OpenAI Grows Computing Power 17-Fold in Two Years
A new tracker by Epoch AI reveals OpenAI expanded its computing capacity 17-fold over two years, highlighting the massive infrastructure scale required to lead the frontier AI race.
Research group Epoch AI has launched its AI Chip Users explorer, a public dataset tracking the hardware footprints of five leading AI labs. The data shows that OpenAI expanded its compute capacity roughly 17-fold over two years, reaching an estimated 1,743,000 Nvidia H100-equivalent (H100e) GPUs by the end of 2025. This growth is mirrored in OpenAI's power consumption, which surged from 0.2 gigawatts at the end of 2023 to 0.6 gigawatts in 2024, and finally to 1.9 gigawatts by the end of 2025—an amount of electricity capable of powering 1.5 million average American homes.
The tracker reveals a highly competitive frontier. While OpenAI holds a narrow lead, Google DeepMind follows closely with a median estimate of 1,583,000 H100e. Anthropic ranks third with 1,190,000 H100e, followed by Meta Superintelligence Labs at 996,000 H100e, and SpaceXAI at 615,000 H100e. However, high uncertainty ranges exist; DeepMind's capacity could span from 1.01 million to 2.55 million H100e, while OpenAI's ranges from 1.25 million to 2.19 million. Anthropic's interval is 842,000 to 1.72 million, Meta's is 606,000 to 1.64 million, and SpaceXAI's is 551,000 to 700,000.
Operational models differ wildly across these firms. OpenAI and Anthropic primarily rent capacity from cloud providers like Microsoft, Amazon, Google, Oracle, and CoreWeave. Meanwhile, DeepMind and Meta draw from parent-company fleets, and SpaceX rents out spare capacity from its Colossus system to Anthropic and Google. To maintain its lead, OpenAI plans to spend approximately $50 billion on compute in 2026, tripling its 2025 budget.
For AI practitioners, these figures provide a realistic baseline to evaluate competitive claims and model future costs. With one million H100e, a lab can run dozens of large training runs on the scale of GPT-4 simultaneously. Epoch notes that OpenAI's workload has shifted to a 50/50 split between research and inference, compared to 2024 when research dominated. This shift helps developers estimate the massive inference capacity required to serve models at scale and anticipate the cost curves behind subscription pricing.
This is our own summary of reporting by AlphaSignal
