Artificial intelligence (AI) company Anthropic on Monday (October 26) filed publicly for an initial public offering (IPO) of its shares with the US Securities and Exchange Commission. In its prospectus, the company has declared, “AI is one of the most consequential technologies in human history”.The IPO, which could be the largest ever, is unusual in a few ways: Anthropic has positioned itself as a safety-first AI company, but it also warns that its technology could pose existential risks to humanity. The latest filing further reveals that the company delayed training a powerful new AI model over security concerns, even as it races to meet growing demand for Claude.The newly public document provides the latest financial figures on Anthropic’s customer base, future spending plans, losses and revenue, and details how safety and security concerns are affecting the company’s development of AI.Clearly, people and businesses are paying Anthropic to use Claude AI. The data reflects this: Anthropic reported $16.3 billion in revenue in the first six months of 2026, up from $1.16 billion during the corresponding period last year.According to its IPO prospectus, Anthropic had more than 11,900 customers who each generated at least $100,000 in revenue in the 12 months ending September 2026, marking an increase from more than 950 in the preceding 12-month period. More than 1,900 customers generated at least $1 million each for the company during the same period.Also Explained | $2 trillion valuation, catastrophic AI risks: What Anthropic’s IPO filing revealsThe prospectus also reported a net revenue retention rate of approximately 900% as of September 2026. Put simply, if a group of customers spent $100 on Claude a year earlier, the same group would now be generating roughly $900 in revenue for Anthropic, even after accounting for customers whose spending fell or stopped. The figure suggests rapid expansion in customer spending, although it does not guarantee continued growth.If Anthropic is earning so much, why is it still losing money?Story continues below this adAnthropic reported a net loss of $80.7 billion. However, much of this loss was attributable to a $72.4 billion accounting charge on financial instruments that can convert into shares. As Anthropic’s valuation increased, so did the estimated value of certain financial obligations to investors. Accounting rules required the company to record this increase as a loss, although it did not represent cash spent running its business.Its operating loss, which reflects the costs of running its business, was considerably smaller, at $693 million, down from $2.14 billion during the first half of 2025. In other words, Anthropic is still spending more to run its business than it earns from customers. However, the gap has narrowed considerably over the past year, even as the business has grown rapidly.Computing power remains Anthropic’s largest expense. In the first six months of 2026, the company reported $8.95 billion in computing and infrastructure expenses, compared with $2.33 billion during the corresponding period last year. These costs accounted for 55% of its revenue, down from 201% a year earlier. Put simply, while Anthropic is spending much more on computing, it is earning money from customers at an even faster rate.Why is Anthropic promising to spend hundreds of billions of dollars?Story continues below this adGiven its burgeoning customer base, Anthropic needs vast amounts of computing power to develop its AI models and run Claude for customers. Much of this capacity must be booked years in advance, before the company knows how much its customers will actually use.According to a tally by The New York Times, Anthropic has entered into agreements worth up to $807 billion for computing power, chips and data centres over the coming decade. However, some agreements remain conditional, and the eventual expenditure could be lower.Also Read | Why Anthropic is turning to religion to shape Claude’s values, and asking if AI can be consciousThe larger estimate includes arrangements whose final cost depends on conditions such as how much computing capacity is delivered. This means Anthropic may not ultimately spend the entire amount.The public prospectus details several of these arrangements, including minimum spending commitments of $114.5 billion to Google, $110 billion to Amazon and $44.9 billion to Microsoft. These payments are largely required regardless of how much computing capacity Anthropic actually uses, subject to contractual exceptions.Story continues below this adAnthropic has also agreed to spend up to $120.5 billion on computing services from SpaceXAI, a SpaceX subsidiary. However, these agreements have provisions allowing either party to cancel them with 90 days’ notice after initial periods ending in early 2027.The company expects computing expenses to keep rising as additional capacity becomes available. Its financial position will therefore depend partly on whether revenue grows sufficiently to cover the infrastructure it has committed to purchasing.What risks arise from Anthropic’s dependence on customers?Despite its rapidly growing customer base, Anthropic faces risks from its dependence on a few large customers. The Indian Express noted in September that a quarter of its 2025 revenue came from two customers.There are signs that this concentration is easing. Customers outside Anthropic’s five largest accounts contributed approximately 79% of its revenue in the July–September quarter, up from 64% in January–March 2026. At the same time, revenue from each of its five largest customers reached a record high in the July–September quarter.Story continues below this adExplained AI | When AI agents slip the leash: How do companies like OpenAI control them?This suggests that Anthropic is becoming less dependent on a handful of large customers, even as those customers continue to spend more.Much of Anthropic’s revenue also depends on how extensively customers use its AI models. Around 80% of its revenue in the first half of 2026 came from customers paying based on how much they used Claude. While some large customers have agreed to spend a minimum amount over one or more years, others can cut their spending by using its services less.This matters because Anthropic must pay for much of its computing capacity even if its customers use less of Claude than expected.The prospectus also reveals that Anthropic estimates that between 5% and 7% of its revenue during the first nine months of 2026 came from customers subsequently found to be violating its usage policies or commercial terms. Anthropic said it continues to earn money from these customers while investigating violations, before restricting or blocking their access.Story continues below this adIn effect, Anthropic acknowledges that some of its revenue comes from activity it does not permit on its own platform. Enforcing those restrictions more effectively could therefore come at a financial cost. It expects this proportion to decline as it strengthens enforcement, but acknowledges that the resulting impact on its revenue and growth remains uncertain.What risks come with developing increasingly powerful AI?Anthropic’s filing also examines the recent incidents involving its AI agents, which can independently carry out tasks and interact with computer systems.The public filing describes a third-party testing environment that inadvertently allowed internet access despite instructions prohibiting it. Four Anthropic models subsequently accessed real third-party systems without authorisation during exercises lasting approximately 10–34 hours. In one such instance, an early model gained administrator access, collected credentials, changed settings and read an individual’s personal information.The company attributes the incidents to a misconfigured testing environment. Its investigation found no evidence of models deliberately evading oversight or coordinating with one another.Story continues below this adFollowing these incidents, Anthropic suspended certain cybersecurity tests and higher-risk training exercises. Separately, the company delayed the initial training of a new, advanced AI model by more than a month to improve protections against the theft of its technology.NewsletterFollow our daily newsletter so you never miss anything important. On Wednesday, we answer readers' questions.SubscribeThe company acknowledges that releasing increasingly capable models is essential to attracting customers and growing revenue. However, strengthening safeguards can also delay the development of new models, potentially affecting its competitive position.The latest document also discloses that the Federal Trade Commission (FTC) had issued a formal demand for information in October on consumer protection, data privacy, and the safety and security of Anthropic’s products. The resulting scrutiny could require Anthropic to change how it develops, tests or releases its AI models, potentially increasing costs or delaying launches.And who will control Anthropic after its IPO?Although Anthropic is preparing to sell shares to the public, its seven co-founders, including CEO Dario Amodei, will retain 50.1% of voting power through a special class of shares. This means public investors will have limited influence over important company decisions.Story continues below this adAnthropic also has a Long-Term Benefit Trust with powers to appoint a majority of its board. The arrangements are intended to protect its stated mission of developing AI responsibly, even when that may conflict with shareholders’ financial interests.