The IEA projection cited across this literature has global data-centre electricity demand reaching nearly 945 TWh by 2030, more than double the 415 TWh used in 2024, with AI workloads the principal driver of that growth.
IEA, cited in Chen et al., 2025
CorroboratedData centres drew 415 TWh in 2024, about 1.5% of world electricity, and demand is growing roughly 12–15% a year: several times faster than electricity demand overall.
IEA, cited in Chen et al., 2025
Forecasts converge on +128% by 2030, from 415 to 945 TWh. Of the 210 GW of generation planned to replace retiring plants and meet new load, DOE finds only 22 GW will be firm, dispatchable capacity.
IEA, cited in Chen et al., 2025; US DOE, Resource Adequacy fact sheet, 2025
The median interconnection wait is now 5 years, up from under two a decade ago, with over 2,300 GW of generation and storage stalled in US queues.
Mehboob et al., 2026
Efficiency, clean-power contracts, privately paid transmission and flexible load: the gains the research can put a number on.
Google's ML-optimised TPUv4 cut training energy 13.7-fold against the unoptimised 2017 GPU it replaced. Hardware is one of four factors the authors say combine to cut ML carbon footprints up to 747-fold.
Patterson et al. (Google / UC Berkeley), 2022
Single-sourceMicrosoft (by 2024) and Amazon (by January 2025) each report more than 34 gigawatts of power-purchase agreements for clean generation, larger than the contracted renewable portfolios of most national utilities.
Huang, Deb & Zareipour, 2026, Table I (company sustainability reports)
Company-reportedSB Energy and AEP Ohio partnered to build $4.2 billion of new transmission for a Portsmouth AI campus, with SB Energy, not ratepayers, committed to paying for it.
US DOE, Portsmouth fact sheet, 2026
CorroboratedSwitching from air to direct-to-chip or immersion liquid cooling can cut cooling-related electricity by roughly 40 to 50%, with water use falling 20 to 90% depending on climate and design.
Patel, Mahalingam & Patel, 2026
Single-sourceIn the first live demonstration of its kind, Emerald AI cut an Oracle Cloud AI cluster's power draw 25% for three hours during grid peaks in Phoenix, using workload orchestration alone: no batteries, no hardware changes.
Colangelo et al. (Emerald AI / EPRI DCFlex), 2025
Company-reportedGoogle's agreement with Kairos Power will deploy small modular reactors supplying up to 500 MW of new clean power to its AI data centres: new generation, not relabelled existing supply.
Huang, Deb & Zareipour, 2026
Single-sourceLocation and grid mix shape outcomes as strongly as hardware efficiency: placing AI data centres in favourable regions rather than unfavourable ones can cut their carbon emissions by up to 49%.
Patel, Mahalingam & Patel, 2026
Single-sourceThe same choice cuts the water footprint by about 52%. Advanced cooling plus low-carbon, water-secure siting can nearly halve the combined footprint.
Patel, Mahalingam & Patel, 2026
Single-sourceNo verified figure exists for the net emissions displaced by workloads that moved from on-premise hardware into these facilities. Vendor claims range from substantial to negligible and share no common methodology; the figure was excluded rather than approximated.
No verified figure exists for how much contracted clean capacity has actually energised. Contract announcements are public; commissioning dates largely are not.
Modelled, newly measured or still moving: findings that point a direction without closing the question.
Power usage effectiveness of 1.09 (Google) and 1.08 (Meta) means 8–9% extra energy for cooling and power delivery; at the ~1.6 industry average the overhead is 60%. Fleet averages can hide inefficient sites.
Huang, Deb & Zareipour, 2026, Table I; EPRI, Powering Intelligence, 2024
Company-reportedA CAISO-scale grid model holds 770 MW of frequency-regulation reserve at all times, 5.5% of modelled capacity; the authors propose flexible GPU data centres as a new source for it.
Jahanshahi et al., 2026, ACM
Peer-reviewedEPRI's four growth scenarios take data centres from about 4% of US electricity in 2023 to between 4.6% and 9.1% by 2030.
EPRI, Powering Intelligence, 2024
Single-sourceMeasured against rated power, inference servers leave about 21% headroom; training leaves about 3%, because synchronous power peaks run close to the limit. Inference is the candidate for packing more servers onto the same supply.
Measurement study cited in Vercellino et al., 2026
Single-sourceUS data centres used 176 TWh in 2023, 4.4% of the country's electricity, up from 1.9% six years earlier.
Chen et al., 2025; Golden et al., 2026
CorroboratedNo verified figure exists for inference energy per query at production scale. Every published number is either a vendor estimate or an extrapolation from a single benchmark run; none were reproducible.
No verified figure exists for the fraction of accelerator idle time that could be eliminated by scheduling alone. The telemetry that would answer it is not published by any operator.
Emissions, water, capacity prices, permits and land: the costs, measured at the same weight as the gains.
US AI servers are projected to add 24 to 44 million metric tons of CO₂-equivalent emissions a year by 2030, depending on the carbon intensity of the grid supplying them.
Patel, Mahalingam & Patel, 2026
Single-sourceUS AI servers are projected to add 200 to 300 billion gallons of water consumption a year by 2030, driven mainly by evaporative cooling.
Patel, Mahalingam & Patel, 2026
Single-sourcePJM's capacity price, the charge utilities pay to guarantee supply and pass through to customer bills, jumped from $30 to $270 per megawatt-day between 2023 and 2024 amid data-centre load growth.
Luby et al. 2025, cited in Golden et al., 2026
CorroboratedSatellite monitoring found the gas turbines powering xAI's Colossus 2 data centre emitting NOₓ at roughly 16 times its permitted rate: 730 kg an hour measured against a permitted 47.
Gauld et al. (MIT / Carbon Mapper), 2026
Single-sourceThe proposed Digital Gateway campus in Prince William County, Virginia would rezone roughly 2,100 acres (850 hectares) of rural land, one of several AI-driven greenfield conversions alongside sites in Indiana and Nebraska.
Bolaños-Zuñiga & Lamadrid, 2026
Single-sourceGlobal data-centre server manufacturing reached an estimated 12 million units in 2021; the same authors argue that the lifecycle cost of building this hardware, not operational energy, is the bigger long-term climate risk.
Patterson et al. (Google / UC Berkeley), 2022
Single-sourceDOE counts 104 GW of announced US plant closures by 2030, to be met by 210 GW of new generation of which only 22 GW is firm, dispatchable capacity; it has issued emergency orders keeping at least one coal plant running past its retirement date.
US DOE, Resource Adequacy fact sheet, 2025
CorroboratedWhen new AI load is met by marginal coal generation, the authors model roughly 800 g CO₂/kWh, against about 100 g/kWh in a high-renewable system.
Huang, Deb & Zareipour, 2026
Single-sourceMore than 2,300 GW of generation and storage is stalled in US interconnection queues; the median project now waits about five years, up from under two a decade ago.
Mehboob et al., 2026
CorroboratedThe mirror of the siting gain: an unfavourable region can raise an AI data centre's carbon emissions by up to 90%.
Patel, Mahalingam & Patel, 2026
Single-sourceWater footprints can rise by more than 350% in an unfavourable region, the widest swing of any factor the study measured.
Patel, Mahalingam & Patel, 2026
Single-sourceWide ranges, concentrated risk and thin disclosure: the places where the record is late, partial or inconsistent.
From 176 TWh in 2023, US data-centre use is projected to reach anywhere between 325 and 580 TWh by 2028. The width of the range is the finding: it depends on growth scenarios nobody can yet pin down.
Chen et al., 2025
Single-sourceFifteen states hosted 80% of US data-centre load in 2023. EPRI flags two of them, Arizona and Nevada, for water scarcity as a top challenge, yet siting keeps following cheap land and power there.
EPRI, Powering Intelligence, 2024
CorroboratedData centres can be built in 18 months, but bringing new generation online to power them takes more than three times as long: a structural lag between the AI build-out and the grid.
US DOE, Resource Adequacy fact sheet, 2025
CorroboratedIn 2024 a single Virginia data centre dropped 1.5 GW of load from the grid within 82 seconds of a transmission fault. Hyperscalers rarely disclose facility operating data, so regulators often learn of such swings after the fact.
NERC, cited in Vercellino et al., 2026
CorroboratedA review of 258 published estimates of data-centre energy use found methods so inconsistent that results span a wide range: there is no agreed methodology behind even this page's headline numbers.
Mytton review, cited in Vercellino et al., 2026
Single-sourceBoth sides above are real. The gains are measured in performance per watt; the costs in ratepayer bills, withdrawn water and permitted gas. Nothing here decides which matters more: that judgement sits with whoever reads the tariff filing. Every figure carries its source and its confidence; where the record was silent, this page says so rather than estimating.
Four figures are cited second-hand, as their source states: the IEA projection via Chen et al., Luby et al. via Golden et al., and NERC and the Mytton review via Vercellino et al.
