AI Data Centers
Evergreen explainer · neutral data, no opinion

GPUs in largest cluster

Current verified value
200k
xAI (Colossus) · single training cluster
Press release · 2025 Human-verified

What it measures

Number of accelerators wired into the largest single training cluster. A coarse but vivid measure of scale: frontier clusters went from ~25k (2023) to 200k-class (2025), with million-GPU sites announced. Counts mix GPU generations (H100, GB200), so read alongside power.

Glossary →

History

257k135k71k37k19kfrontier cluster, 2023202220242025

By actor

xAIColossus200,000
Microsoft (Azure Quantum)OpenAI training sites150,000
OracleStargate Abilene130,000
Meta AIlargest Llama cluster100,000

Goal / baseline

now200k
goal1M
baselinefrontier cluster, 2023

first million-GPU cluster (~2027)