AI data centres: global power demand

What it is

AI data centres are large-scale facilities housing server hardware, networking infrastructure, and cooling systems, built and operated by hyperscalers (Amazon, Microsoft, Google), specialist cloud providers, and enterprise operators. The energy intensity of AI workloads distinguishes them from conventional data centres: a high-density AI rack running Nvidia H100 or Blackwell GPUs continuously can draw 100 kW or more, compared with roughly 10-15 kW for a typical legacy server rack. As of 2024, data centres consumed an estimated 415 TWh of electricity globally, about 1.5% of total world demand, per the International Energy Agency's January 2025 report "Energy and AI." The United States alone accounted for 176 TWh in 2023, or 4.4% of total US electricity consumption.

History

Commercial data centres date from the 1990s, initially consolidating enterprise mainframe computing. US consumption held at roughly 58 TWh in 2014 despite rising server counts, because virtualisation and efficiency improvements absorbed demand. The inflection came with the deep-learning wave: from 2017 to 2023, US data centre power more than doubled, driven by AI training and inference workloads on GPU hardware. The public release of large language models in late 2022 triggered a hyperscaler capital expenditure surge. Five major technology companies spent more than US$400 billion in capex in 2025 and planned a further 75% increase for 2026, per IEA data.

Current state

Lawrence Berkeley National Laboratory's December 2024 report projects US data centre electricity consumption will reach 325-580 TWh by 2028, representing 6.7-12% of total US electricity. The IEA's central scenario projects global data centre demand reaching roughly 945 TWh by 2030, with AI-focused facilities tripling from 2025 levels. The United States holds 54% of global hyperscale capacity by megawatt. Amazon, Microsoft, and Google together hold 59% of the hyperscale fleet, per Synergy Research Group data from March 2025.

The scale of the buildout has created grid-planning emergencies on two major US regional grids. PJM, which serves 67 million people across 13 US states and Washington DC, attributed 63% of a near-doubling in wholesale power prices in Q1 2026 to data-centre load (see Data centers drive 63% of PJM's power-price surge as backlash hardens). Texas's ERCOT forecast a record summer peak above 92 GW for 2026, partly driven by data-centre interconnection requests that had almost quadrupled in a year (see ERCOT forecasts record 92GW summer peak, says Texas grid will hold).

Water is a secondary constraint gaining legal weight. Google consumed 6.4 billion gallons at its global data centres in 2023, up 14% from 2022. Suits over water access, including a claim for roughly 260 million gallons per year of Colorado River water in California's Imperial Valley (see Imperial Valley data-center developer sues for Colorado River water), mark an emerging regulatory front.

Relationships

The power-demand story connects directly to the US transmission deficit: new HVDC corridors (see US transmission inches forward in 2026 as HVDC corridors advance and the gap stays open) are lagging both data-centre load growth and queued renewable generation. PJM's June 2026 emergency curtailment order (see DOE clears PJM to curtail data centers as the largest US grid edges toward shortfall) was a direct consequence of that supply-side gap. On the generation side, data-centre developers are now the primary contracted buyers of small modular reactors, including Oklo's 12 GW agreement with Switch and the Constellation-Microsoft Three Mile Island Unit 1 restart (see SMR order book swells on AI power demand: Oklo-Switch 12GW, Rolls-Royce wins Sweden). More than 300 US state bills filed by mid-2026 aim to shift grid-upgrade costs from general ratepayers onto large-load customers.

What to watch

• Whether the IEA's 2030 forecast of 945 TWh proves conservative given 2025-2026 hyperscaler capex; a faster ramp would intensify grid and water crises before mitigation measures can arrive.
• State-level large-load tariffs and interconnection-filter regimes: ERCOT's Batch Zero process and PJM's 50 MW tariff threshold set precedents other US and international grids will follow or resist.
• Water-rights litigation in drought-constrained US river basins, particularly the Colorado, as liquid-cooled AI clusters apply for allocations alongside agriculture.
• Whether SMR and repowered-nuclear contracts convert into financed, permitted construction before the demand wave peaks in the late 2020s.
• Policy responses from the EU and India, both large build-out markets with distinct grid and water stress profiles from the United States.