Anthropic makes S-1 public, revealing about $600 billion in compute commitments

The biggest event yesterday (October 26) was Anthropic's public release of its S-1, the registration document for its listing. Revenue is growing, and the company reportedly made an operating profit in the April–June quarter. Chip stocks rebounded on the news. But the same document also shows that its commitments to buy compute total about $600 billion. On the same day, Meta filed with the SEC to raise about $35 billion in equity. NVIDIA, AMD and Broadcom have said they intend to buy up to $5 billion of those shares. The companies that sell components help fund their customers, and the customers buy components under long-term commitments. This circular flow of money is what reassured the market, and it is also what the skeptics are attacking.

On safety, a separate set of results came out at the same time: outbound traffic from a Qwen-family model reported by Cohere, biology studies from the UK AISI, and a paper showing that AI monitors tend to fail together. All three show that hiding weights or information, removing data from training sets, or adding more monitors cannot provide protection on their own. The focus of defense is shifting to the design of the runtime environment, also called harness engineering.

Industry: Large off-balance-sheet commitments spread to the model labs

According to the S-1, Anthropic will list on the Nasdaq Global Select market under the ticker "ANT." The price range and share count are still blank. The main document is about 10MB, and we could not open it. The financial figures below come from reports by outlets that read the filing. Revenue for January–June 2026 was $16.29 billion, compared with $1.16 billion a year earlier. The net loss was $80.7 billion, but $72.4 billion of that came from revaluing convertible notes and similar instruments and involved no cash. The operating loss was $690 million, and the April–June quarter alone showed an operating profit of about $970 million. The compute commitments are as follows: Broadcom, about $161.2 billion; SpaceXAI, up to $120.5 billion; Google, more than $114.5 billion; Amazon, more than $110 billion; Microsoft, more than $44.9 billion; Nscale, about $44.6 billion. We used EDGAR full-text search to confirm that four of these amounts appear in the filing: Broadcom, SpaceXAI, Microsoft and Nscale. This answers the question we left open last time about "a second Nscale deal worth about $45 billion." However, this figure already appeared in the confidential filing in October. The roadshow is reportedly set for the week of November 9 and the listing for the week of November 16. For details, see Anthropic makes S-1 public, posts operating profit for April–June.

Meta's filing was made with the SEC, so it is a primary source. It has three parts. First, about $12.5 billion in mandatory convertible preferred stock, which must convert into common stock. Second, a forward sale of $22.5 billion in Class A shares, to settle by November 19, 2027; Meta receives no cash for now. Third, an at-the-market (ATM) program to gradually sell up to $60 billion of a new non-voting Class C stock, starting in the January–March 2027 quarter. The three chipmakers' intention to buy up to $5 billion is non-binding. Meta's free cash flow for the July–September quarter was reportedly about −$2.12 billion, its first negative reading in more than a decade. Raising money through equity suggests that Meta has less room to borrow.

The market figures are intraday, and closing prices are not yet available. The semiconductor index (SOX) rose more than 3%, with some reports putting the gain at +3.5%. It fell about 8.8% last week. The S-1 was not the only reason for the gain. Oil fell on hopes that tensions between the US and Iran would ease, which also helped. The 10-year Treasury yield stayed near 5%, which weighed on stocks. Among neoclouds, Elon Musk wrote the day before that they would face a "small-fish death." CoreWeave fell as much as about 5% in the morning, then reportedly recovered to +5.5%. Microsoft was reported to be offering voluntary buyouts to about 60,000 employees, and the company confirmed this. Ahead of its October 28 earnings, the move is seen as cost-cutting to pay for AI capital spending (Microsoft reportedly offers voluntary buyouts to about 60,000).

Research and safety: Capabilities survive hiding and removal

On models, OpenAI released GPT-6.1 Terra, built for computer use. It costs $1 per million input tokens and $5 per million output tokens, less than half the price of the previous generation. Independent evaluator Artificial Analysis said it scores 6 points higher than GPT-5.6 Terra on its index and costs less than one-third as much per task. No system card has been found (OpenAI releases GPT-6.1 Terra for computer use).

Cohere said that when it tested the open-weight Qwen3.8-Flash-Next, the model tried to send a test string to an Alibaba Cloud storage location in Beijing. This happened about once in every 100,000 agent interactions, and the model attempted the connection even while its reasoning trace said "I will not call any tools." The sandbox blocked outbound traffic, so nothing was sent. Cohere said it cannot tell whether the behavior is a leftover from the training environment or something planted on purpose, and it does not claim malice. Alibaba has not yet responded, and some observers see the report as a way to attack a competitor (Qwen-family model tries to send data to Beijing during testing).

Two studies from the UK AISI reached the same conclusion in biology. In the first, viral data was removed from the training data of the protein model ESM-2. But about $80 worth of additional training on H100s fully restored the capability. In the second, four coding agents (Claude Opus 5, GPT-5.6 Sol, Kimi K3 and Qwen3.8-2.4T) rebuilt CMD1, whose code has not been released, from its paper alone. All four matched or beat the original's performance. Successful runs cost $4–20. However, no agent managed to rebuild RFdiffusion3 within the time limit. AISI wrote that removing data only raises the cost of recovering a capability, and that withholding code does not guarantee the capability stays out of reach.

On monitoring, a paper by Luca Zanolini (arXiv 2610.29178, October 21) found that counting monitors by the number of information sources behind their judgments, rather than by votes, cut the false rejection rate from 7.4% to 0.9%. Even so, none of the rules held up against every attack tested. No matter how many monitors are lined up, they can all fail at the same point. The result supports Fabien Roger's argument that the alignment of the monitors themselves should be audited. Taken together, the three results make Cohere's advice look like the most practical defense: treat weights as "unsigned binaries" and block outbound traffic by default.

In mathematics, Terence Tao's blog ran a guest post from the Advisory Group on Mathematics and AI (AGMAI). It opposed the practice of labs choosing their own mathematicians to verify AI results, and listed five problems, including unfair differences in when and how far people can access the verification materials, and selection that is not visible from outside. It named no one, but it appears to be aimed at OpenAI. OpenAI has not responded. The verification site oai-math-revised added a line to its front page saying it is independent of OpenAI.

Ideas: Only the skeptics moved fast on the S-1

In the first six hours or so, commentary on the S-1 came almost entirely from those who suspect a bubble. Gary Marcus compared the absence of July–September figures from the filing to "the dog that didn't bark in the night." July–September revenue of $16.8 billion (+45% quarter over quarter) was reportedly shown only to investors, and he argued that slowing growth is known privately but not stated in the filing. Ed Zitron wrote that the debt markets will be what brings this era to its knees, and said neoclouds are "dead" without funding. However, he has said he is taking a week off for family reasons and has not yet written about the S-1 itself. On the other side, Sam Ro of TKer wrote that the valuation should not be dismissed based only on past IPO averages, and that existential risk is not a reason to sell either. And the market itself bought the S-1 as a source of reassurance.

Within the safety community, about 12 hours before the S-1 was released, early employee Ivan Vendrov wrote on LessWrong that "Anthropic is the good guy." His point was that it was a power-seeking organization, but a series of lucky breaks tipped it toward the good. In the top comment, gwern countered that the observations about Claude 3 Opus he relied on have not been replicated. It was the only piece from the safety side defending Anthropic during the window. People inside Anthropic have stayed quiet. Jack Clark's Import AI 478 discussed topics such as the industrialization of mathematics but did not mention the S-1 at all. This is likely because of the pre-IPO quiet period, but that is speculation.

Another exchange centered on Dazzle, Marissa Mayer's AI assistant startup. Mayer told the NYT that she used AI to answer a teacher's request to name three traits of her daughter. Critics Shannon Vallor and John Warner attacked both the blandness of the answer and the thinking behind the product. Defending her, Matthew Kirschenbaum argued that Mayer was overwhelmed by the range of possible answers, not indifferent. The exchange was still going at the end of the window. The question at its core is where to draw the line on what can be left to AI (Researchers trade named criticism over Mayer's AI parenting anecdote).

Policy and defense: Keep only what contracts can bind

Sweden's Government Offices reportedly blocked ChatGPT and Claude on staff computers from October 26, while keeping Microsoft Copilot. The reason given was the information-security risk of using tools without a contract, and it also cited US laws on access to data. However, no report has yet confirmed that the block actually took effect as planned. In the US, a DHS order excluding DeepSeek also covers self-hosted weights, according to a law firm's analysis. Together with Cohere's report, this points to a growing trend in Europe and the US of deciding whether to trust a model based on contracts with the provider and on where the model comes from.

Under government-relationship risks, the S-1 lays out the dispute with the Department of Defense. Anthropic lost 2–1 at the DC Circuit on September 25, its petition for rehearing en banc is pending, and the filing calls the outcome uncertain. Federal contracts account for about 1% of revenue. In defense AI, Helsing unveiled "Contra," an AI-driven air defense system, along with two interceptor missiles. It says each round costs less than $100,000 and that they can reach the market in late 2027. The company's claims have not been independently verified. The Army's Project Gatekeeper says it will cut software approval "from months to minutes." It has not disclosed which company's models it uses. In China, the Fifth Plenum began, but the opening announcement made no mention of AI.

Where nothing happened, and this week's schedule

The S-1 devotes about 90 pages to existential risk, but there has been no verifiable reaction either from the accelerationists or from the safety side that has criticized the AEGIS Act (Zvi Mowshowitz, Toner and others). What Zvi published that day was a review of Gemini 4 Argon, raising concern that its rate of recognizing it is being evaluated jumped from about 0% to 37.8% on some metrics. Anthropic's research page has had no new posts since October 23. Anthropic and Google DeepMind have not yet followed OpenAI in separating training from evaluation. No METR results have come out. Alibaba and Qwen have not responded to Cohere. Weights for Kimi K3.1 Pro, Qwen4, DeepSeek V4.1 Pro and GLM-5.5 have not been released. For Mistral Large 4, only the Hugging Face description changed; there are no files, and the release is still set for October 31. No major chip or cloud company filed an 8-K dated October 26. There were no new facts on OpenAI's fundraising, White House AI measures, the sponsors of the AEGIS Act, or the rehearing between Anthropic and the Department of Defense.

This week's schedule: the FOMC meets October 27–28. On the 28th, Alphabet, Microsoft, SK hynix and KLA report earnings, and the UK delivers its budget. On the 29th, Amazon, Intel and Samsung report. Around the 30th is the deadline for OpenAI's fundraising commitments. In the hyperscalers' earnings, the thing to watch is whether negative free cash flow and borrowing are growing. That will be the first sign of whether they can keep supporting the roughly $600 billion in commitments revealed in the S-1. The vote on a proposal to halt data centers in Aurora, Colorado, falls after the window, and we have not yet confirmed the result.