Technology
Verified dataTraining AI doubles every six months
The computing power used to train notable AI systems has grown astronomically — doubling roughly every six months since 2010, and about a billionfold since 2012. Measured in floating-point operations (FLOP).
CSV · 535 rows · 1950–2025
Since 2010, the compute used to train leading AI systems has doubled roughly every six months — far faster than Moore's Law. Frontier models in 2025 used around 10^26 operations, about a billion times more than AlexNet in 2012.
Key findings
- Since 2010, training compute for leading AI systems has doubled roughly every six months.
- Frontier 2025 models used about a billion times more compute than AlexNet did in 2012.
- The trend far outpaces Moore's Law, which doubled roughly every two years.
Explore the data
Coverage: Jan 1, 1950 – Dec 31, 2025 · Global
Free to download and reuse. Get the data behind every new chart →
Sources
Primary source
Epoch AI
via Our World in Data
Computation used to train notable AI systems
Accessed Aug 22, 2026
Training compute (FLOP) of notable AI systems by publication date, 1950–2025.
Methodology
Each point is a notable AI system, plotted by its publication date and the estimated compute used to train it, in floating-point operations (FLOP), on a logarithmic scale. Data are from Epoch AI, distributed via Our World in Data.
What this does not show
Training compute measures an AI system's scale, not its capability or efficiency; newer methods can do more with less. Estimates for many systems are approximate, and “notable” systems are a curated selection rather than every model ever built. The vertical axis is logarithmic — each gridline is 100,000 times the one below.
Creator
Reuse this visualization
Cite this
License: CC BY 4.0
Embed code
<figure style="margin:0;max-width:1600px">
<a href="https://chartive.org/visualizations/ai-training-compute-1950-2025" target="_blank" rel="noopener">
<img src="https://chartive.org/infographics/ai-training-compute-1950-2025.png" alt="Logarithmic scatter chart of training compute for notable AI systems from 1950 to 2025, rising from about 10^2 operations to over 10^26, with landmark models Perceptron (1960), AlexNet (2012), AlphaGo (2016), GPT-3 (2020), GPT-4 (2023) and Grok 4 (2025) highlighted." width="1600" height="2133" style="max-width:100%;height:auto" />
</a>
<figcaption style="font:14px/1.4 system-ui,sans-serif;color:#57534e;margin-top:8px">
Graphic by Mara Ellison, using data from Epoch AI — via <a href="https://chartive.org/visualizations/ai-training-compute-1950-2025" target="_blank" rel="noopener">Chartive</a>
</figcaption>
</figure>Reuse is welcome — the embed keeps a link back to the source.
