US Secures 72% of Global AI Compute Share Since 2022, Epoch AI Data Shows
New research from Epoch AI reveals the United States has acquired 72% of global AI compute since 2022, with its share expanding to 78% in early 2026.
US Secures 72% of Global AI Compute Share Since 2022, Epoch AI Data Shows
New research from Epoch AI reveals the United States has acquired 72% of global AI compute since 2022, with its share expanding to 78% in early 2026.
US Dominance in AI Compute
The United States has acquired 72% of the world's AI compute since 2022, according to new research published by Epoch AI. Built using the organization's AI Chip Imports explorer, the data highlights a widening lead for America in the global infrastructure race.
America's share of newly acquired AI compute reached 78% in the first half of 2026. This indicates that the gap between the United States and other nations is continuing to grow rather than narrow over time.
China Position and Performance Gap
China, frequently framed as the primary rival to the United States in artificial intelligence development, holds approximately 10% of global AI compute. However, examining raw chip counts reveals a different picture, with China accounting for roughly 18% of physical AI chips.
This discrepancy stems from China's reliance on lower-performance chip models and compliance with international export regulations. Analysts suggest that if Chinese manufacturers can bridge this performance gap, the country's compute share could rise without increasing the physical volume of imported units.
Customs Records and Supply Chain Chokepoints
The methodology behind Epoch AI's findings relies on tracking customs records rather than depending solely on corporate disclosures or industry estimates. This tracking is made possible by a specific geographic chokepoint in the global supply chain.
Nearly all advanced AI chips are packaged in Taiwan before being shipped to data centers around the world. By examining customs data through Taiwan, researchers can measure actual processing power rather than just unit volume.
About The Author
