110 MW
estimated power draw while training Grok 3 (Feb 2025); Llama 4 Behemoth needed about 44 MW
Largest training power draw to date (log scale)
20122025
What the chart shows
The power needed to run a frontier training run has grown by several orders of magnitude. Grok 3 (February 2025) drew an estimated 110 megawatts during training, the most of any model in the data. Llama 4 Behemoth needed about 44 MW and Llama 3.1-405B about 23 MW.
The earliest model in the series, a 2011 deep autoencoder, drew about 246 watts. Power draw is not available for many of the newest models, because developers disclose too little.
Key numbers
| Measure | Value |
|---|---|
| Grok 3 (Feb 2025) | 110 MW |
| Llama 4 Behemoth (Apr 2025) | 44 MW |
| Llama 3.1-405B (Jul 2024) | 23 MW |
| GPT-4 (Mar 2023) | 20 MW |
| GPT-3 175B (2020) | 5.1 MW |
How to read it
- Power draw is the estimated electrical load of the training hardware cluster, not total energy used.
- Only 34 models have enough public information for an estimate. Hover over a dot to see the model.
- The vertical axis is logarithmic.
Source: Epoch AI, 2026, via the AI Index Report 2026 (Stanford HAI), figure 1.4.2: Total power draw required to train frontier models, 2011–25. Chart and text by AI Adoption.
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 19 October 2026.
Download the data (CSV) · Cite this chart
Discussion