Anthropic's newly released S-1 filing reveals the company has locked itself into a minimum of roughly $850 billion in compute and infrastructure commitments over the next two decades, with that figure potentially climbing to around $1 trillion once pending transactions close. The filing shows a company that grew its revenue fourteen times to reach $16.3 billion between the first half of 2025 and the first half of 2026, but one that now faces what analysts describe as mandatory growth to meet its fixed obligations. According to the filing, the company has pinned its survival to dramatic expansion, with no option to slow down.
The financial picture shows stark contrasts. While Anthropic reported an $80 billion loss, the filing notes this is largely an accounting matter not tied to core operations. Stripping away that charge along with employee compensation and other items leaves the company with $1.62 billion in adjusted income for the first half of 2026. About three-quarters of customers in that period spent at least $100,000 annually, and the company's net revenue retention hit 900 percent. The company's non-cancelable purchase commitments total $265.6 billion with terms spanning one to 20 years, while lease agreements for compute and infrastructure that haven't yet started add $321.4 billion. After June 30th, Anthropic entered additional agreements worth roughly $245 billion, again with one-to-20-year terms, plus other deals estimated between $125 billion and $150 billion that aren't yet finalized. Model development costs ran $9.7 billion in the first half of this year, while compute and infrastructure tied to customers cost $2.3 billion.
PitchBook analyst Harrison Rolfes models the payment schedule at approximately $90 billion annually from 2027 through 2030, requiring $188 billion in revenue to maintain current cost ratios. "Growth is not optional for this company," he writes, adding that "the contracts assume it." The filing acknowledges 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 company states that more capable models help its researchers find improvements faster, and stronger compute multipliers let it train even more capable models and deploy them more efficiently. Its Claude Opus 5.5 currently leads the Epoch Capabilities Index, though the filing notes that leaderboard changes frequently.
The filing frames the company's strategy around recursive self-improvement as both a business necessity and a calculated risk. Anthropic has focused on enterprise markets and tasks like coding and cybersecurity rather than consumer products, competing mainly against OpenAI and Google. But it faces a distribution disadvantage against rivals like Google, which can bundle AI into existing business software the way Microsoft Teams displaced Slack. Revenue arrives in spurts following new model releases, and the filing notes that Opus 5.5 cut customer costs by 40 percent compared to Opus 5, though the company says lower prices drive higher usage that preserves margins. The core tension is clear: Anthropic warns in its risk factors that it's locked into using AI to build AI faster than it can fully validate safety measures, driven by business pressures to meet massive financial commitments. If dramatic recursive self-improvement beyond human supervision doesn't materialize as the business plan requires, the filing suggests meeting existing obligations and turning a profit becomes far more difficult. The path forward depends entirely on whether the company can continue improving efficiency and sustaining explosive growth for two decades—an open question even for believers in the technology's potential. For enterprises weighing vendor lock-in, the S-1 exposes an uncomfortable reality: betting on a frontier lab now means betting on an industry-wide race that no single player can afford to lose.