How do AI labs actually make money, and why do they still lose so much?

For years, outsiders had to guess how AI companies were doing. That has changed. Anthropic and SpaceX, which now owns xAI, have both published prospectuses with audited numbers, and large partners such as Microsoft disclose part of their dealings with OpenAI. OpenAI itself has not published a prospectus, so its figures below come from press reports or from what it told investors.

The picture that emerges is a business with very fast sales growth, enormous fixed costs and accounting that needs some translation.

The five ways a lab earns money

The mix differs a lot. Anthropic's prospectus shows usage-based fees bringing in $13.1 billion of its $16.3 billion in revenue for the first half of 2026, or 80%, with subscriptions at 20%. No customer exceeded 10% of revenue over the first nine months of 2026, though the top five together made about 25%.

OpenAI leans further toward consumers. CNBC reports that it had 2025 revenue of $13.1 billion and burned about $8 billion in cash that year, citing people familiar with its finances. OpenAI said in March 2026 that it was generating $2 billion a month. "Burn" means the cash a company spends beyond what it takes in.

xAI is different again. In 2025 its segment, which includes the X social network, had revenue of $3.2 billion from ads, subscriptions and data licensing, against an operating loss of $6.4 billion and $12.7 billion of capital spending. In May 2026 it also agreed to rent its Colossus data centres to Anthropic for $1.25 billion a month through May 2029, a deal either side can end on 90 days' notice. So one lab became a landlord to a rival.

Where the money goes

Here is Anthropic's statement in simplified form, because it is the most complete one available, in millions of US dollars:

Two things stand out. Compute cost $1.60 for every dollar of revenue in 2025, then about 55 cents in the first half of 2026. And in the first half of 2026 the net loss is more than 100 times the operating loss.

That gap is mostly a non-cash item. Anthropic owes convertible notes, loans that turn into shares. When its valuation rises, those notes are worth more, and accounting rules record the increase as an expense: $72.4 billion in the first half of 2026. No money left the company for that line. Operating cash flow over those six months was positive, at $3.9 billion, though the filing attributes part of the swing to the timing of payments to suppliers.

Three reasons losses persist

Costs arrive first. A data centre contract is signed years before it earns anything. Training the next model is paid for out of today's sales.

Models age quickly. A leading model may command a premium for only months before a rival or an open-weight model catches up, so research spending does not stop.

Customers can leave. Anthropic warns that its revenue is largely consumption-based, while its compute contracts run for years. A customer can cut usage overnight; a data centre lease cannot be cut so easily.

How a run-rate number is built

Labs love to quote "run-rate revenue." It is an annualized snapshot of recent sales, not what was actually earned. Anthropic's definition takes usage fees from the last 28 days and multiplies by 13, then adds current monthly subscriptions times 12.

Suppose a lab earned $900 million in usage over 28 days and has $100 million in monthly subscriptions:

• Usage: $900 million × 13 = $11.7 billion
• Subscriptions: $100 million × 12 = $1.2 billion
• Run-rate: $12.9 billion

If usage then falls, that figure falls with it. The filing itself warns that the measure is not a forecast. Anthropic also estimates that 5% to 7% of its revenue over nine months came from customers it later determined were breaking its usage policy or commercial terms, a share it expects to decline as enforcement improves. That is a reminder that even audited revenue deserves questions.

How can you tell a real profit from a flattering number?

Compare three lines rather than one. Revenue shows sales actually recognized. Operating income shows whether the business covers its running costs. Cash flow from operations shows whether money really came in, and you should then subtract capital spending.

Be careful with "adjusted" figures, which remove costs the company considers unusual. Anthropic reports an adjusted operating income of $1.6 billion for the first half of 2026 against an operating loss of $693 million; the difference is mainly $2.2 billion of stock paid to employees, which is a real cost to shareholders even if no cash moves. Adjusted numbers can be useful, as long as you read what was removed.

Private companies such as Cohere, Mistral or OpenAI publish far less, so treat any reported figure as a claim until a filing confirms it. Nothing here is a recommendation about any investment; it is a way of reading the numbers.