Compute Power by Geography
The global distribution of AI/ML compute is staggeringly unequal. We mapped every major GPU cluster to show who controls the infrastructure powering the AI revolution.
This story uses Epoch AI’s GPU Clusters dataset — an open database tracking 786 GPU clusters and AI supercomputers worldwide, spanning 2010 to 2025. Each cluster is measured by its peak FP16 (16-bit floating point) throughput, the standard benchmark for AI training performance. Of these, 482 clusters have confirmed locations and performance data, which is what we’ve visualised below.
The United States alone holds 77.6% of global FP16 throughput. China is a distant second at 10.8%. The rest of the world shares scraps.
Before 2020, compute was distributed more evenly. Then the AI scaling race began. The US didn’t just pull ahead — it lapped everyone.
Data: 482 existing GPU clusters from Epoch AI’s dataset. Performance measured in peak FP16 throughput. Some clusters lack coordinates or FLOPS data and are excluded from specific visualizations.
Where the Compute Lives
Global AI compute power by country