Anthropic’s confidential IPO prospectus, as reported by Reuters, is less a victory lap than a balance-sheet reveal. The Claude maker generated nearly $4.6 billion in 2025 revenue, roughly twelve times 2024, while posting more than $8 billion in operating losses, $7.33 billion in compute and infrastructure expense, and about $518 billion in future cloud, computing, and infrastructure obligations. Reuters said the offering could value Anthropic above $2 trillion and come after the November U.S. midterms, though neither is final.
The practical question for investors is not whether frontier-model demand exists. It plainly does. The question is whether Anthropic is becoming a scalable software company or an AI utility whose economics will be set by utilization, contract flexibility, and customer concentration.
Software growth on top, utility economics underneath
The reported filing gives both sides of that argument real weight. Revenue is surging, with about $3.8 billion of 2025 sales coming from usage-based Claude consumption and another $789 million from subscriptions. Anthropic also ended the year with $20.28 billion in cash, cash equivalents, and short-term investments. That is the profile of a company reaching escape velocity in a large market.
But the cost structure looks much closer to industrial infrastructure than conventional software. Total operating expenses were about $12.65 billion, and the $7.33 billion compute and infrastructure line alone was roughly triple 2024. The headline net loss of about $42 billion overstates the immediate operating strain because roughly $34 billion came from an accounting charge tied to financing arrangements that could convert into equity. Even after setting that aside, the operating loss still exceeded $8 billion. For public investors, that is the number that matters more: how much money the business loses before financing effects, and whether scale improves or worsens that relationship.
The $518 billion obligations figure is the sharpest expression of that tension. It is a stack of multi-year commitments, not a one-year spending plan, and the commitments are not economically identical. Public reporting describes more than $111 billion owed to Google through 2033, $110 billion to Amazon through 2036, $31.4 billion to Microsoft through 2033, and about $161.2 billion in Broadcom-related equipment lease obligations. Up to $84.5 billion of Nvidia-based capacity from xAI appears more flexible, with most of it reportedly cancellable on 90 days notice. AMD commitments also involve up to $5 billion of Anthropic stock purchases and more than $20 billion in expected computing capacity. The missing piece is how much of the total is fixed minimum take, how much is reservable capacity, and how heavily Anthropic will actually use it. That is why the prospectus reads less like a software growth story than an infrastructure utilization test.
Partners, channels, and concentration
The filing also shows how much of Anthropic’s growth is intertwined with the same companies that supply its infrastructure. Amazon and Google accounted for about 47% of 2025 sales, according to Reuters’ follow-up reporting, and they collected roughly $351 million in distribution fees on about $2.16 billion of marketplace sales. Those relationships solve real problems. Cloud platforms bring enterprise distribution, access to scarce compute, and financing support at a moment when capacity can be the constraint on growth.
They also concentrate risk. Amazon and Google are investors, cloud suppliers, distributors, and potential competitors at the same time. Marketplace distribution expands reach but takes margin and weakens direct customer ownership. On top of that, two unnamed customers each represented about 12% of revenue, meaning nearly a quarter of sales came from just two buyers. Anthropic reportedly warned that many large customers do not have long-term contracts and could reduce or stop spending. That combination, high concentration with limited contractual lock-in, is unusual for a company that may soon be valued on the stability of its future cash flows. Usage-based revenue can look recurring right up until a customer slows a workload.
What the IPO really puts on trial
This is why the most important questions around Anthropic’s IPO are not the exact valuation or the timing of the listing. They are operational. How much of the obligation stack is cancellable? How full can Anthropic keep the capacity it has reserved? Does more revenue arrive through direct enterprise relationships or through partner marketplaces that take fees and mediate the customer bond? And if model efficiency improves faster than demand, does Anthropic enjoy better margins, or does it wind up holding more contracted capacity than it needs?
The bullish case is easy to see. Anthropic has extraordinary revenue growth, major enterprise demand, substantial cash, and access to infrastructure partners that many rivals cannot match. In a market where compute availability has often been the bottleneck, locking in supply can be rational. The bear case is not that demand is fake. It is that demand may still be too variable, too concentrated, or too weakly contracted to support a cost base built on long-dated infrastructure commitments. Anthropic’s reported prospectus does not settle whether frontier AI is best valued as software or as infrastructure. It does something more useful: it shows that the answer will depend less on abstract excitement about AI and more on whether paid usage can keep an enormous chain of compute obligations productively full.
By
By
By
By
By
