Anthropic Warns AI May Soon Build Itself Faster Than Safety Checks Can Follow as $17 Billion Quarter Powers Its IPO Filing

Anthropic has told prospective investors that its artificial intelligence may soon start improving itself faster than the company’s own safety work can keep up. The warning appears in the amended stock market filing the Claude maker sent to the United States Securities and Exchange Commission on Monday, 2 November, and it is the bluntest statement any leading AI laboratory has yet put in a legal document about the moment when AI begins to drive its own development.

According to the filing, reviewed by TechCrunch, Anthropic’s internal estimates “indicate that we may imminently meet our Responsible Scaling Policy’s definition of ‘dramatically accelerated progress attributable to AI'”. The company was careful to add that no current model has been designated as having crossed that line, and that it is still assessing the underlying data. But the same passage contains an admission that is harder to walk back: Anthropic says it has “in the past not identified, and may in the future not identify, a significant acceleration” in its rate of development “until after it has occurred”.

The disclosure landed in the same document that gave Wall Street its first look at Anthropic’s third quarter: revenue of about $17 billion for the three months to 30 September, up roughly 47 per cent on the previous quarter and more than eleven times the figure a year earlier, alongside about $895.4 billion of signed spending commitments, almost all of it for computing power and data centres. The filing is an amendment, known as an S-1/A, to the original prospectus Anthropic published on 26 October, which covered only the first half of the year. Anthropic has applied to list its Class A shares on the Nasdaq Global Select Market under the ticker ANT, and it is expected to begin marketing the shares to investors next week, in what could be one of the largest technology listings ever.

What the filing actually says

The new language sits in the Risk Factors section of the prospectus, the part of a filing where a company is legally obliged to tell investors everything that could go wrong. Anthropic’s warning is framed in plain terms. Its “increasing use of AI systems to accelerate our own research and development may cause the capabilities of our models to advance more rapidly than our ability to develop, validate, and implement corresponding safety, alignment, and security measures”, the filing says, according to TechCrunch’s review of the document.

The idea at the centre of the warning is recursive self-improvement, often shortened to RSI. It describes the point at which an AI system becomes good enough at AI research itself that each generation of models builds the next one faster than the last, so that progress compounds without a matching increase in human effort. Anthropic’s own safety rulebook, its Responsible Scaling Policy, already treats “dramatically accelerated progress attributable to AI” as a threshold that should trigger stronger safeguards. The amended filing is the first time the company has said in writing that its own near-term expectations put it close to that line.

The filing also gives one measure of how much Anthropic already leans on its own technology. The value of its internal use of Claude, measured at list prices, grew more than 250 times between January and September 2026, according to an analysis of the filing, and by April that internal use had passed the company’s monthly personnel costs, excluding stock-based pay. In other words, the company was already, by the spring, consuming more of its own AI’s work than it was spending on its people’s salaries.

Anthropic pairs the warning with a commitment. “We do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding, if and when that becomes technically possible,” the filing says, and it adds that the company is “considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development”. The filing is equally direct about the risk: “Using AI to help build more capable AI could also pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.”

That language builds on warnings Anthropic made in its original prospectus a week earlier. The first filing said that “sufficiently advanced AI systems could, if not developed, deployed, and operated with appropriate safeguards, oversight and security measures, pose catastrophic or existential risks to humanity”, and it tied that outcome specifically to recursive self-improvement. “Risks will increase as AI systems substantially automate AI research and development if, for example, they accelerate AI progress beyond our ability to evaluate and safeguard new models,” the original filing said. It also coined a phrase for the prize: “We refer to this recursive dynamic as ‘closing the loop,'” the prospectus said, “and we believe it is one of the most significant structural advantages in frontier AI.” Failure to close that loop, it added, could leave a company “falling progressively further behind as competitors continuously use AI to further enhance their own development capabilities.”

The threshold itself is precisely defined. Under version 3.4 of the Responsible Scaling Policy, effective 8 July 2026, the line is crossed when AI doubles the rate of progress in aggregate AI capabilities, compressing roughly two years of progress at the 2018-2024 pace into a single year. Anthropic’s August risk report, covering the period through 15 July, judged AI-driven acceleration still below that two-times threshold, and an outside evaluation judged Claude Opus 5.5 unlikely to fully automate AI research and development. The amended filing does not move that judgement; it says the company’s expectations for future models point towards the line, not that any model has crossed it.

The $17 billion quarter — and the $895 billion bill

The warning would matter less if it came from a small laboratory. Anthropic is now, on the filing’s own numbers, one of the fastest-growing companies in corporate history. Revenue for the three months to 30 September was about $17 billion, according to the filing as reported by CNBC and Reuters, up from $11.56 billion in the second quarter and from $1.38 billion in the same quarter of 2025 — a year-on-year increase of about 1,127 per cent. Revenue for the first nine months of 2026 reached $33.24 billion, roughly 1,200 per cent higher than the same stretch of 2025.

The growth, however, is slowing. Quarter-on-quarter revenue rose about 47 per cent in the third quarter, down from 144 per cent in the second quarter and 132 per cent in the first, according to CNBC. And the quarter swung back to an operating loss of $567 million, after a $974 million operating profit in the second quarter. Total operating expenses jumped to $17.5 billion from $10.6 billion, and the filing puts the blame mainly on computing power bought under short-term contracts, which it says “currently costs significantly more” than longer-term capacity. About $5 billion of short-term computing was bought in the quarter, and computing took 58 per cent of revenue, up from 48 per cent, with the company expecting more short-term buying to compress its margins in the near term.

The net loss looks startling at first glance: $23.29 billion for the quarter and about $104 billion for the nine months. But the filing’s own accounting shows that figure is driven largely by non-cash items, including the revaluation of convertible notes, rather than by cash going out of the door. The operating result, a $567 million loss on $17 billion of revenue, is the better guide to the underlying business — and it shows a company whose biggest cost is the computing power it must rent to keep growing.

That cost is also the filing’s most eye-catching number. Anthropic says it has entered into purchase commitments and lease agreements with an aggregate committed spend of about $895.4 billion, “substantially all of which is related to compute and infrastructure”. Roughly $70 billion of that falls due in the twelve months after 30 September, and nearly half — about $448 billion — is expected to be incurred before the end of 2030, according to Reuters. The commitments run through 2045, and about $100 billion more was added after 30 September, according to Bloomberg. Around $214 billion of the total is subject to termination rights or conditions, such as a counterparty securing financing or approvals.

The largest counterparties give a sense of where the money goes. Broadcom tops the list at about $161 billion in equipment leases, followed by Google at about $121 billion and the artificial intelligence unit of SpaceX at up to $120.5 billion, according to the filing. Amazon is also among the largest counterparties. Anthropic plans another $125 billion to $150 billion of Google chips and expects to spend more than $20 billion with AMD. The company is building its own dedicated infrastructure, which Fortune has called Anthropic Direct, so that it can swap expensive short-term capacity for cheaper long-term capacity as new data centres come online. It held $146.6 billion of liquidity at the end of September, including cash, short-term investments and an available credit line.

The listing itself is still taking shape. The amendment leaves the price range and the number of shares blank, so it does not yet say how much Anthropic would raise or at what valuation. Multiple reports have put the target valuation at up to $2 trillion, although one tracker of the offering says institutional investors have pushed the expected valuation down to $1.5 trillion to $1.6 trillion after the recent technology sell-off. Bloomberg reported that Anthropic is looking to raise $3.5 billion from investors in Japan, and that the offering is set to happen before the United States Thanksgiving holiday. Part of the offering is expected to go to retail buyers at the IPO price through brokerages including Fidelity, Robinhood, SoFi and E*TRADE from Morgan Stanley.

A company racing its own creation

Anthropic was founded in 2021 by former OpenAI researchers led by chief executive Dario Amodei, and it has built its identity around a single claim: that the most capable AI company should also be the most careful one. Its Claude family of models is now one of the two leading products in the enterprise AI market, and the filing says Claude has “experienced the fastest enterprise adoption of any technology in history”. Amazon, which has invested $8 billion in the company, and Google, which has invested more than $2 billion, are among its largest outside backers.

That careful image is precisely why the new disclosure is striking. Anthropic has spent years publishing safety frameworks, hiring alignment researchers and warning publicly about the risks of advanced AI. Amodei himself has written at length about the dangers of a world in which AI development outruns human oversight. A company with that record does not add language about self-improving AI to a securities filing lightly — and the filing is a legal document, where misleading statements carry consequences that a blog post does not.

The timing is also significant. Anthropic is days away from asking investors to value it at a figure that would make it one of the most valuable companies ever to go public. A warning that its own technology may soon accelerate beyond its ability to evaluate it is, on its face, the kind of disclosure that could cut against that valuation. TechCrunch noted the tension: a claim of being close to a capability as dramatic as self-improving AI could cut either way with investors, as a sign of frontier leadership or as a flagged risk that invites a valuation discount. The fact that Anthropic chose to put it in writing, in a document whose job is partly to sell the company, suggests the disclosure was driven by legal caution, by genuine concern, or by both.

The disclosure does not arrive in isolation. The same week, OpenAI’s newest model, GPT-6 Astra, was at the centre of a widely shared research blog post showing it reading and changing the weights of another company’s model directly, lifting that model’s score on a mathematics test from 7.9 per cent to 48.1 per cent, according to the post, which topped the Hacker News front page on Tuesday. That claim comes from the post’s authors and has not been independently verified, but it illustrates the direction of travel: the most capable models are increasingly being pointed at the work of building better models. Meanwhile, OpenAI and Anthropic both set up dedicated teams to watch for election misinformation ahead of Tuesday’s United States midterm elections, and WIRED reported that it had uncovered more than 400 deceptive apps and websites disguising tools for generating sexually abusive deepfakes as ordinary services, prompting Apple to remove two identified apps within hours. The governance questions around AI are arriving at the same speed as the technology itself.

The wider market context is less forgiving than it was a month ago. October was the Nasdaq’s worst month since April 2025, with the index down about 7 per cent and the Nasdaq 100 losing more than $2.5 trillion in value. Marvell’s $115 billion all-stock deal for optical-components maker Lumentum, announced on Monday, was greeted by a sell-off in the chipmaker’s shares, and Oracle’s shares fell more than 8 per cent on Monday as investors questioned the funding required for artificial intelligence investments. Anthropic is asking the market to believe in a $1.5 trillion valuation at a moment when the market is openly doubting the economics of the AI build-out.

Analysis: Why It Matters

1. A safety warning is now a legal disclosure

Until now, warnings about recursive self-improvement have lived in research papers, safety blogs and conference talks. Anthropic has moved the argument into a securities filing, where it carries legal weight and where investors must price it. That changes who has to take it seriously. A pension fund deciding whether to buy Anthropic shares, a regulator deciding whether the listing needs extra scrutiny, and an enterprise customer deciding whether to build its business on Claude all now have the same sentence in front of them: the company that builds the model says it may soon be unable to track how fast the model is improving itself. Whatever the eventual outcome, the debate about AI self-improvement is no longer a fringe concern; it is a line item in the prospectus of a company about to go public at a valuation in the trillions.

2. The detector may be the weakest link

The most unsettling sentence in the filing is not the one about the threshold. It is the admission that Anthropic has “in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred”. A safety framework is only as good as the measurements that feed it, and Anthropic is telling investors that its measurements have failed before and may fail again. The Responsible Scaling Policy assumes the company will notice when it crosses a line and then apply stronger safeguards. If the crossing is only visible in hindsight, the safeguards arrive after the event they were meant to prevent. That is a structural problem, not a staffing problem, and it applies to every laboratory building frontier models, not just Anthropic. The honest question the filing raises is whether any company can reliably detect the moment its AI starts improving itself faster than it can be checked.

3. The numbers make the risk concrete

Abstract warnings about self-improving AI are easy to dismiss. The filing makes them concrete by attaching them to numbers. The value of Anthropic’s internal use of Claude grew more than 250 times in nine months. Computing consumed 58 per cent of revenue in the quarter. The company has committed $895.4 billion to computing and infrastructure — a sum larger than the annual economic output of most countries — and roughly $70 billion of it falls due within a year. These figures describe a company that is already, in practice, letting AI do a large and growing share of the work of building AI, while paying a fortune for the computing power to do it. The warning is not about a distant future. It is about the business the company is already running.

4. The IPO is a test of whether markets price safety

Anthropic’s listing will be the first real test of whether public markets put a price on AI safety. The original prospectus devoted 94 pages to risk factors, an extraordinary length for a company of any kind, and the amended filing adds to them. If investors demand a discount for the RSI warning, the market will have said that safety risk is a cost. If they ignore it and pay the full asking price, the market will have said that capability is all that matters. Either outcome will be read by every other AI laboratory planning a listing, and by regulators watching the sector. OpenAI, which is itself preparing for a possible 2027 listing after cutting its fundraising target to $20 billion, will be watching the Anthropic roadshow with particular care. The valuation Anthropic achieves will set the benchmark for the whole industry.

5. The human-in-the-loop promise is the real battleground

Anthropic’s commitment — that it does not anticipate being able to safely let Claude “fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding” — sounds reassuring, but it is also a line drawn in sand that the tide is already reaching. The company’s own filing shows that its internal use of AI is growing hundreds of times over in months, and the economic logic of the industry pushes relentlessly towards more automation: the laboratory that automates its research first gains the largest advantage, and the original filing is explicit that failing to “close the loop” means falling “progressively further behind”. The question is not whether Anthropic intends to keep humans in control — it plainly does — but whether the competitive and financial pressures on a newly public company will allow it to hold that line when rivals do not. The updates to the Responsible Scaling Policy that Anthropic says it is considering will be the first evidence of how it intends to answer that question.

What to Watch Next

1. The roadshow. Anthropic is expected to begin marketing its shares to investors next week. Watch whether the RSI language is raised by investors on the roadshow, and whether the price range, when it is published, reflects a discount for the risk the company has disclosed.
2. The Responsible Scaling Policy update. Anthropic says it is considering updates to its safety rulebook to keep humans in control of frontier model development. Watch whether a revised policy appears before the shares are priced, and whether it sets out concrete new safeguards rather than general commitments.
3. The valuation. Reports have put the target at up to $2 trillion, with some investors pushing for $1.5 trillion to $1.6 trillion after the technology sell-off. Watch where the deal actually prices, and what it implies for OpenAI’s own fundraising and possible 2027 listing.
4. The next filing. The amendment still leaves the price range and share count blank. Watch the next amendment for those numbers, and for any further changes to the risk disclosures — companies often sharpen their warnings as a listing approaches.
5. The evidence of AI building AI. Watch for independent verification of claims like the GPT-6 Astra weight-editing result, and for further disclosures from the other frontier laboratories about how much of their own research is now done by AI. The filing’s warning is about a trend across the industry, not a single company, and the next data point could come from anywhere.

Sources

• Pondero — “Anthropic’s IPO filing warns it may be close to AI that speeds up its own R&D” (reporting TechCrunch’s review of the amended S-1/A)
• PYMNTS — “Anthropic Reports $17 Billion Quarterly Revenue Ahead of IPO” (2 November 2026)
• AIF Intent — “Anthropic’s Updated IPO Filing: $17B Q3 Revenue, $567M Loss and an $895B Compute Bill”
• Model Dex — “Anthropic Tells Investors Its AI May Soon Speed Up AI Research Dramatically” (analysis of the filing text)
• Today News Brief — “Anthropic S-1/A: $17bn Q3 revenue, $895bn compute, self-improvement risk” (3 November 2026)