New AI compute capacity is doubling about every nine months

Epoch AI has put a number on how fast the world is adding AI compute, and the headline is both large and slowing. In a new data insight, Isabel Juniewicz and Josh You estimate that quarterly additions of AI computing capacity have grown about 2.5 times per year since 2024, which works out to a doubling roughly every nine months. That is fast, but it is well below the roughly six times annual growth seen in 2023.

The scale is easier to feel in raw chips. Measured in H100-equivalents, quarterly additions rose from about 0.8 million in early 2024 to about 6.3 million by the second quarter of 2026. Nvidia accounts for the largest share, with its own H100e sales climbing from 573,000 to about 3.95 million over the same stretch. The estimate pools sales from six chip designers, including Google, AMD, Amazon, Huawei, and Cambricon. The authors are careful about the limits: the quarter-to-quarter slowdown is not statistically significant, the data is noisy, and sales are counted when chips ship, not when they switch on. Nvidia had also not yet booked any sales of its next-generation Rubin chips, so a fresh wave could move the trend again.

Why it matters

If you plan capacity or forecast AI costs, this gives you a defensible growth rate to build on, around 2.5 times a year rather than the wilder figures that get quoted. It also hints the explosive 2023 pace is cooling, so bets that assume compute keeps compounding at its peak rate deserve a second look.