For fifteen years American utilities planned for flat demand. US electricity use barely moved between about 2007 and 2021 as efficiency offset growth, and utility business plans were built on replacing old plants rather than adding new ones. Data centres for artificial intelligence changed that within about two years. Lawrence Berkeley National Laboratory estimated that US data centres used ~176 TWh in 2023, 4.4% of national electricity, and projected 6.7–12% by 2028 (December 2024); the International Energy Agency put global data-centre use at ~415 TWh in 2024 and projected ~945 TWh by 2030 (Energy and AI, April 2025). Those are averages across scenarios that differ by a factor of two, but the local numbers are starker: a single campus may ask for 1 GW at one substation, Texas's grid operator reported ~410 GW of large-load requests by March 2026 against a peak demand record under 90 GW, and Dominion Energy in Virginia reported ~54 GW of data-centre capacity "in various stages of contracting" in October 2026, of which ~12 GW was under signed electric service agreements (reported). This section takes the energy side of the story that AI section 16 tells from the AI side. It covers the load forecasts and why they spread so widely, what a data centre looks like as a load on the grid, how nuclear and gas owners repriced their fleets on hyperscaler contracts (Constellation Energy, Talen Energy, Vistra), the co-location fight at FERC and PJM, large-load interconnection rules (Texas SB 6, PJM's 2026 decision, FERC's June 2026 show-cause orders), behind-the-meter generation, and the question that decides most of the politics: who pays for the grid if the load does not come? In the language of the book's power equation, data-centre demand handed generators a route to market they had lacked for a decade, and made licence to operate — ratepayer consent — the factor most likely to fail.
What the study covers
- 1Overview: the end of flat demand
- 2How much power, and why the forecasts disagree
- 3The data centre as a grid load
- 4The supply response: who sells power to data centres
- 5Co-location: plugging a data centre into a power plant
- 6Interconnecting large loads
- 7Who pays: capacity prices and large-load tariffs
- 8Behind-the-meter and on-site generation
- 9Winners and losers by business model
- 10Outside the United States
- 11Key numbers
- 12Landmark companies and cases
- 13Pitfalls
- 14How an expert thinks about it
- 15Glossary
Technical exhibits
Worked examples, checklists and case studies included with this study.
Large Load Tariff and Bridge Power Worked Example
A stylized 500 MW campus served by a utility: stranded cost under a standard tariff, a minimum-take large-load tariff and up-front payment for dedicated wires, in a plan and an "AI bust" case; then the cost and value of on-site gas bridge power for a 300 MW campus (computed in Python)
Data Centre Power Deal Diligence Checklist
Questions to ask of any data-centre power arrangement — what is secured, who carries the risk, where the energy comes from, price and terms, grid and equipment, licence — with red flags