Infrastructure, Resources, and Constraints

AI infrastructure is the physical base of AI computation: semiconductor design, fabrication and packaging; the manufacturing equipment that makes chips; data centres that house accelerators; and the electricity, cooling and grid connections that power them. This article describes the structure of that base, reports dated energy estimates from primary institutional sources, and states where estimates conflict.

Scaling model training and inference (ECON-COMP-001) converts into demand for chips, floor space and electricity. The binding constraint may differ by region and period: chip supply, grid interconnection, equipment lead times or capital. Because these inputs are geographically concentrated, they are also instruments of national policy (GOV-EXPORT-001). Energy figures below are scenario-based estimates, not measurements, and the IEA itself states there is "substantial uncertainty" about current and future data-centre consumption.

Data-centre electricity use grew modestly in the 2010s as efficiency gains offset rising workloads; growth accelerated with AI-oriented servers. The Lawrence Berkeley National Laboratory (LBNL) 2024 US Data Center Energy Usage Report, published 19 December 2024 by Shehabi et al., reports that US data-centre demand more than doubled between 2017 and 2023, largely due to AI servers, and gives scenario ranges to 2028 (Shehabi et al., 2024). The headline numbers (176 TWh, 4.4% of US electricity in 2023; 325-580 TWh, 6.7-12% in 2028) come from LBNL's news summary as returned by a search tool and were not read in the report itself [VERIFICATION NEEDED].

The International Energy Agency's Energy and AI report (IEA, 2025) provides the most-cited global estimates.

#4.1 Energy

Energy used by a data centre is the product of IT load, a utilisation factor and a power-usage-effectiveness overhead for cooling and power conversion. For AI, accelerated servers (GPU-type) draw more power per rack than conventional servers.

IEA estimates (IEA, 2025; scenario-based):

• Global data-centre electricity consumption was about 415 TWh in 2024, roughly 1.5% of global electricity, having grown about 12% per year over the previous five years. By region, the United States accounted for 45%, China 25% and Europe 15%.
• In the Base Case it reaches about 945 TWh by 2030 (slightly below 3% of global electricity), about 15% per year growth from 2024, with accelerated servers growing about 30% per year and conventional servers about 9% per year; and around 1,200 TWh by 2035.
• Growth by 2030 is about +240 TWh in the US, +175 TWh in China and more than +45 TWh in Europe; the US and China together account for nearly 80% of global growth.
• 2035 sensitivity cases: Lift-Off over 1,700 TWh, High Efficiency about 970 TWh, Headwinds near 700 TWh.
• Globally the share is small; locally it can be large because AI data centres cluster. About 50% of US data centres under development are in existing large clusters; around 20% of planned projects could be delayed absent grid action; and in advanced economies data centres could account for more than 20% of electricity-demand growth to 2030 (IEA, 2025).

These are the IEA's own scenario outputs. Other analyses produce different numbers for AI-specific consumption (one secondary-reported Nature estimate of 239-295 TWh for AI data centres by 2030 is a different scope and was not verified here) [VERIFICATION NEEDED]. The claim that AI energy use is large or small is therefore scope-dependent: small as a global share, material in particular grids.

#4.2 Grid and equipment constraints

The IEA states that wait times for transformers and cables have doubled in the past three years, that gas-turbine deliveries have lead times of several years (potentially pushing new gas capacity beyond 2030), and that building new transmission lines can take four to eight years in advanced economies (IEA, 2025).

#4.3 Semiconductors

Producing AI accelerators involves design, wafer fabrication at leading-edge nodes, advanced packaging and high-bandwidth memory (TECH-HW-001). It is widely reported that leading-edge logic fabrication is concentrated in Taiwan and that extreme-ultraviolet (EUV) lithography tools are supplied by a single company, ASML. Reported shares differ with the node cutoff and year (secondary sources cite figures around 90% for the most advanced logic, and a CSIS 2023 estimate above 90%), and I did not verify them against TSMC or ASML filings, so no percentage is adopted here [VERIFICATION NEEDED: primary source for fabrication concentration and ASML's sole-supplier status].

#4.4 Critical minerals

The IEA states that China supplies around 99% of refined gallium, important for advanced chips and power electronics, and estimates that data-centre demand for gallium could exceed 10% of today's global supply by 2030 (IEA, 2025). China's separate export controls on rare earths and related items are discussed in GOV-EXPORT-001.

Examples in the sources: IEA's statement that a typical AI-focused data centre uses as much electricity as about 100,000 households; clustering of US projects; and long lead times for grid equipment. The dollar value of announced data-centre investment is a fast-changing figure and is omitted (see FACT-001).

• Uncertainty: IEA and LBNL both present scenarios; Base Case numbers should not be read as forecasts.
• Efficiency: hardware and algorithmic efficiency gains could reduce energy per unit of capability (the IEA's High Efficiency case reduces 2035 demand to about 970 TWh against over 1,700 TWh in Lift-Off), but rebound effects are possible and unquantified here.
• Emissions: the IEA reports data centres account for around 1.5% of energy-sector emissions in the period it examines, while being among the fastest-growing sources.
• Unresolved: how much planned capacity will be built; who bears grid-upgrade costs; the share of AI versus other digital services.

As of 2026-10-15 the IEA 2025 scenarios remain the principal global reference reviewed here; later updates were not retrieved [VERIFICATION NEEDED: check for any 2026 IEA update]. It is Well Supported that data-centre electricity demand is rising quickly and is geographically concentrated; the size of future demand is Contested and scenario-dependent.

The IEA's 2035 range (about 700 TWh to over 1,700 TWh) is a scenario envelope, not a registered forecast. No separate forecast is registered here; see FCST-REG-001.

• Foundational Concepts:
  • TECH-HW-001 AI Hardware
  • ECON-COMP-001 The Economics of Compute and Scaling
• Related:
  • ECON-IND-001 The AI Industry Structure
  • GOV-EXPORT-001 Export Controls and AI Geopolitics
  • ECON-MEAS-001 Measuring the AI Economy