Are AI models profitable, and who pays when it is free?

In the 1st quarter of 2026, OpenAI earned about $5.7 billion in revenue with an adjusted operating margin of -122%, excluding stock-based compensation, according to documents reported by The Information: the company spent $2.22 in operating costs for every dollar it took in (our calculation). Its annualised revenue, one month multiplied by 12, was about $30 billion in May; OpenAI is aiming for more than $40 billion by the end of the year. Anthropic's exceeded $65 billion at the end of July 2026, 7 times higher than a year earlier, according to Bloomberg sources cited by The Decoder. These annualised rates are not accounting revenue: they assume that the latest month is repeated 12 times.

At equal performance, the price charged per token fell each year by a factor of 9 to 900 depending on the task, between 2021 and early 2025, according to Epoch AI; this is a price, not the provider's cost. No provider publishes the cost of a response, and usage per subscriber can exceed what the price covers, as Sam Altman said of the Pro plan (see below).

Alphabet plans $195 to $205 billion of capital expenditure in 2026, Amazon about $220 billion, Meta $135 to $145 billion; Microsoft spent $39 billion in the quarter ending in September alone, 2 times more than a year earlier. The 5 largest technology companies invested more than $400 billion in their infrastructure in 2025 and plan more than $700 billion in 2026. This money does not fund AI models alone: it also serves the cloud computing sold to other companies, including the model providers themselves. Google Cloud reports $561 billion of contracts yet to be billed; Microsoft reports $677 billion of remaining commercial performance obligations, all services combined.

This spending weighs on cash: Alphabet had negative free cash flow of $9.1 billion in the 3rd quarter of 2026, and Amazon of $42.4 billion over the last 12 months. In 2024, David Cahn of the Sequoia fund already estimated at $600 billion a year the revenue that AI would need to generate to make the chips bought profitable. In September 2025, the consultancy Bain estimated that $2,000 billion of annual revenue would be needed in 2030 to fund the expected computing, and that $800 billion would be missing, once the savings that AI would allow companies to make are counted.

ChatGPT had about 1 billion weekly active users in August 2026, most of them on the free plan or on Go, the cheapest of the paid plans. Their questions are paid for in 3 ways.

• By investors, who cover the providers' losses. Part of this money comes from chip suppliers: in September 2025, Nvidia had announced an investment of up to $100 billion in OpenAI, by a letter of intent; in February 2026, Reuters reported that Nvidia was close to completing an investment of $30 billion that would replace this commitment.
• By the profits of the large groups, which fund their data centres with the cash from their other activities: advertising, commerce and software. Part also goes through the debt of financial partners: Meta's Hyperion campus in Louisiana relies on about $27 billion of debt carried by the Blue Owl fund, maturing in 2049, with Meta as tenant.
• By advertising, tested in ChatGPT since February 2026 on the free plan and on Go. It brought in more than $1 billion a year at the end of August; OpenAI is aiming for $2.5 billion by the end of the year.

Paying subscribers and businesses fund the rest. The price of a subscription does not say whether it is profitable: in January 2025, Sam Altman, the chief executive of OpenAI, said publicly that the $200-a-month plan was losing money, because its subscribers used it more than expected.

At Alphabet, 65% of the 3rd quarter 2026 infrastructure investment goes to servers and network equipment, depreciated over 6 years; the remaining 35% goes to buildings, depreciated over 7 to 40 years. If demand does not follow, chips lose their value within a few years; buildings, grid connections and the power plants built to supply them remain.

In 2025, OpenAI announced at least 10 GW of Nvidia systems, by a letter of intent, and signed with AMD for 6 GW, with the first gigawatts of each expected in the 2nd half of 2026. Bain estimates at 200 GW the additional computing needed worldwide in 2030, half of it in the United States. In France, data centre projects had reserved nearly 18 GW of grid connection in May 2026; our analysis on electricity explains what this reservation represents for the grid. Some of the centres planned in the United States will be powered by gas, on site: our analysis of gas gives the figures. A gas plant built for a data centre that runs less than planned keeps emitting if it finds other customers, or becomes a stranded asset.