New AI compute capacity is doubling every nine months, down from a 2023 surge, with quarterly H100-equivalent sales broken out by vendor. This grounds expectations about future model scale and availability.
Key takeaways · AI-distilled
Epoch AI estimates quarterly AI chip capacity additions rose from 0.8 million H100-equivalents in Q1 2024 to 6.3 million in Q2 2026, a fitted trend of 2.5x per year and a doubling time of 9.1 months.
Epoch fits quarterly additions rather than cumulative stock because the 2023 surge inflated cumulative growth; by 2026 the two converged, with cumulative capacity up 2.6x year over year versus 3.5x the year before.
Year-over-year growth in quarterly additions fell from 2.8x to 2.2x, but Epoch says the apparent slowdown is not statistically significant (p = 0.26) and may reflect a lull between chip generations.
The data count chips when sold, not when running: sellers like Nvidia typically book revenue on delivery to server assemblers, so capacity can take months more to come online, and Epoch's intervals ignore uncertainty in the sales estimates.
