The stock of AI chips in the world's data centers is growing at a breathtaking pace: roughly 20 million high-performance accelerator-equivalents today, doubling about every nine months and heading toward 200 million by the end of 2028, according to research firm Epoch AI, as reported this week by The New York Times.
Epoch AI's analysis of chip sales data finds that total available computing capacity from AI chips has grown by about 3.3x per year since 2022 — a doubling every seven months. NVIDIA chips currently account for over 60% of total compute, with Google and Amazon making up much of the remainder. The estimates are inferred from revenue data, financial disclosures and analyst reports.
The compounding matters: compute that doubles every nine months grows roughly tenfold every 2.5 years, enabling larger models, more experiments and cheaper inference — a thread running through this week's news, from OpenAI's $2,000 math proofs to gigawatt-scale data center projects.
The surge also shifts the constraints on AI progress. Analysts say the bottleneck moves from scarce compute toward energy, capital and the harder questions of what to do with all that capability: an extraordinary build-out of infrastructure — and the electricity to power it — lies ahead.
