Anthropic is worth $965 billion — and for eighteen days a US export directive switched off its most capable models worldwide

Every figure in this article is drawn from Anthropic's own announcements and funding disclosures, or from CNBC, Bloomberg, Fortune and TechCrunch as attributed. Anthropic is a private company and does not publish audited financial statements; revenue figures are the annualised run-rates the company itself discloses. Sections marked as analysis are identified as such.

At 5.21pm Eastern on 12 June, Anthropic received a directive from the United States government requiring it to suspend all access to its two most capable models — Claude Fable 5 and Claude Mythos 5 — by any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees.

Complying with that instruction, as written, meant switching the models off for everybody. Anthropic did so.

What triggered it

The order followed a report by researchers at Amazon — one of Anthropic's largest investors and its primary cloud and training partner — who had found a way to prompt Fable 5 past its safeguards so that it identified software vulnerabilities. The government treated the finding as evidence that a model with offensive cyber capability was in wide commercial release.

Anthropic disagreed, and said so publicly: a narrow jailbreak, it argued, should not be grounds for recalling a commercial model deployed to hundreds of millions of people, and a standard applied that way would halt new model deployments across the entire frontier industry. It is an argument with obvious self-interest attached and it is also, on the merits, difficult to dismiss.

The controls were lifted on 30 June. Fable 5 returned globally on 1 July; Mythos 5 had already been re-released four days earlier to a restricted set of companies and government agencies. The models were dark for roughly eighteen days.

The sequence before that is worth recording, because it explains why a jailbreak report landed with such force. Mythos first became publicly known in March through a data leak, at a point when Anthropic had told nobody it existed. The company unveiled it in April and deliberately restricted the rollout to a small group of firms under a cybersecurity programme called Project Glasswing, on the grounds that a model unusually good at finding software vulnerabilities is unusually dangerous in the wrong hands. Fable 5, released publicly on 9 June, was the same underlying system with safeguards applied in high-risk domains. Three days later the safeguards were judged insufficient.

The company that was switched off

The scale of what paused is the part that makes this more than a regulatory anecdote. On 28 May, Anthropic closed a $65 billion Series H at a $965 billion post-money valuation — the largest private funding round on record, co-led by Altimeter, Dragoneer, Greenoaks and Sequoia, with Capital Group, Coatue and D1 among the participants. The round placed Anthropic above OpenAI as the most valuable private company in artificial intelligence.

That valuation had nearly tripled in three months. In February the company raised $30 billion at $380 billion. The justification is a revenue line that is difficult to find a precedent for: run-rate revenue of roughly $9 billion at the end of 2025 reached $47 billion by May 2026. More than 1,000 business customers now spend over $1 million a year with the company, double the number in February.

For context on how fast that is: enterprise software businesses that reach $1 billion of annual revenue within five years are rare enough to be studied individually. Anthropic went from roughly $1 billion to a $47 billion run-rate in about seventeen months, and did it selling to large corporations with procurement departments, security reviews and legal teams — the slowest buyers in the economy.

What the money is for

Almost all of it is compute. In October 2025 Anthropic expanded its Google Cloud arrangement to access up to a million TPUs and more than a gigawatt of capacity — a deal CNBC reported as worth tens of billions of dollars. In November it committed $50 billion to American AI infrastructure. In April it deepened the arrangement again with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, deploying from 2027 and sited mostly in the United States, while Amazon Web Services remains its primary cloud provider and training partner. Forbes has reported a further roughly $30 billion of Azure services from Microsoft.

Krishna Rao, Anthropic's chief financial officer, described the Google and Broadcom expansion as "a continuation of our disciplined approach to scaling infrastructure" and "our most significant compute commitment to date." Both phrases are doing a great deal of work.

The strategic logic of spreading compute across Google TPUs, Amazon's silicon and Microsoft's Azure is straightforward: no frontier lab wants a single supplier setting the price of its only input. The financial logic is harder. These are multi-year, multi-gigawatt commitments made against revenue that did not exist two years ago, in a market where the cost per unit of useful computation has been falling fast enough to make any long contract a bet on the direction of that curve.

One structural detail deserves attention. Of the $65 billion Series H, about $15 billion was previously committed money from hyperscalers, including $5 billion from Amazon announced in April. The companies selling Anthropic its compute are also, in part, funding its ability to buy that compute — a circularity now common across the AI buildout, and one that makes both the revenue and the valuation harder to read from outside.

Analysis: the risk nobody priced

Investors in frontier AI have spent three years pricing two risks: that the models stop improving, and that the revenue never justifies the capital. June introduced a third. A company with a $965 billion valuation had its principal product removed from the market, worldwide, on a few hours' notice, by an instrument its investors had almost certainly not modelled.

What makes it genuinely novel is where the trigger came from. Anthropic has built its public identity on candour about danger — publishing capability evaluations, red-team findings and risk assessments that its competitors keep internal. In this instance the machinery it helped construct, applied to a finding by a partner and investor, produced an order it considered disproportionate and had to obey anyway. TechCrunch's framing was blunt: the safety warnings may have backfired.

That is not an argument against disclosure. It is an argument that the disclosure regime and the enforcement regime were built by different people with different objectives, and that a frontier lab now operates under a discretionary authority with no obvious appeal, no published threshold and no notice period. Every AI company's risk register acquired a new line item on 12 June, whether or not their lawyers have written it down yet.

There is a competitive dimension too, and it is uncomfortable. The report that triggered the directive came from Amazon, which is simultaneously Anthropic's primary training partner, one of its largest shareholders and the owner of a rival model family. Nothing in the public record suggests that was anything other than responsible disclosure. But the structure — where your investor, supplier and competitor are the same company, and their safety finding can take your product off the market — is one the industry built without ever examining it.

Analysis: what $965 billion assumes

Set the politics aside and the financial question is stark. At a $47 billion run-rate, the Series H valued Anthropic at roughly twenty times revenue — expensive but not absurd for a business growing at this rate, and cheaper on that multiple than several listed software companies. The difficulty is the denominator underneath: a company that must commit tens of billions of dollars to compute years in advance, against revenue that has existed for eighteen months.

It is also worth stating what is genuinely unusual here, because scepticism about AI valuations has become reflexive. Growing from $9 billion to $47 billion of run-rate revenue in under six months is not a story about hype cycles; it is enterprise software adoption at a speed the industry has not previously recorded, with a thousand corporate customers each paying seven figures. Whatever else is uncertain, the demand is not imaginary.

The counterweight is that none of these numbers are audited. Anthropic is private, publishes no financial statements, and reports revenue as an annualised run-rate — a measure that takes the best recent month and multiplies it, and which flatters any business growing steeply. An initial public offering, widely expected to follow the Series H, would replace all of it with numbers an auditor has signed. That will be the first genuine test of what this company earns.

The eighteen days in June are therefore best read as a preview rather than an aberration. Anthropic's competitors face the same authority, the same undefined threshold and the same partners-who-are-also-investors. The question the industry now has to answer is not whether frontier models can be built profitably. It is who, in the end, decides when they are allowed to be switched on.