Shows final training runs are a small share of lab compute, with most spent on research and experiments. That changes how to read training-cost figures.
Key takeaways · AI-distilled
Reports in The Information and The New York Times put OpenAI's 2024 spend at $3B on training compute, $1.8B on inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition →, and $1B on research compute amortized over multiple years.
Epoch assumed a two-year amortization for research compute, giving $2B in 2024, and grouped it with training as $5B of R&D because how OpenAI separates training and research compute is unknown.
The GPT-4.5 final training run estimate carries a wide 90% confidence interval of $170M to $890M, and final runs for all other released models, including GPT-4o, GPT-4o mini, Sora Turbo and o-series post-training, total about $80M.
That leaves an estimated $4.5B for basic research, derisking runs ahead of final training, and unreleased models, since final-run estimates exclude the experimental runs used to prepare them.
Because OpenAI owned little AI compute in 2024 and rented from Microsoft, these cloud expenses represent essentially all its compute, not data-center capital costs.
Terms in this piece · Glossary
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
