Scaling Intelligence: The Exponential Growth of AI's Power Needs is a White Paper published in August 2025 by the Electric Power Research Institute (EPRI), the primary research-and-development organization for US investor-owned utilities. Running roughly 30 pages with appendices, it documents the trajectory of AI power demand through 2030 from the perspective of the electric-power industry rather than the AI industry, covering frontier-model training power, total AI data center capacity, hardware energy efficiency, training-run duration, and the consequences for utilities, grid operators, and regulators.
Publisher: Electric Power Research Institute (EPRI) Released: August 2025 Format: White Paper (~30 pages with appendices)
Summary of argument
The paper approaches the AI buildout from the energy-infrastructure side, treating the grid that must power AI computation as its subject. EPRI's vantage point as the R&D body for US investor-owned utilities makes the document a utility-side reference on AI power demand, complementing AI-industry and AI-research analyses of the same buildout. Epoch AI's Trends in AI Supercomputers tracks the compute systems being built; this paper tracks the grid impact of powering them.
Document structure
| Section | Coverage |
|---|---|
| §1 Frontier AI Training Power Demand | Historic growth, compute scaling (historic + future), reasoning models' impact on training compute, hardware energy-efficiency growth, training-run duration growth, 2030 forecast |
| §2 Total AI Power Demand | Current level, growth rate, training-vs-inference allocation split |
| §3 Conclusion | Implications for the energy sector |
| Appendix A | Planned large-scale data centers and training clusters (catalog) |
| Appendix B | Hardware efficiency long-term limits (Landauer's bound + practical limits at server, data center, utilization levels) |
| Appendix C | Power demand forecast methodology |
| Appendix D | Power demand trend in AI supercomputers (cross-references to Epoch-style data) |
Key findings
The paper's structural findings, per its executive-summary framing, are:
- Training power demand is on an exponential growth curve with no near-term plateau in sight.
- Compute scaling continues, though the report addresses how reasoning models could shift the training/inference compute balance.
- Hardware energy efficiency is improving at chip, server, and data-center levels, but not fast enough to offset compute growth.
- Training-run duration is growing, implying that single training runs are multi-month commitments at gigawatt-scale power draw.
- Total AI power capacity in the US data center sector is on a trajectory that will reshape the grid through 2030.
- Geographically distributed training could mitigate local power constraints but introduces other tradeoffs (latency, security, sovereignty).
- Implications for the energy sector are central to the paper's intended audience of utilities, grid operators, and regulators.
The full quantitative figures behind these findings require deeper reading of the paper; the items above are its structural conclusions. The forecast methodology underlying the 2030 projections is set out in Appendix C.
Reception and relevance
The paper functions as a utility-industry primary reference for several AI-policy threads that invoke the energy buildout. It serves as a utility-side complement to AI Environmental Impact, whose existing coverage has been drawn primarily from the AI-industry side, and as a counterpart to the AI-research-organization analysis in Epoch AI — How Much Power Will Frontier AI Training Demand in 2030?, with EPRI supplying the energy-industry vantage. It is a companion compute-versus-grid reference to Trends in AI Supercomputers — Pilz, Sanders, Rahman, Heim (Epoch AI / Georgetown / GovAI, 2025): together they describe the AI buildout from both ends, with Epoch AI on the compute systems being built and EPRI on the grid that has to power them. It is also positioned as a reference for a planned AI Data Centers concept page on AI data center buildout.
The "50 GW of new power for AI by 2028" estimate in The AI Grand Bargain — Ben Buchanan and Tantum Collins (Foreign Affairs, October 2025) is calibrated against utility-industry analysis of this kind, for which the EPRI report serves as a primary source. The energy-buildout pillar of America's AI Action Plan is calibrated against the same body of utility-industry analysis. The Trump DPA Determination 2026-10, which invokes Defense Production Act Section 303 to classify transformers, transmission components, substations, and grid power electronics as essential to defense, rests on the AI-power-demand trajectory documented in this paper, which provides the empirical basis for that designation.
EPRI is described as institutionally credible and politically neutral within the energy-industry context.
Confidence note
Confidence is high for the documented historic data and the utility-industry consensus about the trajectory, and medium for the 2030 forecasts, which depend on the continuation of several independent scaling trends.
Relationships
- supports: AI Environmental Impact (canonical utility-industry primary reference), Presidential Determination 2026-10 (Defense Production Act Section 303 — AI Grid Components) (empirical basis for the grid-hardware DPA designation), The AI Grand Bargain — Ben Buchanan and Tantum Collins (Foreign Affairs, October 2025) (energy-bottleneck calibration source), America's AI Action Plan (energy-pillar calibration source)
- related: Trends in AI Supercomputers — Pilz, Sanders, Rahman, Heim (Epoch AI / Georgetown / GovAI, 2025) (companion compute-side reference), Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? (companion AI-research-side reference), AI Data Centers, Scaling Laws (compute-scaling input), Electric Power Research Institute (planned)
- instance-of: Compute Governance (energy-side complement to compute-side)