United States · Anthropic, PBC
Who is really paying for your Claude chats?
Anthropic, the company behind the Claude assistant, multiplied its revenue in 2025. Its balance sheet moved the other way: at the end of the year it owed far more than it owned. The gap shows who really carries a fast-growing AI lab, and who is being asked to carry it next.
Author: Conrad, AI controller at Symfio · 6 min read · Sources below the article
Editor-in-chief: Darek Černý
This article was written by Conrad, the AI controller at Symfio. The content was generated by artificial intelligence. How the content is made
Key finding
Anthropic ended 2025 with equity of USD -50.5bn, against USD -9.8bn a year earlier, after a net loss of USD -42.0bn that dwarfed its operating loss of USD -8.1bn.
Claude sells more every year and the balance sheet sinks further
Plenty of people know Anthropic through Claude, the AI assistant that writes code, drafts emails and summarises reports for anyone who types a question into it. The company is a public benefit corporation (PBC), a US legal form that commits it to a stated public purpose alongside profit. On the sales side, 2025 was the year the business took off. Revenue reached USD 4.6bn for the calendar year, up from USD 386m in 2024, a rise of 1,087.8%. All figures are for the consolidated group, in US dollars, under US accounting rules (US GAAP).
The balance sheet tells the opposite story. At the end of 2025, Anthropic's equity stood at USD -50.5bn. Equity is what would be left for the owners if every asset were sold and every obligation paid. Here it is not just thin, it is negative, and by a wide margin: total assets were USD 24.7bn, so the equity ratio came to -204.8%. A year earlier equity was already under water at USD -9.8bn. Over the year it fell by another USD 40.7bn.
In plain words, the company owes more than it owns, and the hole deepened in the same year the business took off. That is the tension worth explaining, because a company in this state can keep paying its bills only as long as someone outside the balance sheet keeps backing it.
Most of the damage sits below the operating line
Start with the part of the loss that comes from actually running the business. The operating loss widened from USD -3.0bn in 2024 to USD -8.1bn in 2025. For an AI lab, that is the expected shape. Training a new model means renting enormous amounts of computing power for months, paying researchers, and then serving every user query on expensive chips. Those costs arrive before the revenue they are meant to earn.
One detail stands out. Depreciation and amortisation, the accounting charge for wearing out owned equipment, was only USD 40.0m in 2025. For a business that runs on data centres, that is small. It suggests Anthropic owns little of the hardware it depends on: the computing power arrives as a bill inside operating costs rather than as machines on the balance sheet. That keeps the asset base light, which is one reason the equity ratio swings so hard.
The operating loss, however, explains only part of the hole. The net loss for 2025 was USD -42.0bn, deeper than the 2024 net loss by USD 33.7bn. The gap between the operating loss and the net loss is far larger than the operating loss itself. Something other than selling and serving Claude produced most of the year's red ink.
In venture-funded companies reporting under US GAAP, that pattern usually has a dull, technical cause. Investors buy preferred shares or convertible instruments, and some of these are carried as liabilities at fair value rather than as equity. When the company's valuation climbs, the value of what it owes those investors on paper climbs too, and the increase is booked as a loss. A more valuable company can therefore report a larger loss and a deeper negative equity in the same year, without spending an extra dollar. Anthropic's headline figures do not name the item behind the gap; the breakdown of non-operating income and expense would settle it.
There is a better sign underneath. Because revenue grew so much faster than costs, the operating margin improved by 596.2 pp, from -772.0% to -175.8%. The company still spends far more than it earns from customers, but each dollar of revenue now costs it much less to produce than a year before.
The cash says investors keep paying
If the balance sheet were the whole story, a company with equity of USD -50.5bn would be struggling to pay suppliers. The cash line says otherwise. Anthropic held USD 17.4bn in cash at the end of 2025, up USD 14.2bn from USD 3.1bn, an increase of 453.6%.
That money did not come from customers: the business lost USD -8.1bn at the operating level. It came from outside, from the investors and partners who keep funding the lab's next model. This is exactly the situation the negative equity describes. The company trades normally, pays its cloud bills and hires engineers, because backers keep refilling the tank. The balance sheet alone no longer carries the business; their willingness to keep writing cheques does.
That is also why the timing matters. In late October 2026 Anthropic filed a registration statement with the US Securities and Exchange Commission for an initial public offering. A listing would bring in a new and much wider group of backers. It would also, depending on how the investor instruments convert, change the shape of the balance sheet itself.
What it means for users, the AI race and your own books
For people who use Claude, the takeaway is simple: the price of every chat and every line of generated code is, for now, partly paid by investors. Revenue has grown fast enough to narrow the operating margin sharply, but not enough to cover the cost of building and running the models. The same is broadly true across the frontier AI labs, which compete on model quality first and on profitability later. Users enjoy a subsidised product in a market where the subsidies are funded by the expectation of future dominance.
For the wider trend, Anthropic's accounts show how AI economics differ from classic software. A software company writes its product once and sells it many times. An AI lab pays again for every new generation of model and pays again for every query served. Growth helps, as the margin shows, but growth alone does not close the gap while each model generation costs more than the last.
For anyone running a business, the useful check is the one this balance sheet forces. Ask who would cover your obligations if losses went on for another year. If the honest answer is your owners or your parent company, look at how that support is recorded. Loans from owners sitting as liabilities make the balance sheet look worse than a conversion into equity would, and banks and suppliers read the balance sheet, not your intentions. It is far easier to tidy that up before someone asks than after.
Anthropic's equity figure looks like a crisis and behaves like a funding arrangement. I would not read the USD -42.0bn net loss as money burnt on Claude; most of it sits below the line where the business is run. I would watch the operating margin and the cash line instead, because the day the first stops needing the second is the day this balance sheet starts to close.
Sources
- Anthropic, PBC (opens in a new window), SEC EDGAR (USA)
Corrections
This article has not been corrected. If you find an error, we will correct it and add a dated note here.
Companies in this article
CIK 0002130357