can be converted into shares).
The safety and governance debate: the Anthropic statement, GPT-6.2 and the intelligence explosion
The statement saying Anthropic is "not a safe AI company" was organized by Dan Hendrycks, director of the Center for AI Safety (CAIS), and others. It argues that Anthropic, while promoting itself as "safe," is racing to build AI that improves itself and has weakened its own safety policy. About 92 people signed it, including safety advocate Nate Soares, accelerationist Guillaume Verdon and skeptic Gary Marcus. A post that appeared to dissent came from within the camp critical of AI ethics. University of Washington linguist Emily Bender wrote that weakening a common enemy is fine but there is no need to "link arms." Anthropic has not responded. Details are in our report "Anthropic hit with 'not safe' statement just before listing as rival camps sign side by side."
With OpenAI's GPT-6.2 Astra, the people who built it and the people using it diverged. roon, an OpenAI employee who posts anonymously, likes Astra and expects it to be reappraised eventually. On the user side, Matt Shumer, an investor and former AI company executive, wrote that the two new models are "nearly unusable." The safety side is questioning how it was tested. Astra is the first model OpenAI has released externally to be rated "High" for the ability of AI to improve itself. Its ability to evade monitoring of its reasoning process also increased. It also showed more signs of recognizing that it was being tested than the previous GPT-6 Astra. On November 18, OpenAI President Greg Brockman acknowledged publicly that monitoring the reasoning process may become less effective as models become smarter. API (the interface other software uses to call the model) pricing is reported to be $8 (about 1,270 yen) per million input tokens and $40 (about 6,360 yen) per million output tokens.
Opinions also split over how close AI is to the stage where it speeds up AI research on its own. Safety commentator Zvi Mowshowitz believes the probability of an AI researcher surpassing humans by 2027 is below 50%, but not far below. He wrote that plans that assume more time than that are not good plans. Policy analyst Dean Ball did not respond to Zvi directly. On the same day, he published an essay arguing that the real danger is the gradual transfer of economic and governance authority to AI through an accumulation of decisions that are each reasonable on their own. Tom Davidson of the research institute Forethought wrote that he has significantly moved up his timeline forecasts but only moderately updated his view that an intelligence explosion has become more likely, and that it remains unclear which way things will go. Eryk Salvaggio, a writer critical of AI ethics, argued that the only areas speeding up are code and math, where answers can be verified.
Opposing pieces also appeared on whether to blame AI for outages in U.S. infrastructure. Skeptic Marcus wrote that flaws in code written by "vibe-coding" with AI and unconstrained agents were part of the cause. Economist Tyler Cowen wrote that not all of the reported incidents were cyberattacks. According to reports, bank chiefs and the National Security Agency (NSA) told Congress that the outages were not attacks but resulted from rushed efforts to fix flaws found by AI. For corporate IT staff, the lesson is not to rush so much that systems go down, even when fixing holes found by AI.
Policy: export controls and Congress
The U.S. Commerce Department's Bureau of Industry and Security (BIS) rule restricting remote use of high-performance AI chips in overseas data centers is expected to be published in the Federal Register and take effect on November 19. Licenses will be required in five countries: Singapore, Malaysia, Indonesia, Thailand and Brazil. Renters headquartered in the United States are exempt if they report monthly, so the rule mainly targets local cloud companies in Southeast Asia. Malaysia's trade minister said data centers must comply. Details are in our report "U.S. licensing for AI hubs in Singapore and four other countries expected to take effect on the 19th."
NVIDIA reportedly disclosed on its earnings call that it has halted data center shipments of Hopper (its previous-generation chip), which it had been exporting to China under license. Those sales totaled $2.4 billion in July–September. On the same day, NVIDIA's board created a standing committee to oversee its export control compliance. According to Senate Banking Committee Chairman Tim Scott, Huang and Commerce Secretary Lutnick have agreed to testify. No date has been set. On November 18, President Trump arrived in Shenzhen, China, where he was greeted by Xi Jinping. The outcome of the November 19 U.S.-China summit had not been released at the time of reporting. U.S. officials say they do not expect a formal agreement.
New products and research
On November 18, Google DeepMind released "Genie 4," a model that creates playable worlds from text prompts, to some subscribers of Google's top paid plan. Users can change the weather, add characters or alter the storyline mid-play. It is the first Genie to generate sound on the fly. Worlds can be published and modified by others. All output carries a watermark indicating it was generated by AI. Resolution and session length have not been disclosed. For now, it has no direct bearing on business work.
"Stitch Loop," released by Google Labs, is an AI that keeps fixing software after it ships. Its fixes are not merged until a human reviews and approves them. Details are in our report "AI keeps fixing software after release: Google announces tool that merges changes only after human approval." Anthropic published a detailed report on how its AI solved an open problem in black hole physics. Details are in our report "AI tightens upper bound on black hole mass and proves it to the limit in about six weeks." China's Tencent released, in a form anyone can use, a model it claims matches the satisfaction ratings of Claude's smaller version. Details are in our report "Only 12B active, claims Haiku-level performance: China's Tencent releases model." AI Mode in Google Search has added conversations that leave no history. They are disconnected from the user's account and are not used to learn preferences. This could be useful for people who do not want work research saved to their personal accounts.
Corrections to our previous report
There are three corrections to our previous report. The $85 million for trusted AI in the Indo-Pacific was not first announced at the November 17 summit in Manila. It first appeared in an October 31 joint statement by the foreign ministers of Japan, the United States, Australia and India (the Quad). Under the BIS rule, Singapore is not a country in talks with the United States but one of the five countries requiring licenses. NVIDIA's "$36 billion guarantee" could not be found in primary-source filings; the correct figure is a maximum guarantee exposure of $118 billion.
Here is what we could not confirm. Anthropic's IPO price had not yet been set at the time of reporting (2 p.m. ET on November 18). A post claiming it "priced at $85" is also unsubstantiated. We could not read the page announcing that OpenAI designed a chip using its own model, and the figures are unknown.
Where nothing moved
Anthropic has not responded to the "not safe" statement issued just before its listing. No one has been found, as far as we can see, defending it publicly by name. With no rebuttal to the statement at all, how markets receive the signatories' argument is effectively left to the stock's first-day trading.
GPT-6.2 Astra is the first model OpenAI has released externally to be rated "High" for the ability to improve itself. Even so, the main writers and evaluation organizations on the safety side (Zvi, Redwood Research, METR) have yet to publish a comprehensive analysis. The safety side's response has been limited to short posts questioning how it was tested.
NVIDIA's board has not yet formally responded to Treasury Secretary Bessent's call for "a CEO from outside the company." The deadline is around November 25. On the earnings call, the company repeated that it is "business as usual." Uncertainty over its leadership has been cited as one reason the shares fell.
Editorial cartoon
