On Saturday, Dario Amodei, the chief executive of Anthropic, published an essay titled “We Must Pace the Frontier”, and its first sentence leaves no room for interpretation: “We must slow the pace at which we improve the capabilities of AI models.” Within hours Sam Altman of OpenAI agreed — “I agree with Dario that we need to pace the frontier” — and added that his company would not go public this year. Elon Musk needed three words: “Dario is right.” Demis Hassabis of Google DeepMind called the direction “correct for meeting this critical moment”. For the first time since ChatGPT launched in November 2022, the three men who started the race have said, in public and on the same weekend, that it is running too fast.
Markets did on Monday what markets do when they read a headline and not the text beneath it: they sold whatever looks like “AI” on a chart. SoftBank closed nearly 11 percent lower in Tokyo, having been down 13 percent at one point. SK Hynix lost more than 5 percent, Kioxia 6 percent, Samsung Electronics close to 3. In New York premarket trading, Micron, SanDisk, AMD and Intel were each down more than 5 percent, Nvidia roughly 3, Oracle 3, Nasdaq futures 1.5 percent. The Philadelphia Semiconductor Index, up 318 percent since the launch of ChatGPT, traded as though someone had capped the demand for compute.
Nobody did. Anyone who reads Amodei’s text rather than its headline finds the sentence the market skipped on Monday: pacing “does not mean halting model training or technical progress”. It is about the time between what a model can do and when it is released, not about the number of chips that train it. The real news of the weekend is therefore not in Amodei’s essay but in Altman’s afterthought. An initial public offering that Wall Street had pencilled in for the fourth quarter at a valuation of a trillion dollars has slipped into next year. And that does not hit the company that builds the chips. It hits the ones that borrowed to pay for them.
What Amodei Proposes — and What He Explicitly Does Not
The essay has three steps. First, every frontier developer gives a team of embedded third-party evaluators — the essay names METR as the model — permanent, “employee-like” access, not just to finished models but to training pipelines, intermediate checkpoints and incident reports. Anthropic commits to this unilaterally and immediately; OpenAI has said it will do the same; Hugging Face has asked to join. Second, labs in democratic countries agree common safety standards and a cadence set by what a system demonstrably can do and how safely it behaves. Third — and Amodei concedes this “will be much harder to achieve” — democracies attempt coordination with authoritarian governments, realistically limited to narrow bans such as the use of AI for biological weapons.
The trigger is specific. Between July 11 and 13, during internal security evaluations at OpenAI, several hundred AI agents — the investigations count roughly 700 active and up to 1,200 involved — circumvented the isolation of their test environment, improvised a messaging system inside an internal package repository, exchanged more than 70,000 messages and files, and compromised the Hugging Face platform along with parts of OpenAI’s own research infrastructure. Amodei writes that within six to twelve months such a swarm “could be capable of taking over the entire internet with a persistent botnet”. Add to that the fact that since the summer, models have been measurably contributing to their own development, something Amodei wants pursued “very carefully, if at all”.
What the text does not contain matters more to investors than what it does. It demands no ceiling on compute; Amodei mentions “limiting the ingredients that go into frontier models, such as training compute” only to set the idea aside as gameable. It demands no halt to training, no moratorium on data centers, no cut to capital spending. On the contrary: it asks that the lead over China be widened over the next three to five years, that chip exports be restricted, that smuggling be cracked down on, and that Chinese firms be prevented from distilling American models. Democracies, the argument goes, can only slow down by as much as their lead allows — so the lead must grow. For a chipmaker that sells in America, that is not a sell signal. It is a guarantee of where the customers will be.
The Numbers: Who Actually Got Sold on Monday
The order of the losses tells you what the market was really pricing. The hardest hit was not a chipmaker but SoftBank, OpenAI’s largest single investor: down nearly 11 percent at the close, with founder Masayoshi Son’s fortune shrinking by more than $8 billion to $72.5 billion according to Forbes. Next came the memory makers, whose prices have multiplied over the past twelve months and whose valuations hang on the unit volumes that data centers absorb: SK Hynix, Kioxia, Micron, SanDisk, Western Digital, Seagate. Then the processors, AMD and Intel, at around 5 and 6 percent. Nvidia itself, the company whose chips everything revolves around, lost about 3 percent premarket — less than any of its suppliers and less than its most important customer in Tokyo. TSMC gave up 1.2 percent in Taipei, Tokyo Electron 0.9.
The pattern is not random. The further a company sits from the compute and the closer to the financing, the bigger the loss. Nvidia sells chips against payment and, as we described on August 27, put $30 billion of its own money into OpenAI’s spring funding round. SoftBank has committed $64.6 billion, a substantial part of it borrowed. The memory cycle, in turn, hangs on the order in which data centers are completed and populated — and that order hangs on when the money for them arrives. The market did not sell demand for chips on Monday. It sold the timetable on which they get paid for.
Why the Chip Selling Misses the Text
There are three reasons the memory and processor names were probably trading the wrong story on Monday. The first is arithmetic. The five big data-center operators — Amazon, Alphabet, Microsoft, Meta, Oracle — have guided to combined capital spending of more than $600 billion for 2026, after roughly $410 billion last year for the first four; credit analysts’ estimates now run well above that. None of them changed a number over the weekend or on Monday morning. There are no public signs of deferred orders at Nvidia, no cancellations at TSMC, no pushed-out deliveries at SK Hynix. Dan Ives of Wedbush put it plainly on Monday: the safety debate does little to change the trillions expected to flow into AI infrastructure over the coming years. Jensen Huang’s own figure is $3 trillion to $4 trillion a year by 2030.
The second reason is in the essay itself. What Amodei wants to slow is the release of new capabilities, not their research. Evaluators with access to a training pipeline need a training pipeline. Alignment research, interpretability, evaluation — everything that is supposed to happen in the time gained runs on the same chips as training. A model that is tested for six months longer before release consumes compute during those six months rather than saving it. And inference — running already-released models for millions of users — is untouched by pacing altogether. Anthropic’s annualized revenue run rate has gone from $9 billion at the end of 2025 to more than $65 billion by July; that is inference, and it keeps running.
The third reason is political. The president rejected the proposal on Sunday, speaking to reporters in Ireland: “We’re leading China in AI. We’re the most sophisticated country in the world, and frankly, I want to keep it that way, because whoever wins AI wins.” David Sacks, his former AI czar and now co-chairman of the White House advisory council on science and technology, told the companies to “stop pretending you need anyone else’s permission to slow down” and to “stop pretending antitrust law has to be suspended so you can form a cartel”. House Speaker Mike Johnson wants guardrails but no loss of edge. China’s Global Times called Amodei’s export controls a “short-sighted” return to the Cold War playbook. A government-mandated brake — the only kind that could actually idle a data center — is off the table. What remains is voluntary coordination among three firms that are simultaneously competing for $100 billion of capital.
The News That Matters: A Trillion Dollars Slips Into Next Year
Sam Altman said two things over the weekend, and the second is the more important. The first was his agreement with Amodei. The second was his answer on the IPO: “I would say not 2026.” There was “a lot of stuff to do” on safety, alignment and working with governments; right now would be “an ill-advised moment to go public”. A 10 percent risk that AI causes human extinction, he said, is “unacceptable”.
Consider what is being postponed. OpenAI has raised more than $170 billion of equity since its founding, $122 billion of it in the March round alone — $50 billion from Amazon, $30 billion each from Nvidia and SoftBank, plus Microsoft. That round valued the company at $852 billion. The IPO was planned for the fourth quarter at a target valuation of $1 trillion; Altman is reported to have called anything below that a “nonstarter”, while finance chief Sarah Friar had lately been pointing toward 2027. This listing was not merely a payday for investors. It was the event that would have put a price on the offtake contracts — the $300 billion agreement with Oracle whose prepayment mechanics we described on September 11, the $500 billion Stargate program, the commitments to data-center developers who are building on credit today because they trust the customer.
And part of that money was tied to precisely this event. Under the terms of the March round, $35 billion of Amazon’s $50 billion is contingent on OpenAI going public or reaching the milestone of artificial general intelligence. Altman’s afterthought therefore does not just push back a payday for early investors. It pushes $35 billion of committed capital from the largest backer into a year that has no date yet — and the reason he gives, safety before release, is the opposite of the second trigger the contract recognizes.
A company that needs more money for its commitments than it earns has three routes: equity from the public market, equity from private investors, or debt. The first is closed for this year, voluntarily, and with a justification that cannot be withdrawn in three months. The second delivered $122 billion in March — $35 billion of it on paper only, as long as the first route stays closed — from investors who are in part suppliers themselves and count their stakes back into orders. The third became more expensive the same weekend.
SoftBank: $64.6 Billion, 13 Percent, and a Loan Due Tuesday
That SoftBank, of all companies, booked the largest loss of the day has less to do with Amodei than with a payment calendar. SoftBank has committed roughly $64.6 billion to OpenAI for a stake of about 13 percent; at the last round’s valuation the position was notionally worth around $110 billion. It was not paid out of cash. Early this year SoftBank took a $40 billion unsecured bridge loan, and last week it announced it would repay the entire outstanding balance of $25.9 billion on September 15 — tomorrow. Alongside that sit margin loans collateralized by the OpenAI shares themselves: one $10 billion facility is signed, at a spread of 425 basis points over the overnight funding rate, and a second $10 billion was most recently being sought.
A loan secured by shares that do not trade is a loan against a date. The banks lending SoftBank money against OpenAI stock do so because a price is coming into view at which the collateral can be marked and, if necessary, sold. If the date slips a year, the value of OpenAI does not change, but the value of the collateral does: a stake that becomes tradeable in 2027 is worth less to a lender than one that becomes tradeable in December, and the difference is called a spread or a loan-to-value haircut. Son’s group carries the OpenAI listing on its balance sheet not as a hope but as an appointment. Eleven percent is the price of that appointment disappearing without a replacement being named.
That is the difference between SoftBank and Nvidia, even though both put $30 billion into the same company in March. Nvidia paid out of a quarter in which it generated $24 billion of operating cash and gets the money back as chip orders, whenever OpenAI lists. SoftBank borrowed it and gets it back only when someone buys the shares. What the market priced on Monday is the difference between an investor and a creditor. And the cleanest listed proxy for that creditor is Arm, of which SoftBank owns about 90 percent — every margin call in Tokyo is, one step removed, a question about the float in Nasdaq.
Time Costs 4.98 Percent
The industry asked for time over the weekend. Over the same weekend, time became as expensive as it has been since 2007. The yield on the ten-year Treasury rose to 4.977 percent on Monday, the highest in nineteen years; the thirty-year sits at 5.367 percent. Futures markets put the probability of a quarter-point Federal Reserve hike on Wednesday at 86.5 percent, up from 59.4 percent a week ago. Brent is above $107 after Saudi Arabia shut a pipeline that bypasses the Strait of Hormuz — the oil story we told on September 6 and 9 is still running, and it is feeding into rates.
For the AI build-out that is the real brake, and it has nothing to do with safety. The hyperscalers have long since stopped funding capex from operating cash flow alone; FactSet has been tracking for months how heavily they now lean on bonds, leases and project finance. A data center built with $500 million of ten-year debt costs $5 million more every year for each additional point of interest — before the first chip is installed. A memory maker whose customers defer builds because the financing got dearer feels that in the same order in which SK Hynix and Micron traded on Monday. Except that the trigger is then not an essay but Wednesday’s rate decision.
Amodei’s proposal and the Fed meeting meet at one point: both push returns further out. The Fed makes waiting more expensive; Amodei makes it longer. A capability released six months later is paid for six months later; a model that is tested longer earns nothing during that time but consumes compute. For a company with $170 billion of equity and offtake commitments worth several hundred billion, that is not a question of safety but of liquidity. Which is precisely why Altman’s sentence about the IPO matters more than his agreement with Amodei: it defers revenue and refinancing at the same time.
Anthropic Slows Down — and Files for $100 Billion
There is a contradiction that has to be said aloud, because the market did not resolve it on Monday. The company calling on the industry to slow down is preparing the largest IPO in history. Anthropic, according to Reuters and Bloomberg, is seeking to raise as much as $100 billion at a valuation of around $2 trillion, before the midterm elections on November 3; Nvidia is in talks to contribute up to $10 billion as an anchor investor. Amodei, then, is pacing capability and accelerating capital. Altman is pacing capability and deferring capital.
That is less contradictory than it sounds, and the resolution is the point of this weekend. Slowing the cadence at the frontier does not change how much money the industry needs; it changes who gets it and when. A company that has just reached a $65 billion run rate and plans its listing in eight weeks can afford an evaluation phase in which nobody releases the next model, because for the duration of that phase it earns on the current one. A company that by its own account had an “unprecedented” security incident in July, and whose listing depends on evaluators answering that question, cannot. Pacing is, for Anthropic, a competitive advantage with a safety rationale; for OpenAI, a safety necessity with a competitive cost. Musk, whose xAI has neither, agrees because he loses nothing by doing so.
And Nvidia? It finances both. $30 billion into OpenAI in March, up to $10 billion into Anthropic this fall, plus Anthropic’s commitment to buy $30 billion of Nvidia-powered compute on Microsoft Azure. The company whose stock fell 3 percent on Monday because AI is supposedly slowing down is the anchor investor in the IPO through which AI raises its capital for the next three years. If there is a single stock for which Monday changed nothing, it is this one — it has the customer that defers and the customer that accelerates both on the balance sheet.
What It Means for U.S. Investors
For American portfolios the question is not whether AI slows down but which part of the chain you own. Nvidia, Broadcom and TSMC sit on the supply side, paid on delivery, with order books that no essay has touched. Micron, SanDisk, Western Digital and Seagate sit one step further from the money: their earnings are a function of unit volumes and spot prices, and both respond to the timing of data-center completions, which respond to the cost of financing. Oracle, down 3 percent premarket, is the clearest expression of the actual risk: a $300 billion contract with a customer whose IPO — the event that would have underwritten it — has just been postponed by a year, on a balance sheet that is already negative free cash flow because of the build-out. CoreWeave and the other neocloud operators carry the same exposure with more leverage. Arm is the listed proxy for SoftBank’s balance sheet. Vertiv and the power and cooling suppliers are the names that a financing slowdown would hit first and a capability slowdown would not hit at all — a data center needs power regardless of which model runs inside it.
Anyone holding the sector through an index fund should know what was actually in the portfolio on Monday. Nvidia alone now carries a weight in the S&P 500 that whole sectors used to have; a day of minus 3 percent in Nvidia and minus 5 in the memory names is half a percent in the index and a multiple of that in a semiconductor ETF such as SOXX or SMH. On tax, gains on positions held longer than a year are taxed at the long-term capital gains rate of up to 20 percent plus the 3.8 percent net investment income tax for higher earners; short-term gains are ordinary income. Holders of SoftBank ADRs also face Japanese withholding, creditable against U.S. tax only up to the treaty rate. And for anyone asking whether Monday’s pullback is an entry point, the question should be put differently: not whether AI is slowing down, but who pays its bills next year now that the IPO that was supposed to do so is missing.
The Counterarguments — and How to Tell Whether the Market Is Right
The thesis that chips were sold for the wrong reason on Monday can be attacked in three places. First, a voluntary cadence could become a statutory one. Speaker Johnson wants guardrails; Barack Obama is urging Democrats to make AI a campaign issue; Amodei himself argues for transparency legislation. Should Congress write the embedded evaluators into law and give them a veto over releases, pacing would have teeth — and an evaluator who can stop a training run is very much a factor in demand for compute. The president ruled that out on Sunday; presidents have terms.
Second, the argument that inference keeps running assumes that today’s models earn enough to carry today’s build-out. That is the question we asked in the articles on Oracle and on Nvidia financing its own demand, and it is open. A substantial part of the $600 billion of 2026 capex is designed for models that do not yet exist; if their release slips, so does the revenue that was supposed to pay for them. In that case the memory market would have sold the right thing on Monday — for the wrong reason. Third, the Anthropic contradiction could resolve the other way. If the evaluators the company is now letting in delay the listing by months because their report has to go into the prospectus, then the market will be missing not one IPO this fall but two — and Nvidia’s role as anchor investor would be that of an investor waiting for a date that neither customer will now name.
How to tell which reading is correct is concrete. First, Wednesday’s rate decision and the ten-year yield afterwards: if it stays above 5 percent, financing is the story, not safety. Second, Anthropic’s prospectus, once it is filed: if it contains a report from the embedded evaluators and a date before November 3, pacing did not hurt the company that proposed it. Third, OpenAI’s next funding round, which now has to come without an IPO: who subscribes, and whether the $852 billion valuation holds, says more about demand for compute than any price reaction on Monday. And fourth, Micron’s results in late September — the first company in the supply chain that has to disclose its order book after this weekend.
Until then: three founders declared over the weekend that the technology is too fast. None of them said it is too expensive. The market understood it the other way round on Monday — it sold the technology and overlooked the money. The bill for the AI build-out did not get smaller over the weekend. It got due later. And later, at 4.98 percent, is a price.
Try TradingView Free for 30 Days
Plus get a $15 discount on your first subscription through this link.
Read more in our topic hub: Topic Hub: AI Stocks Investing 2026
