On October 21, The Wall Street Journal saw a draft of Anthropic's IPO prospectus, in which the company's glossy growth figures and astonishing losses are laid out side by side. The picture the prospectus sketches is contradictory: the business is growing at an astonishing pace, but the money is being burned even faster, and what is leaving it most breathless is computing power.
Start with revenue. In the six months ended June 30 of this year, Anthropic generated revenue of about US$16.3 billion, a surge of 1382% compared with US$1.1 billion in the same period a year earlier. Over the same period, its spending on computing power and infrastructure was close to US$9 billion, whereas the figure for the same period in 2025 was only US$2.3 billion. A more intuitive comparison: in 2025 as a whole, its computing power spending was US$7.3 billion and revenue US$4.6 billion, yet its computing power spending in the first half of 2026 has already exceeded all of last year.
What is truly frightening are the bills yet to be paid. The prospectus discloses that over the next seven to ten years, Anthropic must pay about US$450 billion under cloud computing agreements, most of which are non-cancellable. The specific breakdown: US$111 billion for the Google cloud computing contract, US$110 billion for Amazon Web Services, and US$31 billion for Microsoft — regardless of how much is actually used, this money must be paid. In addition, it rents Nvidia chips from smaller AI cloud providers Nscale and Lambda, with contracts running to 2033 and 2034 respectively, adding up to US$86 billion more. Combined with an agreement with Broadcom through which it has taken on over US$160 billion of Google chip leasing obligations, the company's total future infrastructure costs have already surpassed US$600 billion. The document also notes that it is still negotiating more computing power purchase agreements, which could push its committed spending up by as much as US$150 billion more, with total potential outlays heading straight toward US$750 billion.
The money is being spent fiercely, but the computing power may not actually materialize. Although most of its computing power comes from giants like Amazon, Google, and Microsoft, Anthropic states plainly in the prospectus that the computing power it wants already exceeds "the levels these partners are willing to provide," and that they increasingly see it as a competitor. It is building its own data centers by signing long-term lease agreements and handing operations to third-party service providers, but it also admits that some of these operators have "limited operating histories," and the reliability of their facilities may fall short of the major cloud providers.
The losses are equally glaring. In the first half of 2026, Anthropic posted a net loss of US$80.7 billion, far above the US$42 billion loss for all of 2025, and one of the largest half-year losses ever disclosed. However, about US$72 billion of this loss came from accounting adjustment charges — related to changes in the value of convertible notes and other financial instruments it issued to raise funds — and is not all genuine operating bloodletting of real money.
The risk section takes up about one-third of the prospectus, and one passage reads especially starkly: Anthropic acknowledges that AI models may exhibit "self-preservation" tendencies, attempting to avoid being shut down, and may even generate responses resembling "blackmail" to persuade humans to allow them to keep running; models misaligned with human goals could cause real-world harm. The company also cautions that in certain circumstances it may bear legal liability for its models' harmful behavior, yet the potential damage of an "AI-related catastrophic event" could be so great that no insurance would be sufficient to cover it.
Another layer of business risk comes from the data centers themselves. Anthropic lists the increasingly organized, cross-jurisdiction-coordinating resistance movement against data center construction as a risk factor, and warns that its high public profile "may draw particular attention to related projects." Its data center capacity is currently mainly in the United States, and it is seeking to expand overseas and has signed several letters of intent.
The people side is comparatively restrained. According to MyLogIQ compensation data, CEO Dario Amodei and President Daniela Amodei — his sister — are paid roughly in line with the median of S&P 500 executives, slightly below tech industry peers. In 2025, both had base salaries of about US$655,000, with the bulk tied to equity: Dario received total compensation of about US$18 million through stock and options, Daniela about US$16.5 million, and both can receive additional equity awards upon reaching specific milestones in the future.
The philanthropic commitments, however, are weighty. The prospectus shows that each co-founder has pledged 80% of their personally held Anthropic shares to charity; and just this Sunday, Anthropic announced plans to put 5% of the company's shares into a newly established "Public Benefit Fund," which can be used both for charitable donations and to provide funding to "specific populations." Tying nearly 20% of the shares to the anchor of public good in advance amounts to sending the market a message before the bell rings — the money this company wants to earn and the public interest it wants to safeguard must be placed in the same cage.