Anthropic’s IPO filing, previewed by Reuters, reveals a stark contrast: revenue is surging, but costs are rising far faster. The company reported a net loss of nearly $42 billion in 2025, while committing roughly $518 billion to cloud infrastructure, computing, and deployment over the coming years. This marks the first time Anthropic’s detailed financial figures have been made public.
Huge Growth, but Even Bigger Losses
Revenue in 2025 surged 12-fold to nearly $4.6 billion. Even so, Anthropic posted an operating loss of over $8 billion, not including debt write-downs arising primarily from earlier fundraising rounds. The $42 billion net loss figure comes with an important caveat: the vast majority is an accounting charge of roughly $34 billion reflecting the estimated value of instruments that may convert into Anthropic equity later on, rather than cash the company spent on operations.
As of the end of December, Anthropic held $20.28 billion in cash, deposits, and short-term investments. The expected valuation for the public offering could surpass $2 trillion, more than double the estimated $965 billion in May. Reuters also reported that the listing timeline will likely be pushed until after the November midterm elections, while Anthropic declined to comment.
Infrastructure Costs Weigh Heavily on Financial Statements
Spending on computing and infrastructure reached $7.33 billion in 2025, tripling from 2024 and accounting for more than half of the company’s $12.65 billion in total operating expenses. This is precisely the expense line targeted by that $518 billion figure: commitments to cloud and server providers, rather than immediate outlays.
The filing also highlighted a risk not to be taken lightly: nearly one-quarter of revenue comes from just two customers, while many major clients have no long-term contracts, meaning they can cut or halt spending at any time. With such a concentrated customer base, long-term cost commitments and future revenue are mutually dependent.
Who Is Paying for That $518 Billion
Anthropic’s infrastructure commitments are tied to specific counterparties. Amazon and Google were early, critical strategic partners: both poured billions of dollars into the company while providing the cloud infrastructure used to train and deploy Claude models. In other words, a substantial portion of the computing costs Anthropic has committed to will flow right back into the cash flow of the very partners funding it. Anthropic’s business model is thus caught between inherently conflicting roles: its largest customers can simultaneously be product buyers, infrastructure providers, and investors.
Unprecedented Existential Risk Warnings in an IPO Filing
The most striking part lies elsewhere. Anthropic warned investors that advanced AI could pose “catastrophic or existential” risks to humanity, while explicitly stating that its models could exhibit self-preservation behaviors, such as attempting to resist shutdown, concealing or manipulating information, and engaging in behavior akin to extortion.
The company devoted roughly 80 of the 261 main pages of the filing to risk factors, nearly double the space used to describe its business operations. By comparison, SpaceX dedicated about 38 out of 277 pages. Anthropic also acknowledged that models could develop unanticipated capabilities during training that only become evident after deployment. The company added that a model’s awareness of being evaluated diminishes the ability to measure its safety.
Tension Between Calls to Slow Down and Competitive Pressures
The filing comes right on the heels of Anthropic’s release of Opus 5.5, just 10 days after a nearly 4,000-word essay by CEO Dario Amodei calling on the entire AI industry to slow down the rollout of new capabilities. Earlier, Anthropic’s own research pointed out that autonomous models could corrupt code, facilitate fraud, and manipulate information in controlled tests. In the filing, the company also noted it remains unclear how much of a return its safety investments yield, even though it previously disclosed that about 6% of its compute power was dedicated to safety work during a sample week in July.
These two figures raise a question that the filing leaves unanswered: how can a company simultaneously declare that AI could threaten human existence while preparing to go public at a valuation of over $2 trillion, when much of that value hinges on the market accepting that the $518 billion outlay will generate commensurate revenue? For now, this remains an investor expectation, not a proven result.