The Demand Curve
The International Energy Agency's April 2026 assessment is the most cited number in this thread: global data centre electricity consumption grew 17% in 2025, with AI-focused facilities growing roughly 50% IEA, April 2026. Data centres now consume roughly 460 TWh per year globally — more than the entire electricity use of countries like Turkey or France — and the IEA projects that to roughly double to ~945 TWh by 2030 IEA, April 2026 Partly substantiated.
The Doubling
Global data centre electricity consumption, TWh · solid = observed (IEA), shaded band = 2030 projection range · AI-focused facilities grew ~50% in 2025 alone
In the United States, data centres already consume about 180 TWh per year — roughly 4.4% of national electricity — and projections for 2030 range from 400 to 600 TWh depending on the forecaster EPA / DOE / Goldman Sachs / EPRI estimates Partly substantiated. At the high end, US data centres would consume more electricity than all American households' lighting, or roughly the output of 100 large power plants running flat out.
When One Building Out-Eats One City
The scale is easier to grasp at the project level:
- Stargate (OpenAI / Oracle / SoftBank), Abilene, Texas: designed for over 1 gigawatt of draw — the equivalent consumption of roughly 750,000 homes Project announcements / The Register Company claim
- Meta's Hyperion, Louisiana: projected to consume more electricity than the city of New Orleans at full buildout Meta / press reporting, 2025 Company claim
- Northern Virginia, the world's largest data centre market, has repeatedly delayed new connections because Dominion Energy cannot deliver power fast enough Utility filings / press
In PJM — the grid operator covering 13 states including Virginia — interconnection queues stretch four to seven years for new large loads. Multiple hyperscaler projects have been announced with completion dates that precede their legally available grid connection dates. The gap between announced capacity and deliverable power is one of the least-discussed risks in the entire buildout.
Water: The Invisible Bill
A modern AI data centre using evaporative cooling can consume millions of litres of water per day — a single large facility's cooling load can rival a small city's municipal supply Company sustainability reports / academic estimates. Microsoft's water withdrawals rose 34% between 2021 and 2022 as its AI buildout began Microsoft environmental report; Google reported similar increases. Both companies attribute the rise primarily to data centre expansion.
The tension is geographical: many of the cheapest, most buildable sites (Phoenix, Atlanta, Dallas, parts of the Middle East) sit in water-stressed basins. The industry's response has been a shift toward closed-loop and liquid cooling, which cuts water use dramatically but raises capital cost and shifts the constraint to electricity — liquid-cooled racks run hotter and denser, demanding more power infrastructure per square metre Industry engineering literature.
Carbon and the Nuclear Pivot
The climate arithmetic is uncomfortable. Microsoft's emissions rose roughly 30% above its 2020 baseline by 2023, effectively abandoning its 2030 carbon-negative target in practice; Google's emissions rose a similar magnitude over the same period Company sustainability reports Substantiated. Both had committed to carbon-free or carbon-negative operations by 2030. AI's energy demand made those pledges arithmetically impossible without new supply.
The industry's answer has been nuclear:
- Microsoft signed a 20-year PPA to restart Three Mile Island Unit 1 (835 MW) by 2028 Constellation Energy / Microsoft, 2024
- Amazon bought a nuclear-powered data centre campus (Talawanta Ridge, Pennsylvania) and has invested in small modular reactors Amazon / press, 2024–25
- Google signed with Kairos Power for up to 500 MW of SMRs through 2035 Google / Kairos, 2024
- Meta issued an RFP for 1–4 GW of nuclear capacity Meta, 2024
Nuclear plants take a decade or more to license and build; SMRs have never been commercially deployed at scale anywhere. Data centres take 18–24 months. The mismatch means the 2026–2030 gap will be filled by gas — which is precisely why several states have seen data-centre-driven gas plant proposals and rising wholesale power prices. EIA / grid operator filings
Confidence Assessment
| Claim | Status | Confidence |
|---|---|---|
| Data centre electricity grew 17% in 2025; AI facilities ~50% | IEA, April 2026 | Substantiated |
| Global consumption doubles to ~945 TWh by 2030 | IEA projection; depends on buildout pace | Partly substantiated |
| US data centres reach 400–600 TWh by 2030 | Wide range across EPA/DOE/GS/EPRI forecasts | Partly substantiated |
| Stargate will draw 1GW+ (750K homes equivalent) | Company announcements; not yet independently metered | Company claim |
| Grid queues, not chips, are the binding constraint through 2028 | Utility filings; contested by some operators | Partly substantiated |
| AI will force a rollback of corporate climate pledges | Already visible in 2023–24 sustainability reports | Substantiated |