ai accelerator As of August 25, 2026

AWS Trainium & Inferentia

3.28 zettaFLOP/s

4.1% of the world's compute · accreted

Amazon's answer to buying NVIDIA — in-house accelerators deployed at hyperscale across AWS regions, and the fastest-compounding vendor slice in the partition. Project Rainier alone is one of the largest single AI clusters on the planet.

Installed compute

3.28 ZFLOP/s

peak dense BF16 FLOP/s

Change, 1 year

+256.9%

3.3M H100e today

Share of world

4.1%

of total installed compute

In the headline

Yes

~30.0% utilized

← All constituents

AWS Trainium & Inferentia over time

Installed compute in zettaFLOP/s, log scale, 2008 → present. A straight line on this axis means steady exponential growth; the slope is the doubling rate. History before the 2026 anchor is reconstructed — see the methodology.

How it's counted

Epoch AI fleet estimate x Amazon custom-silicon share, converted at dense BF16.

~1.2M H100-equivalents, dominated by Trainium2 and the Trainium3 ramp.

Overlap treatment: Amazon's very large NVIDIA fleet is counted in the NVIDIA line; this is Annapurna-designed silicon only.

Growth: compounding at 4.80× a year (5.3-month doubling) between published vintages.

Source Epoch AI — Data on AI Chip Sales

Full methodology →