The United States has acquired 72% of the world's AI compute since 2022, according to new research from AI research organization Epoch AI. The findings, built from Epoch AI's AI Chip Imports explorer and published on October 27, 2026, were reported by CryptoBriefing.

The data also points to a widening rather than a closing gap. America's share of newly acquired compute reached 78% in the first half of 2026, up from 72% in 2025—a level that matched its cumulative share since 2022.

What the numbers say

In this context, compute refers to the processing power supplied by advanced AI chips—the physical input behind training and running large models. Epoch AI's findings break down into a few key figures. The United States took 72% of newly acquired compute in 2025, matching its cumulative share since 2022. Then came 2026: in the first half of the year, the US share of new compute climbed to 78%.

China, the country most often framed as America's AI rival, sits far behind. Epoch AI puts its share of global AI compute at about 10%.

That figure becomes more interesting next to a second one. By raw chip count, China accounts for around 18% of AI chips. By actual compute performance, however, it holds only 10%.

According to the research, the discrepancy comes from China's reliance on lower-performance chip models, along with its compliance with export regulations.

Following the chips through customs

The methodology is arguably as notable as the results. Rather than relying on company disclosures or industry estimates, Epoch AI built its analysis on customs records.

The approach works because of a geographic quirk in the supply chain. Nearly all advanced AI chips are packaged in Taiwan, meaning they pass through a common chokepoint before shipping out to data centers around the world.

Customs data also makes the chip-versus-compute distinction measurable. Knowing what kind of chips went where allows analysts to estimate actual processing power, not just unit volume.

The new figures line up with Epoch AI's earlier work. The group previously estimated that the United States held approximately 75% of global frontier AI compute capacity as of late 2025.

Why compute concentration matters

A 72% share since 2022 means the bulk of the world's capacity to train and run advanced AI systems physically sits in one country. Frontier-scale model training is among the most compute-intensive work in the industry, which is why the geographic spread of chips has drawn attention from policymakers and researchers alike.

The 18% versus 10% split is the key detail to watch for China. If Chinese chipmakers can close the performance gap, China's compute share could rise even without importing additional units.

Export controls sit at the center of that dynamic. The research ties China's lower compute share partly to compliance with export regulations, which means policy decisions in Washington directly shape the global distribution of AI capacity. Washington has imposed successive restrictions on exports of advanced AI chips to China in recent years.

The research also highlights how much of the advanced chip pipeline runs through Taiwan, making that single packaging hub a point of concentration in its own right.

The trend line is what makes this data notable. A 72% share is already dominant, and watching it tick up to 78% in early 2026 suggests the AI compute race is not converging.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see the Editorial Policy.