Key points from the chart
- The chart plots the training computation of notable AI systems from 1950 to 2025, on a logarithmic scale from 100 to 100 septillion FLOP.
- From 1950 to 2010, training computation doubled about every 21 months.
- Since 2010, with the rise of deep learning, it has doubled about every 6 months.
- Labelled systems run from Theseus (1950) to GPT-4 and Gemini 1.0 Ultra (2023), the highest labelled points.
All figures are from the graphic, which cites Epoch (2024).
The AI training compute growth chart in text
Title: “The computation used to train notable AI systems has doubled every ~6 months since 2010”. Subtitle: “Training computation is measured in total floating-point operations (FLOP). Each FLOP represents a single arithmetic calculation, such as multiplication. Shown on a logarithmic scale.” Logo: Our World in Data.
Each AI system is a dot. The horizontal axis is the year (1950, 1965, 1980, 1995, 2010, 2025) and the vertical axis is total training computation, labelled from the top:
- 100 septillion*, 10 to the power 26
- 10 sextillion, 10 to the power 22
- 1 quintillion, 10 to the power 18
- 100 trillion, 10 to the power 14
- 10 billion, 10 to the power 10
- 1 million, 10 to the power 6
- 100
Two dashed trend lines are annotated: “From 1950 to 2010 it doubled every ~21 months” and “With the rise of deep learning since 2010, training computation has doubled every ~6 months”. The period from 2010 is shaded. Eight systems are named on the chart; it prints no value for them, so the table gives only their position in time.
| Labelled system | Position on time axis | Note |
|---|---|---|
| Theseus | about 1950 | Earliest system on the chart |
| Perceptron Mark 1 | about 1957 | |
| Back-propagation | about 1986 | |
| TD-Gammon | about 1992 | |
| AlexNet | 2012 | Breakthrough in computer vision [1] |
| AlphaGo Zero | 2017 | |
| GPT-4 | 2023 | |
| Gemini 1.0 Ultra | 2023 | Highest labelled system |
Footnote: “*For comparison, 1 septillion (1,000,000,000,000,000,000,000,000) is the estimated number of stars in the universe.” Source line: “Data source: Epoch (2024)”. Licence: “CC BY”, by the author Charlie Giattino.
How fast AI training compute has grown
Our World in Data published the chart as a data insight by Charlie Giattino and Veronika Samborska in January 2025. It says that before deep learning took off around 2010, training computation doubled about every 21 months, and since then roughly every six months. As an example it compares AlexNet, the 2012 breakthrough in computer vision, with Google’s Gemini 1.0 Ultra eleven years later, which used 100 million times more training computation [1].
The data comes from Epoch AI. Our World in Data measures it in petaFLOP, where one petaFLOP is 10 to the power 15 floating point operations, estimated from the AI literature with some uncertainty [2]. Epoch’s own analysis found that training compute had grown by a factor of 10 billion since 2010, doubling about every 5 to 6 months [4].
Why compute keeps growing
Our World in Data’s article on scaling up AI explains that larger training runs have been a main driver of more capable AI. It also notes that the hardware used for training, GPUs, gets cheaper and more powerful, with computing speed per dollar doubling roughly every 2.5 years [3]. The same article tracks training data, which has doubled about every nine to ten months since 2010, and model parameters, which have roughly doubled every year [3].
Notes on this infographic
- Log scale. Each labelled line on the vertical axis is 10,000 times the one below it, so a straight line on the chart means steady exponential growth.
- Estimates. Training compute for most systems is estimated from published details rather than reported by developers, so individual points carry some uncertainty [2].
- Data date. The chart uses Epoch data from 2024; Our World in Data’s chart page is updated from newer Epoch releases [2].
Questions about AI training compute
How much is one septillion?
One septillion is 10 to the power 24, written out on the chart as 1,000,000,000,000,000,000,000,000. The chart’s footnote compares it to the estimated number of stars in the universe.
Sources
- Our World in Data, Since 2010, the training computation of notable AI systems has doubled every six months, the original post for this chart.
- Our World in Data, Computation used to train notable artificial intelligence systems.
- Our World in Data, Scaling up: how increasing inputs has made artificial intelligence more capable.
- Epoch AI, Compute Trends Across Three Eras of Machine Learning.
Last reviewed 29 October 2026
