Data centres are the fastest-growing electricity load of the decade, and AI is the main driver. US data-centre demand is on track to more than triple, from about 23 GW in 2023 toward 75-80 GW by 2026. The IEA projects global data-centre electricity use — all data centres, not only AI — to more than double by 2030 to around 945 TWh, close to the entire electricity consumption of Japan today. AI is the fastest-growing slice of that: a single AI-optimised rack now draws 50-100 kW against 5-10 kW for a traditional rack, which is why power — not chips — has become the binding constraint on AI build-out.
| US data-centre demand 2023 | ~23 GW |
| US data-centre demand 2026 | ~75-80 GW (>3× 2023) |
| Global DC electricity by 2030 — all data centres (IEA) | ~945 TWh |
| AI rack power | 50-100 kW |
| Traditional rack | 5-10 kW |
| Share of US power-demand growth to 2030 from DCs | ~50% |
Sources: IEA Electricity 2024 / Energy & AI (global TWh); EIA & industry analyses (US GW). Figures are projections and revise frequently.
The metric the industry now watches is "tokens per watt per dollar" — squeezing more AI output from each unit of power, because power availability caps growth. In the US, data centres are expected to drive roughly half of all electricity-demand growth to 2030, met near-term mostly by natural gas — which runs on an ageing turbine fleet with 5-7 year lead times for new units.
| Unit | Electricity |
|---|---|
| Traditional server rack | 5-10 kW |
| AI / GPU rack | 50-100 kW |
| Global data-centre electricity, 2024 | ~415 TWh |
| Global data-centre electricity, 2030 (IEA proj.) | ~945 TWh ≈ Japan today |
| Share of US power-demand growth to 2030 from data centres | ~50% |
Per-query energy figures are widely contested (they vary 10×+ by model, hardware and method), so we omit a single headline number; rack power density and aggregate TWh are the robust, citable comparisons. Sources: IEA, US EIA, industry analyses.
Every gas turbine, HRSG and steam line in this fleet — and every kW of data-centre waste heat — is a hot surface. Inzonex modular removable insulation cuts surface heat loss up to 96% while keeping flanges accessible for maintenance.
See Inzonex modular insulation →The IEA projects global data-centre electricity use will more than double by 2030 to ~945 TWh, similar to Japan's total consumption today. US demand alone is heading from ~23 GW (2023) toward 75-80 GW by 2026.
AI training and inference run on dense GPU racks drawing 50-100 kW each — roughly 10x a traditional 5-10 kW rack — so the same floor space needs far more power and cooling.
Increasingly power: interconnection queues, gas-turbine lead times and grid capacity now gate data-centre build-out more than chip supply.
IEA (global data-centre electricity, ~945 TWh by 2030); US EIA and industry analyses (US GW, rack power density). All figures externally sourced and attributed; projections revise frequently. Author: Dmytro Aheiev (ORCID 0009-0001-5512-0291).
Externally-sourced figures are attributed and link to primary sources; deal terms are public company/trade-press announcements. Author: Dmytro Aheiev (ORCID 0009-0001-5512-0291).