What exactly is a computing bill?
OpenAI does not own most of its computers. It rents them by the hour from Microsoft, Oracle and Amazon, the way you hire a taxi instead of buying a car. The rental price rolls everything into one: the NVIDIA chips, the building, the electricity, the cooling. That rent is the computing bill, and it runs on two meters.
Meter one is training: building the brain. Before a new model can answer anyone, thousands of chips run non-stop for months teaching it. Like school fees, paid before the student earns anything. Verified figures: about 9 billion dollars in 2025, about 19 billion planned for 2026, rising every year after.
Meter two is inference: running the brain. Every single question from 900 million weekly users costs a few paise of computing to answer. Tiny per question, enormous when multiplied by billions of questions a day. Like petrol: cheap per trip, huge across a fleet. This meter quadrupled in 2025 because ChatGPT grew faster than expected, and OpenAI expects over 150 billion dollars of it between 2025 and 2030. Note the trap: unlike a normal shop, more customers means a bigger bill, because every answer costs money to produce.
Why do the bills “land” in 2027?
OpenAI has signed giant rental contracts, about 250 billion dollars with Microsoft, roughly 300 billion with Oracle, 38 billion with Amazon, for data centres that are being built right now.
Rent starts the moment each building switches on, whether OpenAI has enough customers yet or not. Many of these switch on in 2026 and 2027. It is like signing leases for ten shops that all open the same year: the rent explodes before the sales do.
How was 122 billion estimated?
Honest answer: it is simple addition from OpenAI’s own investor documents, reported by The Information. The company told investors it expects about 62 billion of income in 2027 and a cash shortfall of roughly 57 to 63 billion that year. Money out = money in plus the shortfall: 62 + 60 = about 122 billion.
ChatGPT Business Model and how it loses Billions