Anthropic’s S-1 reveals an $850 billion compute gamble

Anthropic’s S-1 puts roughly $850 billion in compute and infrastructure commitments on the company through 2045. That minimum, calculated from disclosed contracts, dwarfs the scale of Anthropic’s current business.

The stakes are direct: Anthropic needs breakneck revenue growth to meet fixed obligations, keep training and eventually turn a profit. Anthropic’s own risk factors say AI-assisted development may advance faster than its ability to validate safety, alignment and security controls.

What does Anthropic’s S-1 show?

Anthropic reported $16.3 billion in revenue for the first half of 2026, fourteen times the first half of 2025. The S-1 also records an $80 billion loss. The analysis describes that loss as largely an accounting item, not a reflection of underlying operations.

Removing that loss and other charges, including employee compensation, the analysis calculates $1.62 billion in adjusted income for the first half of 2026 from the filing. On that basis, Anthropic appears able to pay for its current operations through AI revenue, at least for now.

The filing names only OpenAI and Google as rivals. The source credits that narrow focus to discipline greater than OpenAI or Elon Musk. Anthropic has never released an AI image generator, an AI TikTok clone or an AI “waifu.”

Instead, Anthropic concentrated on coding and cybersecurity, then sold those capabilities to enterprises rather than fickle consumers. That discipline shows up in retention: Anthropic reports net revenue retention of 900 percent. About three-quarters of its first-half 2026 customers spend at least $100,000 a year, according to the S-1.

Can growth outrun competition?

Anthropic still faces pressure from rival frontier labs and open-source models. Claude Opus 5.5 currently leads the Epoch Capabilities Index, but that leaderboard changes frequently. Open-source options serve buyers willing to run software that trails the frontier by four months.

Revenue arrives in bursts after major model launches, so Anthropic cannot stop training between releases. It also has to operate more efficiently than its much larger competitors.

Google has an advantage Anthropic lacks: multiple paths to revenue and AI distribution through products businesses already buy. The source invokes Microsoft Teams’ displacement of Slack as a distribution lesson. Bundling can carry a product even when a rival offers the better standalone tool.

Model quality matters more in cybersecurity than in instant messaging, Anthropic’s chosen field, while Google’s existing distribution still gives it room to pressure Claude’s pricing and reach among enterprises.

What do the contracts demand?

The filing’s “Commitments and contingencies” section supplies the usable numbers after two conspicuous omissions. It leaves purchase commitments blank on page 64 and expected spending before December 31, 2030, blank on page 137. It still says the commitments run through 2045.

• $265.6 billion in fixed, non-cancelable purchase commitments, with terms of one to 20 years.
• $321.4 billion in leases that have not yet commenced, “primarily related to compute and infrastructure capacity,” based on unaudited June 30 figures.
• About $245 billion in new compute and infrastructure agreements, also with one-to-20-year terms.

Unfinished transactions carry an estimated cost of $125 billion to $150 billion, while other compute commitments remain variable. Adding the three disclosed categories yields the roughly $850 billion minimum. Including the unfinished transactions at their current estimate pushes the total toward $1 trillion.

The source cautions that its calculation may omit commitments, and Anthropic supplies no payment schedule for the future obligations. PitchBook analyst Harrison Rolfes models about $90 billion a year in commitments from 2027 through 2030. He says that pace would require $188 billion in revenue to keep the current cost-to-compute ratio stable.

“Growth is not optional for this company,” Rolfes writes. “The contracts assume it.”

Can efficiency support those commitments?

Anthropic spent $2.3 billion in the first half of 2026 on compute and infrastructure associated with customers, plus a few other items. Model development and related costs reached $9.7 billion over the same period.

Rolfes says Anthropic also used less compute for each dollar of revenue. The filing does not provide third-quarter figures.

“The S-1 does freeze in time a happy period when their initial models skyrocketed in usage and they didn’t have enough compute so they are able to charge more for what they had,” says Brendan Burke, an analyst at Futurum Group.

Why does cybersecurity matter?

As enterprise software, Anthropic can target companies’ IT budgets. Cybersecurity forces customers to update constantly because attackers can use AI to brute-force attacks that previously resisted them. A partner’s sloppy handling of security protocols around AI agents also creates product problems, according to the source.

Burke says frontier labs need high-margin products to avoid a race to the bottom on token prices. Anthropic gives it one potential tool: Opus 5.5 reduced customer costs by 40 percent compared with Opus 5.

Anthropic says lower prices produce more usage, which helps preserve margins. If every model release repeats that reduction, customers soon pay a fraction of today’s price for comparable tasks.

Gil Luria, an analyst at DA Davidson, offers the counterweight. “Amazon Web Services and other cloud providers have become far more efficient every year and reduce their prices every year and they have had steady growth and rising margins,” he says.

His example shows why efficiency does not automatically erase profitability. Cloud companies can cut prices while gaining usage and still posting steady growth and rising margins, an outcome Anthropic hopes to reproduce with Claude.

What does recursive self-improvement mean?

Growth pressure leads Anthropic toward recursive self-improvement with minimal or no human supervision. That means using AI to improve the AI-development process itself. The company says: “More capable models help our researchers discover these improvements faster; in turn, more powerful compute multipliers enable us to train even more capable models and deploy them more efficiently.”

The source frames the strategy as building a machine god or failing financially. Anthropic’s general rhetoric about apocalypse leads the source to read the company as expecting dramatic recursive self-improvement, perhaps soon. In practical terms, Claude already helps build Claude under human direction, and the filing says the company has locked itself into that approach.

Anthropic warns that using AI to develop more AI “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 source reads that warning as a business constraint: Anthropic may need to move faster than it can validate systems because its commitments require it.

That risk sits beside the filing’s references to “catastrophic” harm and the development of “forms of sentience.” If dramatic recursive self-improvement outside human supervision proves impossible, Anthropic will have trouble meeting its compute commitments and turning a profit.

The next model cycle will test whether cheaper Claude systems can drive enough new usage to support roughly $90 billion a year in commitments from 2027. Anthropic’s own filing leaves little room for a pause.