Quarterly AI chip compute additions grow about 2.5× a year, doubling every 9 months
Summary
The computing capacity of AI chips sold worldwide each quarter has been rising about 2.5× per year since 2024, an analysis by Epoch AI finds. That implies a doubling time of 9.1 months, with a 90% confidence interval of 8.4 to 10.0 months and a growth range of 2.3× to 2.7×. New capacity measured in H100-equivalents (H100e) climbed from 0.8 million in Q1 2024 to 6.3 million in Q2 2026.
The current pace is well below 2023. That year, sales of Nvidia GPUs and Google TPUs surged after the launch of ChatGPT and GPT-4, with Nvidia's H100 leading the wave, and growth ran at roughly 6×. Year-over-year growth in quarterly additions slipped from 2.8× in Q2 2025 to 2.2× in Q2 2026. A regression of quarter-to-quarter growth shows a decline from about 32% to 19% per quarter, but Epoch AI says the trend is not statistically significant (p = 0.26). The group adds that the dip may partly reflect a gap between chip generations. Nvidia has begun shipping its Rubin chips but had recognized no Rubin revenue as of its fiscal quarter ending in July.
The estimate draws on quarterly sales figures for six chip designers: Nvidia, Google, AMD, Amazon, Huawei and Cambricon. Those figures come from the AI Chip Sales hub, which relies on company disclosures, analyst estimates and media reports. Epoch AI used the hub's median estimates as of October 19, 2026, and fit a log-linear trend to the ten quarters from Q1 2024 through Q2 2026.
The analysis tracks quarterly additions rather than the installed base, because cumulative figures can overstate the current pace after a sales spike. Installed capacity grew 2.6× between Q2 2025 and Q2 2026, close to the additions trend and down from 3.5× the year before.
Why it matters
Chip shipments are the upstream signal for what data center operators must house, power and cool. A steady growth rate of about 2.5× a year in new compute gives capacity planners a firmer baseline than the 2023 surge, when 6× growth distorted forecasts. The finding that installed-base growth has converged with the additions trend also matters for investors. Models that extrapolate from the 2024–2026 rise in installed capacity would overstate demand for new halls and power. The possible slowdown, from about 32% to 19% quarterly growth, is not statistically confirmed. Even so, it is a reason to stress-test aggressive buildout assumptions.
Expert Lens
Each quarter's chip additions translate into a power and cooling requirement that has to be ready on arrival. A sustained doubling roughly every nine months likely keeps pressure on interconnection queues and ready-for-service dates more than on chip supply itself. The generation gap Epoch AI flags may also bunch deliveries. If Rubin volumes ramp sharply, operators could face a sudden wave of higher-density racks rather than smooth growth, which favors sites with liquid cooling already in place. The signal to watch is the first quarter in which Nvidia recognizes Rubin sales, and whether quarterly additions jump back above the 2.5× trend.
Sources
This brief was drafted with AI assistance from public sources; every number is attributed to the sources above. See the AI-assisted Disclosure and Methodology.