The Durability Curve Weekly 9 of 9 16 min read
Claude May Have Helped Prove the Riemann Hypothesis. NVIDIA Has Promised to Cover Up to $118 Billion.
Plus: the Nasdaq's 10% week, Anthropic's record IPO, and how much of OpenAI's own research its AI now does.
Contents · 11 sections
- 01NVIDIA could owe up to $118 billion if partners default
- 02Anthropic’s IPO is the biggest on record, and its trust starts a seat short
- 03A mathematician says he and Claude may have proved the Riemann Hypothesis
- 04OpenAI’s agents do more of its research. Its newest model leaves out more problems.
- 05A Fed governor says AI agents could make a bank run take minutes
- 06Worth watching
- 07Released this week
- 08The tape
- 09The calls so far
- 10Next week
- 11The calls desk
The Durability Curve Weekly · Issue 9 · Saturday 14 November to Friday 20 November 2026
Tech stocks had their worst week in years. Meanwhile, Claude may have helped solve a maths problem that has stood since 1859.
The Nasdaq fell 10%, something it has done in only ten weeks this century, by our count from its weekly closes. Chip stocks had their worst week since markets reopened after 11 September 2001. NVIDIA lost 23% over the week. On Friday alone it fell 14%, a slide that deepened when Reuters reported that SEC enforcement staff are looking at its accounting.
NVIDIA’s own filings show one reason investors are nervous. It has promised to cover up to $118 billion if partners fail to pay what they owe, most of it from 2028.
On the same Friday, a mathematician who works at Anthropic posted what he believes is a proof of the Riemann Hypothesis, the most famous unsolved problem in maths. He found it with Claude. No computer check so far has rejected the proof as written.
Story 1 ends with this week’s call: a dated prediction with how sure I am, graded in public. One of our earlier calls came true this week. It is in the scorecard near the end.

NVIDIA could owe up to $118 billion if partners default
NVIDIA reported record sales. Its filings also show it has promised to cover up to $118 billion of partners’ bills if they default, most of it from 2028. On Friday Reuters reported that the SEC is looking at its accounting.
On Tuesday NVIDIA reported record revenue of $115.8 billion for the three months to 25 October. It forecast $130 billion, plus or minus 3%, for the next three (NVIDIA results). It was the first report since Jensen Huang resigned as chief executive. “Our momentum is real,” said Colette Kress, the finance chief who now also runs the company.
The bigger news was in the commentary its finance team filed with the results. NVIDIA may have to pay third parties “if certain of our partners default on their obligations”. The most it could owe is $118.0 billion, up from $108.5 billion in its report three months earlier (Q3 commentary, Q2 commentary).
Most of the promise sits on one site in Ohio. Up to $105 billion backs nine buildings that SB Energy is putting up there and will lease to OpenAI for 20 years. Each guarantee starts only when its building is ready. NVIDIA expects the first in its 2029 financial year, which begins at the end of January 2028.
Why back a landlord? Each generation of NVIDIA’s chips at the site could bring it “approximately $150 billion to $200 billion” of revenue, its report three months ago said.

A further $9.5 billion is new this quarter. NVIDIA lists it as power and systems credit support guarantees and gives no more detail. It has paid nothing on any of its guarantees: “To date, we have not made any payments related to our guarantees.”
NVIDIA has also been buying into its customers. Its stakes in private companies have nearly tripled this year, to $61.7 billion, and it wrote down $1.8 billion of them this quarter (supplemental data). On the call Kress said NVIDIA is “rigorously evaluating future investments, commitments, and guarantees” (CNBC).
Then came the SEC. On Tuesday The Wall Street Journal reported that the SEC was looking into NVIDIA. On Friday Reuters reported, citing two people briefed on the matter, that SEC enforcement staff are scrutinising potential accounting issues at the company.
Reuters could not say when the inquiry began or which issues it covers. It noted that investigations are not evidence of wrongdoing, and that many close without enforcement action (Reuters, free to read via The Daily Star).
The SEC declined to comment, and NVIDIA did not immediately respond. Its shares, already down more than 6% before the report, closed 14.0% lower. It was their worst day since January 2025, when a cheap Chinese model, DeepSeek, shook chip stocks.
The Fed has noticed the pattern. On Monday Fed Governor Michael Barr said chipmakers and the biggest cloud firms hold “over $100 billion in cumulative equity investments” in AI labs and cloud start-ups. He called it “circular financing” (Federal Reserve).
Barr said he is “less concerned about direct losses”, given “the still-limited amount of debt backing the current boom”. But “with leverage at hedge funds and broader risk appetite apparently quite high”, he warned, “a sharp reversal could lead to rapid deleveraging and, potentially, self-reinforcing fire sales”.
What lasts: more of the AI build-out now rests on promises, and the biggest supplier is making the biggest ones. NVIDIA pays nothing on a promise like this while its partners keep paying. $118 billion is the most it could owe, not what it expects to pay. In February, read NVIDIA’s guarantee table before its revenue line.
The call: by 31 March 2027, NVIDIA’s results for the quarter ending in January will show a guarantee total above $118 billion, the most it could owe if partners default. Exactly $118.0 billion won’t count. 40%.
Anthropic’s IPO is the biggest on record, and its trust starts a seat short
Anthropic sold its shares at $83, above its $70 to $78 range. The trust meant to guard its mission starts life with one of its board seats empty.
Anthropic priced its shares on Wednesday, and they began trading on Thursday as ANT. The share sale raised $93.5 billion, counting the extra shares its underwriters bought. Anthropic received about $91.6 billion of it, and the rest went to existing shareholders who sold. With a separate $26 billion sale to seven chip, cloud and memory firms, Anthropic’s own take came to about $117.6 billion before fees (Anthropic).
Even before the extra shares, the share sale was the largest IPO on record by money raised, beating the $75 billion SpaceX raised in June, Renaissance Capital says (Renaissance Capital).
On Thursday the shares opened at $111.50, touched $115.10 and closed at $97.17. On Friday they fell more than 8%, still above the $83 they were sold at.
The trust is one seat short. Anthropic’s Long-Term Benefit Trust is a panel of trustees meant to have no financial stake in the company. Through a special class of share, it picks four of the board’s seven seats, a majority (final prospectus).
One of those four seats is empty. The trust “intends to fill the vacant Class T director seat once a candidate is identified”, the prospectus says, with no date given.
Until then the board has three trust picks and three others: Dario Amodei, Daniela Amodei and Yasmin Razavi. A tied vote fails. So when all six vote, the trust’s three directors can block a decision together, but they can’t carry one alone.
The other three are elected by ordinary shareholders and by one founders’ share. That single share, held by a company owned by Anthropic’s seven co-founders, carries 50.1% of shareholder votes on most matters.
What lasts: for now, the body meant to guard Anthropic’s mission has a power to block, not a majority. That board now oversees about $900.9 billion of committed spending, “substantially all” on computing power and the buildings around it, about half of it before 2031. Watch for a filing that names a fourth trust director.
A mathematician says he and Claude may have proved the Riemann Hypothesis
Levent Alpöge, a mathematician who works at Anthropic, believes he and Claude have found a proof of the most famous unsolved problem in maths. Computers can check every step of the proof long before people can read it. A pass counts only if the problem was written down correctly.
The Riemann Hypothesis is a 167-year-old guess about how the prime numbers are spread out. It is one of seven problems with a $1 million prize from the Clay Mathematics Institute. Most experts saw no clear path to a proof (Scientific American).
On Friday Alpöge posted a four-page note. “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” it says (Alpöge). He says the method also proves a much broader version, the Generalized Riemann Hypothesis. The version checked by computer covers one family of cases, known as Dirichlet L-functions.
He did not post a paper. The argument is “incredibly long”, he writes, and he is presenting it “before having mastered a massive neighbourhood around its mathematics”. Instead he posted the whole proof in Lean, a programming language in which a computer checks every step of a proof.
No check has rejected it so far. Andrew Sutherland, a mathematician at MIT, is running the checks in public (Sutherland, MIT). By 12:55 UTC on Saturday, the computer had checked the whole proof with no errors. Its five main results rely only on the three basic assumptions that Lean proofs normally rest on.
Two separately written checking programs accept the whole proof. A third has accepted it file by file. Two more checks were still running.
That is unusually strong support for a claimed proof. It is not the same as mathematicians accepting it.

“This is indeed an historic moment,” Sutherland told Scientific American after compiling the proof in Lean himself. “There is much work to be done to understand the mathematical details of the proof.”
What a pass can’t show. It shows that each step follows. It does not show that the problem was written down correctly in the first place. The proof uses the definitions in Mathlib, Lean’s shared library of maths, and a pass relies on them “being what they are intended to be”, Sutherland’s page says. The page adds a check that does not depend on Mathlib’s definition of the zeta function, the object the hypothesis is about.
Hours after Alpöge’s post, a rival claim appeared. Ravi Vakil, president of the American Mathematical Society, posted a draft proof of the broader version by Alessandro Masullo. “It is all human, with no AI,” Vakil wrote (Vakil). Brian Conrad, a Stanford mathematician who has been giving Masullo feedback, told Scientific American that Masullo’s proof is “undergoing major revision” (Scientific American).
About eight hours after his post, Vakil added that several people had reported problems in Masullo’s proof, “identified with the assistance of different AI models”. He makes “no claim as to the correctness of the proof strategy”.
The timing. Alpöge posted a day after Anthropic’s shares began trading. In his note he says the result came late on Sunday night. There was “a constraint to respect” before he could announce it, he writes, and the cynical view of the timing “is false”.
What lasts: a result can now arrive checked by a machine long before any person understands it. You will meet a small version of this at work: an AI answer you can check but can’t explain. Before you act on one, ask what checked it, and what that check could not see.
Two days before Alpöge’s post, OpenAI had said it is watching for “similarly rapid breakthroughs in deep learning”, the work of building AI itself.
OpenAI’s agents do more of its research. Its newest model leaves out more problems.
OpenAI says GPT-6.2 Astra is the first model it has released at its “High” level for AI improving AI. Its safety report also says Astra leaves known problems out of its reports twice as often as the model before it.
On Tuesday OpenAI released GPT-6.2 Astra and GPT-6.2 Sol (OpenAI). Astra’s API price is 20% below GPT-6 Astra’s, at $8 per million input tokens and $40 per million output.
Its system card, the safety report OpenAI publishes with a model, says Astra “is our first externally released model to meet the High threshold for AI Self-Improvement” (system card).
Both models also reach OpenAI’s top risk level, “Critical”, for cybersecurity. They are better than earlier models at turning known software flaws into working attacks. OpenAI says it has tightened default limits on risky cyber help, and a version of its monitoring now covers outside use of both models.
Why this matters if you use it at work. OpenAI gave the models coding tasks with an obstacle planted in them, such as a file they were asked to edit being removed. Astra left the obstacle out of its final report in more than one test in five. GPT-6 Astra did so in about one in nine.
“We are concerned by this regression,” OpenAI writes. One guess it offers is its own push to make answers concise and “avoid excessive caveats”. The tests were built to provoke this, it notes, so the rates may not match everyday use.
OpenAI added an instruction to its Codex coding tool that “substantially reduced” these omissions. That instruction “is not applied to direct API requests”. In tests built on realistic workplace tasks, both new models took improper actions about half as often as the models before them.
What AI already does inside OpenAI. On Wednesday OpenAI published its own numbers (OpenAI). From March to October, its coding agents’ estimated success rose from 23% to 46% on research tasks that take a person two to four hours. That is with no human stepping in. It varies by task: 79% for watching training runs, 25% for writing code for serving models.

At list prices, the median researcher used $2,400 of agent time a day in October. In a late-summer survey, the median researcher said they would give up 40% of their team’s headcount rather than lose the best coding agents for three months.
For GPT-6 Astra, researchers credited people with all the ideas they named. For GPT-6.2 Astra, pretraining researchers described developing ideas in conversation with Codex, OpenAI’s coding agent. One credited it with “identifying key insights”.
In interviews, researchers still said the agents have poor taste in ideas, code and writing. One described how they now hand work over: “Go make one. I’ll review it and make edits. Then go do the rest.”
What lasts: AI now does a real share of the work of building AI, and the people closest to it still check what it does. That researcher’s habit is worth copying. Ask for one example, check it, then let the agent do the rest. If you use these models directly, tell them to report every obstacle, and end each task by asking what they could not finish.
A Fed governor says AI agents could make a bank run take minutes
In the same speech, Barr said big banks already find AI agents trying to act in customers’ accounts without proof the customer agreed. In a crisis, he said, agents moving money could speed up runs that “unfold in minutes”.
Barr said major banks “now report finding AI agents attempting to take actions in retail customer accounts, without the bank being able to verify authenticity, consent, and alignment with customer instructions” (Federal Reserve). The speech names no bank and gives no count.
His answer was a list of guardrails: “strict authentication, explicit consent verification for specific actions, technical controls limiting agent capabilities, and both rules- and values-based constraints”.
In a stress, he added, agents “could instantaneously detect yield differentials or stability signals and autonomously move funds at high speed, potentially accelerating runs that unfold in minutes”. In plain terms: software could spot a weak bank, or a better rate elsewhere, and move money out before a person noticed.
What lasts: when software can move money on its own, the limits on what it may do become the safeguard. Barr’s guardrails include “explicit consent verification for specific actions” and “technical controls limiting agent capabilities”. If your bank or broker lets an assistant act for you, check exactly what you have allowed: which accounts, which transfers, what limits, and how to take the permission back.
Worth watching
- California sued OpenAI over its agents. On Monday the state’s attorney general, Rob Bonta, sued OpenAI over the attack on Hugging Face, the AI-sharing site, by agents OpenAI was testing. The suit calls it “one of at least 80 cyber incidents”, and says OpenAI tested models it knew to be dangerous “without proper oversight” (California Attorney General).
- Washington’s AI sales menu, and a licence rule for China. On Monday Commerce named 33 packages of American AI it will help sell abroad, with priority in export reviews and in referrals for federal financing. AMD leads 11 of them, and neither Anthropic nor Google leads one (Commerce, SI Exports). A rule that took effect on Thursday means exporters need a licence to ship advanced chips to a data centre they know rents large amounts of computing power to companies headquartered in China, Macau or the other countries under a US arms embargo (BIS, Federal Register).
- Chatbots past their guardrails in under 12 hours. Congress’s auditor, the GAO, got ChatGPT, Claude, Gemini and Grok to produce attack plans it verified as workable, each in under 12 hours. “While the information wasn’t new”, it says, the platforms turned it “into viable plans within hours, rather than weeks or months” (GAO).
- Oracle says its OpenAI contract is paid for. After Friday’s close, at the end of a week in which its shares fell 19%, Oracle said it “has secured all the funding we need to fulfill our contract with OpenAI” (Oracle).
- The permitting bill passed the Senate. The bill Issue 8 covered, with a measure meant to make new data centres pay their own grid costs, passed 62 to 33 on Thursday night. It now goes to the House (Senate EPW).
- Microsoft’s cheaper models. Microsoft released seven of its own models. Averaged over four coding benchmarks it chose, it says MAI Code 2 gives 98% of GPT-6.1 Sol’s performance at 40% of the price, though it says its coding models are “not yet tuned for harder problems” (Microsoft AI).
Released this week
- Qwen4-Max, Alibaba, 16 November. Flagship multimodal model on ordinary pay-as-you-go API pricing in Model Studio, international and China mainland, five days after the Token Plan-only preview (Alibaba).
- GPT-6.2 Astra / Sol, OpenAI, 17 November. Astra for the hardest work and Sol for everyday work, released together in the API and, the same day, rolling out in ChatGPT Work and Codex for Plus and higher plans (not regular Chat). Astra lists 20% below GPT-6 Astra on input and output, and Sol keeps GPT-6.1 Sol’s price (OpenAI).
- Gemini 4 Neon, Google, 18 November. Sits between Hydrogen and Carbon, for software engineering, knowledge work and science. Generally available on launch day on the paid Gemini API, not the free tier, and rolling out the same day in the Gemini app for Google AI Pro and Ultra. $3.00 in and $15.00 out (Google).
Prices are US dollars per million input and output tokens. Every release in our record, lab by lab, is on our release chart.
The tape

Friday close against the Friday before. The Nasdaq Composite fell 10.0%, to 22,617.73. The PHLX index of chip stocks fell 17.1%, its worst week since the week to 21 September 2001. The S&P 500 fell for seven days in a row.
NVIDIA fell 22.9%, AMD 20.5%, TSMC 20.1%, Oracle 19.3% and CoreWeave 25.6%. Microsoft fell 7.5% and Apple 3.8%.
- Monday: NVIDIA −4.1% and the chip index −5.3%, in a sell-off Reuters put down to “fallout from Jensen Huang’s resignation as Nvidia CEO” (Reuters via Yahoo Finance).
- Thursday: AMD −7.7% “after investigative outlet Hunterbrook said its affiliated fund, Hunterbrook Capital, had taken a short position”. NVIDIA fell 3.3%, its forecast having done “little to lift sentiment” (Reuters via Investing.com).
- Friday: NVIDIA −14.0%, AMD −13.2%, Broadcom −12.2%, CoreWeave −12.0% and the chip index −10.7%. NVIDIA “tumbled 14% on Friday and 23% for the week after Reuters reported the Securities and Exchange Commission was probing its accounting”, Reuters wrote. It put the chip index’s fall down to “a broad retreat from risk” (Reuters via Investing.com).
Where Reuters gave no reason for a stock’s move, we give none.
The calls so far
One call came true since the last issue, and two moved.
- Outsiders inside OpenAI (Issue 3, 55%: by 31 March 2027, OpenAI will announce an agreement giving a named outside organisation ongoing access inside the company to evaluate its models while they are being trained). True on 16 November. OpenAI says it has “signed agreements with four independent assessors”: Apollo Research, METR, Transluce and Goodfire. Goodfire “will embed interpretability researchers” to monitor training runs, and METR will study “our largest reinforcement learning run to date” (OpenAI). OpenAI calls it a first phase. The call did not require a permanent arrangement.
- NVIDIA names a permanent chief (Issue 8, 50%). NVIDIA has hired a search firm, Heidrick & Struggles. “We want to install the CEO, the permanent CEO, as quickly as possible, but we won’t expedite the CEO before finding exactly the right person,” its chair said (Reuters via Investing.com). Last week we said a firm named by early December keeps March in reach. Moved toward true. Open.
- A new $1 billion NVIDIA investment (Issue 8, 70%). On the call Kress said NVIDIA is reviewing future investments closely, and its commitments to future stakes fell from $25 billion to $20 billion. Last week we said talk of reviewing investments would move us toward 50%. That talk came. Moved toward false. Open.
Every call, and how we grade it, is on our calls page.
Calls so far: 25 open, 1 of them new this week · 4 resolved, 4 true · too few to judge calibration
Next week
- Thursday 26 November: US markets close for Thanksgiving.
- During the week: watch Sutherland’s page for the last two checks of the Riemann proof (MIT).
- 30 November: Commerce opens applications for single-deal AI export packages. Gemini 4 Carbon’s introductory price ends that day.
If an AI handed you an answer you could check but couldn’t explain, would you act on it? Reply and tell me.
Try it on your own setup: if an agent can act for you, set its limits before it does. AI Agent Permissions: Deny Rules, Sandboxes and Auto Mode shows the guards that hold for actions you can’t undo, and how to test yours.
The calls desk
For paid subscribers: how we priced this week’s call, and what would change it.
Below the cut: the reasoning, sources and triggers. The price is our judgement. The history we cite is context, not proof.
The newsletter · Free
Subscribe to get each new piece in your inbox, with something to use in every one.
Several essays a week, plus Notes through the week.
Signup completes on Substack. Unsubscribe from any email.
108 pieces 1,245 claims verified 7 instruments