The compute behind AI

What is measured

To train an AI model, a computer performs an enormous number of mathematical operations. That number, training compute, is the most widely used measure of a model’s size. It does not measure what the model can do, but models with more compute tend to be more capable.

A peer-reviewed study gathered the compute of models since 1950. Before 2010 it doubled every 20 months, in line with Moore’s law; with the arrival of deep learning it began doubling roughly every 6 months.^{Source 1}

Today’s pace

Epoch AI, the group that maintains that database, estimated in 2024 that the compute of notable models grew about 4.1 times a year between 2010 and May 2024, and that of frontier models about 5.3 times. Since 2018 the frontier has grown closer to 4 times a year, and language models about 5 times since mid-2020. Their summary: 4 to 5 times a year.^{Source 2}

Growing 4 to 5 times a year for a decade multiplies compute by between one and ten million.

Where it comes from

Chips improve, but much more slowly: the performance per dollar of AI GPUs doubles every 2.1 years, less than 1.4 times a year.^{Source 3} If the price of each operation only falls at that pace, the rest of the growth has to come from devoting more chips, more energy and more money to each training run.

Where the chips are made

More than 90% of the most advanced chips are made in Taiwan.^{Source 4} The United States restricts their sale to China, and in January 2026 it allowed some models to be exported under quotas.^{Source 5} Compute depends not only on technology but also on geopolitics.

What it means for the futures

In the futures model, the maximum frontier is today’s pace, 4.5 times a year, and the minimum, about 1.35 times a year, is what chip performance alone improves: it doubles every 2.3 years. That is why the today point sits at the right edge of the matrix. The model does not allow compute to grow faster than today, and it does not count algorithmic improvements, which make the same compute go further.

What we do not know

No exponential trend lasts forever. Energy, chip manufacturing and money set limits, and we do not know when they will bite. Nor do we know how much more capable a model becomes for each multiplication of its compute: compute measures effort, not results.