★
Menu
Accueil Comment c'est construit Watchlist La pile
Matériel PucesComparer les puces Serveurs & systèmesRéseau MémoireFonderies & Packaging Refroidissement
Infrastructures Data centersAlimentation ProjetsCloud et calcul GPU
Intelligence Modèles IALogiciels Produits IABenchmarks Articles de rechercheJeux de données Concepts
Écosystème EntreprisesActions PersonnesChronologie TendancesAvenir de l'IA
Formation Actualités Interroger les données Agents IA
À propos À propos de nous Méthodologie Données & API Avertissement
Options
DONNÉES & API

Accès JSON gratuit en lecture seule aux données en cache du site.

Mode sombre

Vue guidée
Nouveau dans l'univers des puces, des data centers et du jargon des modèles ? Nous expliquons les termes clés au fil de votre navigation, en langage clair. Les mêmes données, avec l'aide intégrée.

Vue expert
Vous connaissez déjà le sujet. Juste les données : claires, rapides et concises, sans explications superflues. Il s'agit de la vue par défaut.

Langue de l'interface
‹ Retour
Exploration approfondie 8 min de lecture · ← Tous les articles explicatifs

Every abstraction in AI bottoms out in electrons. The industry's constraint has shifted — chips in 2023, now increasingly the watts behind them. 'Speed to power' replaced 'GPU allocation' as the sector's scarcest commodity.

The honest numbers

Global data centers consumed roughly 415 TWh in 2024 — about 1.5% of world electricity (IEA estimate), with the IEA's base case near 945 TWh by 2030, AI the dominant driver. In the US, data centers were ~4% of power demand in 2024 with credible projections toward 7–12% by 2030 — the first real load growth American utilities have planned for in twenty years. A single frontier campus now asks for what a mid-size city draws: Hyperion's announced 5 GW is five large reactors' worth. (All projections carry wide bands — see Trends for sources.)

Where it actually binds

Generation fuel is not the bottleneck; connection is. Interconnection queues in ERCOT and PJM hold hundreds of gigawatts of requested data-center load, studies take years, and the equipment is worse: large transformers quote 3–5-year lead times, and heavy gas turbines are sold out into the 2029+ window at GE Vernova, Siemens Energy and Mitsubishi. Hence the era's signature move: bring your own power.

The build-around strategies

Behind-the-meter gas: turbines at the fence line (Colossus, Homer City's 4.5 GW plan). Nuclear money: restarts and PPAs — see the nuclear revival. Renewables + storage: the fastest megawatts in Texas, with hyperscalers the biggest corporate buyers. Fuel cells (Bloom) where combustion permits lag. And load flexibility — pausing training during grid peaks — which studies suggest could unlock existing headroom without new iron.

Efficiency won't save you (alone)

Chips improve perf-per-watt ~25–40% per generation and PUE approaches 1.1 — yet totals keep climbing, because demand elasticity eats the gains (Jevons). The IEA's scenarios differ by hundreds of TWh on exactly this question. Grid operators now plan for the high case; so should readers of this database — while remembering it is a scenario, not a fact.

The politics arriving with the load

Who pays for grid upgrades (the AEP Ohio tariff fights), local air permits (Memphis), water optics, and rising retail rates in data-center regions — the social license questions are now part of the engineering. Records here flag them site by site.

8 Prochaine étape du parcours Inside an AI Data Center An AI data center is a giant building full of computer shelves, thick cables, plumbing for coolant, and enough electricity for a whole city.

Plus dans Alimentation

The Nuclear Revival Lire → Interconnection: The Longest Queue in Tech Lire →