A 2025 Epoch AI blog post projecting the site-level electricity demand of the largest individual AI training runs through 2030. The piece extrapolates from historical training-compute growth to argue that a single frontier run could require 4–16 GW of continuous site power by 2030, up from more than 100 MW for the largest known runs at the time of writing.
Summary of argument
Epoch's central projection is that straightforward extrapolation of historical trends yields frontier training runs drawing 4–16 GW by 2030, a range the piece assigns medium-to-high confidence and attributes to differing assumptions about efficiency and utilization. The post observes that historical training-run power has grown at roughly 2.2× per year, faster than hardware efficiency gains of about 1.3× per year in FLOP/watt, so that rising compute demand outpaces the per-watt improvements that would otherwise hold power flat.
The argument the piece advances is that power, rather than chips, may become the binding constraint on frontier scaling by the late 2020s. It notes that this would shift data-center siting from tax-driven considerations toward interconnect-queue-driven ones, and identifies behind-the-meter generation (gas, nuclear power-purchase agreements, and small modular reactors) as a plausible response that faces its own supply timelines.
Key claims
- The largest known single training runs drew more than 100 MW of continuous site power at the time of writing, a figure Epoch assigns high confidence.
- Historical training-run power has grown at roughly 2.2× per year, against hardware efficiency gains of about 1.3× per year in FLOP/watt.
- Extrapolation yields frontier training runs of 4–16 GW by 2030 (medium-to-high confidence), the range reflecting efficiency and utilization assumptions.
- Global AI data-center capacity was approximately 30 GW as of late 2025; a single 2030 frontier site at the upper bound of the projection would equal roughly 50% of the entire global AI data-center fleet of that period.
Methodology
Epoch combines its Notable AI Models training-compute series with accelerator FLOP/watt curves (H100 to B200 to next-generation parts), utilization assumptions of 35–50% of nameplate capacity, and power usage effectiveness (PUE) of 1.1–1.3, then amortizes the result over a 3–9 month wall-clock training window.
Caveats stated in the source
The piece notes several factors that could change its framing. Distributed training across multiple sites could undercut the single-site power assumption. Algorithmic efficiency gains, including Chinchilla-style rebalancing of model and data scale and increased test-time compute, could reduce training power for a fixed capability level. And it observes that aggregate inference power already exceeds training power and is growing faster.
Relation to other wiki pages
The projection provides a quantified trajectory on the power axis of the compute-chokepoint question and supplies the empirical baseline for the scaling-continuity analysis in Epoch AI — Can AI Scaling Continue Through 2030?. It supports the energy-availability element of Compute Governance by quantifying the energy-bottleneck hypothesis, and it prices the physical cost of continued compute scaling examined in Scaling Laws. The 4–16 GW per-run bracket sits inside the 2–10 GW central single-run range in RAND — AI's Power Requirements Under Exponential Growth (2025), though above RAND's median. On AI Environmental Impact, the per-run figure partially supersedes the previously anchoring single AI-Index figure of 29.6 GW, which measures aggregate fleet capacity rather than per-run demand. The Least Understood Driver of AI Progress, from the same author organization, treats software-efficiency progress as the counterweight to the power-growth curve described here.
Relationships
- supports: RAND — AI's Power Requirements Under Exponential Growth (2025) — same direction of travel on the largest-run power numbers (Epoch's 4–16 GW bracket sits inside RAND's 2–10 GW central range for single-run but above RAND's median).
- supports: Compute Governance — quantifies the energy-bottleneck hypothesis.
- supports: Scaling Laws — prices the physical cost of continued compute scaling.
- supersedes: (partial) the AI-Index 29.6 GW single-figure claim on AI Environmental Impact, which is aggregate-fleet rather than per-run.
- related: The Least Understood Driver of AI Progress — same author org; software-efficiency progress is the counterweight to the power-growth curve here.
Provenance
Epoch AI. "How Much Power Will Frontier AI Training Demand in 2030?" epoch.ai/blog/power-demands-of-frontier-ai-training, 2025.