The draft that Reuters reported on September 28 has not been officially published by the company. On one side, growth the industry had never seen. On the other, commitments of several hundred billion dollars and a highly concentrated customer base. The title of this article raises the idea of an impossible debt to repay. I take the question seriously, but I qualify it: what the document describes is mainly commitments to purchase computing power, not conventional debt.
Key takeaways
- Anthropic’s prospectus has leaked. It remains a non-public draft.
- In 2025, the company lost $42 billion on $4.6 billion in revenue.
- It plans $518 billion in spending on computing power.
- Nearly a quarter of its 2025 revenue came from two customers.
- The IPO is expected in November.
The Anthropic prospectus, an X-ray of an AI giant
According to figures reported by the financial press, here are the essentials:
- 2025 revenue of $4.6 billion, against operating expenses of $13 billion.
- Operating loss of more than $8 billion and net loss of $42 billion.
- Cash of $20.28 billion.
- Revenue of $11.5 billion in the second quarter of 2026, versus $4.73 billion in the first.
- Planned spending of $518 billion on cloud services and data centers.
Another telling detail: the risk factors take up 80 pages, compared with 48 for the description of the business. Rarely has a company insisted so much on what could go wrong.
Growth unprecedented in economic history
2025 revenue is said to be twelve times higher than in 2024. On the projections side, the annualized pace reportedly rose from about $9 billion at the end of 2025 to about $65 billion at the end of July 2026. I have not made an exhaustive comparison with other companies, but I know of no equivalent in software. What remains to be seen is whether this trajectory can continue beyond the phase of mass adoption.
Anthropic’s losses, an accounting illusion that tells a story
The gap between an operating loss of about $8 billion and a net loss of $42 billion is considerable. The reports I consulted do not detail its composition, and I do not want to speculate. What is reported, however, is that the company would be operationally profitable for a second consecutive quarter. The snapshot of 2025 therefore does not tell the whole story of the current situation. The public version of the prospectus will need to clarify this point.
Claude Code, the engine of the enterprise conquest
Launched in May 2025, the agentic coding tool has become a pillar. Its annualized pace exceeded $2.5 billion in early 2026, and businesses accounted for more than half of it. At that date, it represented only about 18% of the annualized total. The strategic stake lies elsewhere: by selling a finished product rather than simple access to a model, Anthropic seeks to capture more value from its customers.
Run rate vs actual revenue, beware the mirage of projections
Recognized revenue must be distinguished from “run rate.” According to a person close to the matter cited by Reuters Breakingviews, Anthropic multiplies usage-based sales from the last 28 days by 13 and monthly subscriptions by 12. In other words, it is an extrapolation, not a result. OpenAI has criticized this method, even though US accounting rules allow both approaches. My own simple, approximate calculation: the $11.5 billion of the second quarter, annualized, gives about $46 billion, noticeably less than the $65 billion run rate at the end of July.
Open source and new rivals threaten margins
The pressure does not come only from head-on competition. It also comes from prices. Meta, long the standard-bearer of open models, has changed course: its Muse Spark model is proprietary, contrary to its past commitments. In this landscape, the value of a model quickly becomes commoditized, and margin becomes the real battleground.
Anthropic wants to replace all human intellectual work
Reuters describes a very broad vision of AI in the prospectus, and the company’s latest models are presented as designed for coding and knowledge work. I read an economic logic in this ambition: to justify a valuation of several hundred billion dollars, you have to target a market as vast as intellectual work itself. The same document nevertheless admits that its models could pose a catastrophic, even existential, risk to humanity.
Meta Muse and the SaaS apocalypse, who disappears first?
The “SaaS apocalypse” refers to the fear that AI agents will hollow out software billed per user. I have not verified the current state of the markets on this point. The Meta signal, however, is documented: the personal agent app Muse, launched on September 8, sent the stock up 11.43% on September 21. Over 13 days, Muse accumulated 2.5 million downloads versus 400,000 for Claude, with staggered launches that limit the comparison. And according to Bloomberg, Meta’s API would cost about a quarter of the price of OpenAI and Anthropic. In my view, the first to be weakened will be intermediaries without their own distribution.
When Anthropic’s biggest customers can walk away
Nearly a quarter of 2025 revenue comes from two unnamed customers, about $1.1 billion, and many large accounts are not bound by long-term contracts. The hypothesis that these two customers are developing their own solutions is circulating, but I found no verifiable statement confirming it. The risk is real without being documented: the equation changes if one of them reduces its purchases.
Can Anthropic’s debt really be repaid?
Let us first recall that $518 billion is not financial debt. These are purchase commitments. Reuters also reportedly corrected its text to refer to a horizon of “coming years” rather than “coming year,” according to observers: the nuance is major. Relative to the $65 billion run rate, these commitments represent about eight years of current revenue, a rough order of magnitude. Everything will depend on whether growth continues and prices hold. If both falter, the burden will become heavy.
Anthropic vs OpenAI, two financial worlds that are hard to compare
Comparing the two groups requires harmonizing their revenue recognition methods. The controversy over run rate shows this: an “annualized” figure only makes sense if the formula is identical. Lacking audited data for OpenAI, I refrain from quantifying the gap.
The Anthropic IPO, an animal Wall Street has never seen
The last private valuation was $965 billion in May, and many anticipate about $2 trillion at the IPO. The timetable, initially planned for October, has reportedly slipped to November. On the legal side, I have not seen details on exposure to product liability, a point I will watch in the public version. The market will have to judge whether growth justifies the gap between net loss and promise. And if the battle is fought on margins rather than on models, who will really come out the winner?
FAQ
Is Anthropic’s prospectus public?
No. It is a leaked draft, reported by Reuters.
Is Anthropic profitable?
Not for 2025, with a net loss of $42 billion. It would, however, be operationally profitable in the second quarter of 2026.
When will the IPO take place?
It has reportedly been postponed to November, according to the press. No official date has been published.
