Nvidia Chip Production

In a February 2025 data insight, Epoch AI's Luke Emberson and David Owen estimated the total computing power available across Nvidia's data-center GPUs, counting chips still in service rather than sales in a given year (epoch.ai). They found that this stock has grown about 2.3x a year since 2019 (90% interval 2.2x to 2.5x), which the headline expresses as doubling every 10 months. At the time it stood at about 4e21 FLOP/s, or roughly four million H100-equivalents, and the Hopper generation accounted for 77% of it. Older generations typically fall below half of cumulative compute about four years after their introduction.

The estimate starts from a dataset of known AI clusters, which captures only about 20% of spending, and scales it up using Nvidia's reported data-center revenue, assuming 80% of that revenue comes from accelerators. It then depreciates chip prices by 30% a year and removes GPUs as they fail, assuming a five-year average lifespan. Epoch reports that varying these assumptions leaves the growth rate at 2.3x to 2.4x a year. Gaming sales are excluded, and Blackwell GPUs did not yet appear in the data when it was published. A separate Epoch series instead counts the computing capacity of AI chips sold across vendors.