How much compute the top five AI labs actually run

Epoch AI put numbers on a question the industry usually answers with rumor: how much computing power does each frontier lab actually control? Its new AI Chip Users Explorer estimates holdings across five labs in Nvidia H100-equivalent GPUs as of December 2025. OpenAI comes out on top at roughly 1.74 million H100-equivalents, just ahead of Google DeepMind at about 1.58 million. Anthropic follows near 1.19 million, Meta's Superintelligence Labs around 996,000, and SpaceXAI about 615,000.

The estimates come with wide error bars, which is the honest part. Epoch says OpenAI likely held a narrow lead over DeepMind, but the ranges overlap enough that the order is not certain. DeepMind's true figure, they note, could sit anywhere from about 1 million to 2.5 million H100-equivalents. The trajectory is clearer than the ranking: OpenAI roughly tripled its compute each year from 2023 to 2025 and reached about 1.9 gigawatts of power capacity, which Epoch compares to powering around 1.5 million American homes at once. The lab splits that capacity roughly evenly between research and training on one side and inference for its billion-plus users on the other.

Why it matters

If you track the AI race, this replaces guesswork about who has the biggest cluster with sourced estimates and explicit uncertainty, so you can stop treating leaked GPU counts as fact. The even split between training and serving is a useful reminder that a lab's compute lead does not translate directly into faster model progress, because much of it goes to keeping current products running.