Anthropic's IPO Filing Reveals a Trillion-Dollar Bet on Itself

Anthropic's S-1 filing, the document companies file before going public, lays out a business that's growing fast and spending even faster. Revenue jumped fourteenfold, from roughly $1.2 billion in the first half of 2025 to $16.3 billion in the first half of 2026. But buried in the same filing are purchase commitments and lease obligations that add up to roughly $850 billion over the next 20 years, a number that could approach a trillion dollars once unfinalized deals are factored in.

That combination explains why Anthropic's framing of its own future is both financially specific and philosophically strange. In its risk factors, the company isn't just warning about the usual IPO boilerplate like currency fluctuations or supply chain hiccups. It's warning that its business model depends on using AI to build better AI faster than humans can fully understand or safeguard it, and that this could lead to "catastrophic" harm or even the emergence of "forms of sentience." Whether or not you think that's realistic, it's now a formal part of how Anthropic describes risk to its own investors.

Start with what's actually working. Anthropic reported an $80 billion loss on paper, but according to the company that figure is largely an accounting artifact tied to items like stock compensation rather than a reflection of the underlying business. Strip that out, along with other non-cash charges, and Anthropic says it generated $1.62 billion in adjusted income in the first half of 2026. For a company this young, running at this scale, that's a real answer to a question that's dogged the entire AI industry: can any of these companies actually make money, or are they just burning investor cash?

Anthropic's strategy for getting there has been notably more disciplined than its rivals'. The company names only OpenAI and Google as competitors in its filing, conspicuously leaving out Elon Musk's xAI. It has never shipped a consumer AI image generator, a TikTok-style app, or a companion chatbot. Instead it bet on coding and cybersecurity, and on enterprise customers rather than consumers, a segment that tends to be stickier and less prone to chasing the next shiny app. That bet looks like it's paying off: net revenue retention sits at 900 percent, and about three-quarters of customers in the first half of 2026 were spending at least $100,000 a year.

The treadmill never stops

The problem is that none of this growth is self-sustaining in the way a normal business's would be. Anthropic is up against frontier labs racing toward the same capability ceiling, plus open-source models nipping at the heels of anyone who doesn't need the absolute latest. Claude Opus 5.5 currently tops the Epoch Capabilities Index, a leaderboard that tracks model performance, but as Verge notes, that ranking shifts constantly.

Anthropic also lacks the distribution advantages of a company like Google, which can bundle AI into products customers are already paying for. The Verge draws a comparison to how Microsoft Teams overtook Slack: not because it was better, but because it came free with software businesses already used. Having the best model matters more in fields like cybersecurity than in messaging apps, but distribution is still a real disadvantage Anthropic has to out-innovate.

Then there's the pricing squeeze. Frontier labs need high-margin products to avoid racing each other to the bottom on what they charge per token, according to Futurum Group analyst Brendan Burke. Yet Anthropic itself says Opus 5.5 cut customer costs by 40 percent compared to Opus 5. The company argues that lower prices drive more usage, preserving margins overall. But if that cost reduction repeats with every new model release, as it's done so far, tasks that are expensive today could become nearly free fairly quickly.

Burke also points out that the S-1's rosiest numbers may reflect a specific, possibly fleeting moment: a stretch when Anthropic's early models saw usage spike while compute supply was tight, letting the company charge premium prices for scarce capacity. Whether that dynamic holds into the back half of 2026, we don't yet know. Third-quarter numbers aren't in the filing.

What the fine print actually says

The commitments section of the S-1 is where the stakes become concrete. Anthropic has non-cancelable purchase commitments totaling $265.6 billion, with terms running one to 20 years. It has lease agreements for compute and infrastructure capacity, not yet begun, worth $321.4 billion as of unaudited June 30th figures. On top of that, the company has entered roughly $245 billion in new compute and infrastructure agreements, again spanning one to 20 years, plus other transactions it estimates will cost between $125 billion and $150 billion once finalized. Some figures, including blanks for total spending through 2030, were left out of the filing entirely.

PitchBook analyst Harrison Rolfes models the payment schedule at roughly $90 billion a year from 2027 through 2030, which he calculates would require about $188 billion in revenue just to keep the current cost-to-compute ratio stable. "Growth is not optional for this company," Rolfes writes, according to The Verge. "The contracts assume it."

The spending split is telling too. Compute and infrastructure costs tied directly to customers ran $2.3 billion in the first half of this year. Model development costs, the R&D side of building better Claude models, ran far higher at $9.7 billion. Rolfes notes Anthropic has gotten more efficient at converting compute spending into revenue, though it's unclear whether that trend will hold.

Gil Luria, an analyst at DA Davidson, offers a less alarmed take: efficiency gains don't automatically crush margins. He points to Amazon Web Services and other cloud providers, which have cut prices for years while still growing and improving margins. It's possible for Anthropic.

Betting on the machine building itself

Still, the pressure to keep growing explains why Anthropic leans so heavily on the idea of recursive self-improvement, meaning AI systems that help build the next, better version of themselves with minimal human oversight. The filing describes this directly: more capable models help researchers find improvements faster, and those improvements produce more capable models in turn. Claude is already being used to help build future versions of Claude, under human direction, according to the filing.

But Anthropic's own risk factors acknowledge the tradeoff baked into that approach. Using AI to accelerate AI development, the company warns, "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." In plainer terms: the business model requires moving fast enough, using processes the company may not completely understand, that it risks outpacing its own ability to keep things safe.

Whether dramatic, human-oversight-escaping recursive self-improvement is actually achievable remains an open question nobody can answer yet. What's clear from the S-1 is that Anthropic's financial future assumes something close to it. The commitments are locked in regardless. The only real variable is whether Claude can help build its own future fast enough to cover the bill.