The compute used to train notable AI models has doubled about every 5.7 months since 2010, against roughly every 20 months in the decades before. Researchers count that compute in FLOP (floating point operations), the total number of arithmetic operations performed during training. This article uses an open dataset of 457 notable models published from 2010 to 2025 to show how fast it has grown.

The short version

• Doubling every 5.7 months. Across the notable models with a compute estimate, training compute grew about 4.3 times a year from 2010 to 2025 (a doubling every 5.7 months, 95% bootstrap interval 5.4 to 6.1 months).
• Much faster than before. For the 62 models published before 2010 the same method gives a doubling every 20.5 months (interval 17 to 25.5), roughly the pace of Moore's law for chips.
• A very large step in a decade. The median model with a compute estimate used about 7 x 10¹⁸ FLOP in 2015 (11 models) and about 3.9 x 10²⁴ FLOP in 2025 (40 models), around 550,000 times more.
• The frontier grew at the same rate. The record holders for "largest model so far" (24 of them since 2010) also double about every 5.6 months.
• Who builds them changed. Models developed by industry alone went from 1 of 10 notable models in 2010 to 98 of 107 in 2025 (92%).

What a number like 10²⁵ means

Large numbers are hard to feel, so here is one way to scale them. If all 8 billion people on Earth each carried out one calculation per second, without stopping, they would manage about 2.5 x 10¹⁷ operations a year. A model trained with 10²⁵ FLOP would take them about 40 million years. The largest estimate in 2025 (5 x 10²⁶ FLOP) would take about 2 billion years.

The very large models have become routine

A decade ago almost no model reached 10²³ FLOP. The table counts how many notable models per year passed three thresholds (only models with a compute estimate are counted).

From university labs to company labs

Epoch AI records the type of organisation behind each notable model. In 2010 most (8 of 10) came from academia alone. By 2025 almost all came from industry alone, which fits the cost of the compute: the largest training runs need hardware budgets that few university groups have.

What this does and does not show

• Only about half of the models have a compute estimate. Of 868 notable models published from 2010 to 2025, 457 (53%) have a training compute figure. Developers disclose less about the largest systems, so the biggest recent models may be under-represented.
• The figures are estimates. Epoch AI derives compute from reported hardware, training time, parameters and training data, and rates its confidence. Small differences are not meaningful. The doubling times describe a trend fitted through scattered points, not a law.
• "Notable" is a selection. Models enter the dataset through criteria such as state of the art results, citations or historical significance, so it is a curated list, not every model ever trained.
• Compute is not capability. More FLOP does not mean a better model. Better algorithms, data and chips change how much compute a given result needs, and this analysis says nothing about that.
• Precision and hardware differ. FLOP counts are not always measured at the same numerical precision, which adds noise.
• No forecast. A trend this steep cannot continue forever. Chips, electricity, data and money all limit it, and this article does not predict when or how.
• Time window. Models are dated by publication. Because the dataset is updated continuously, we excluded the partial year 2026 and used full years 2010 to 2025.

Method and data

Data: Epoch AI's notable AI models table, downloaded in October 2026. For each model we used the publication date and the "Training compute (FLOP)" column, keeping the 457 models published from 2010 through 2025 that have an estimate (and 62 published before 2010 for comparison). We fitted a straight line to the base 10 logarithm of compute against publication date (ordinary least squares). A doubling time of d months follows from the slope s in orders of magnitude per year as 12 x log₁₀(2) / s. Intervals come from 2,000 bootstrap resamples of the models. Record holders are the models that exceeded every earlier model's compute. Developer type uses Epoch's organisation categories ("Industry only" means no academic organisation is listed).

Citation and licence: Epoch AI, "Data on AI models", published online at epoch.ai, https://epoch.ai/data/ai-models, notable models CSV updated 24 October 2026, accessed 24 October 2026. The data is licensed by Epoch AI under Creative Commons Attribution 4.0 (CC BY 4.0). Swift Lake filtered the data, fitted the trend lines, computed the counts and shares, and drew both charts, so all derived figures and charts are our own calculations and are not endorsed by Epoch AI or the authors of the underlying models.

Sources and further reading:

• Epoch AI, Data on AI Models (the dataset page) and the notable models CSV used here.
• Epoch AI, AI models documentation (definitions, notability criteria and how compute is estimated).
• Sevilla, Heim, Ho and others, Compute Trends Across Three Eras of Machine Learning (arXiv, 2022). It reports that before 2010 training compute doubled roughly every 20 months, in line with Moore's law, and about every 6 months after the start of deep learning, which is consistent with the figures above. A summary is on the Epoch AI blog.
• Background on Moore's law (Wikipedia).