AI’s electricity
Data centres use about 1.5% of the world’s electricity today and could approach 3% by 2030. Why it grows less than compute, and what the model’s ceiling means.
How much electricity
The International Energy Agency estimates that data centres used about 415 TWh in 2024, around 1.5% of the world’s electricity. In its base case they will reach about 945 TWh in 2030, just under 3% of the total.Source 1
Not all of that electricity is AI: data centres also serve the web, email and video. But AI is the fastest-growing part.
Why it grows less than compute
The compute of the most advanced models grows 4 to 5 times a year,Source 2 but that figure measures the largest training run at any given time, not everything data centres do. Their electricity, which includes the web, video and everyday use of AI, grows by just under 15% a year in the IEA base case: from 415 TWh in 2024 to 945 in 2030.Source 1 They are different measures, and there is no reason to expect them to grow at the same pace.
Chips also use less and less energy per operation: their energy efficiency doubles every 3 years, about 26% a year.Source 3 That explains only part of the gap.
Efficiency, however, does not guarantee that consumption falls. If each operation gets cheaper, training gets cheaper, and more training happens.
Clean or fossil
What matters most for the climate is not how much electricity AI uses, but where it comes from. That is decided by policy, not technology: in the paths it is one of the decisions.
In the model
In the futures model, AI electricity starts at about 550 TWh in 2026, our interpolation between the IEA figures for 2024 and 2030. Those figures cover all data centres, so at first it overstates AI alone. It grows with AI capability, is tempered by efficiency, and cannot exceed 20% of the world’s electricity. That ceiling is our assumption, not data. In the faster futures the ceiling rules: Peaceful frontier and Race end with similar AI electricity, as the energy topic shows.
What we do not know
The projections depend on how much compute keeps growing, how much chips improve and how much AI is used, not just how much is trained. Beyond 2030, nobody has good figures.
Sources
- International Energy Agency (2025). Energy and AI.
- Epoch AI (2024). Training compute of frontier AI models grows by 4–5x per year.
- Epoch AI (2023). Trends in Machine Learning Hardware.