Topicspowering-ai
A sourced reference collection on the electricity AI actually consumes — what the measured figures are, how far the published estimates disagree, and why grid connection rather than capital has become the binding constraint on building.
8 sources sit behind this page — including any that arrive with a concept this page shares with another collection. 6 were retrieved and read in full, and only those can back an assertion. 2 could not be retrieved. Every one of them is named in the register below, with the reason in view. How we source this.
The vocabulary this subject is built from, and what we can show about each.
AI racks draw 30 to over 100 kW against 7-10 kW for traditional servers, and synchronised GPU clusters swing load by hundreds of megawatts within seconds — a ramp-rate problem distinct from, and not addressed by, average consumption figures.
Data centres were just over 1% of global electricity demand and 0.5% of CO2 emissions in 2024 (US nearly half, China ~25%, Europe ~15%); to 2030 they add a projected 530 TWh of demand against 838 TWh for electric vehicles and 651 TWh for air conditioning.
Estimates of China's data centre electricity use disagree by more than a factor of two: 200 TWh for 2020 (State Council), 77 TWh for 2022 (Development Research Centre), roughly 100 TWh for 2024 at a share stated between 0.9% and 2.7%, and 2030 projections from 400 TWh to about 600 TWh.
US data centres used 58 TWh in 2014 and 176 TWh in 2023 (4.4% of national electricity); Berkeley Lab projects 325-580 TWh by 2028, or 6.7-12% — a range spanning nearly a factor of two, which is the laboratory's own stated uncertainty, not a conflict between sources.
Fossil fuels supply nearly 60% of electricity reaching data centres globally, against 27% renewables and 15% nuclear; in China, where coal is 60.5% of the power mix, eastern data centres source roughly 70% of their electricity from coal.
Grid connection, not capital or chips, is now the binding constraint: transmission and distribution upgrades take five to ten years, and PJM capacity prices rose from $28.92 per MW-day for 2024-2025 to $329.17 for 2026-2027, more than a tenfold increase.
Cost allocation is unresolved: Dominion Energy proposed its first base-rate increase since 1992 — about $8.51 per month in 2026 and $2.00 more in 2027 for a typical household — while special tariffs risk shifting costs to other consumers and utilities face stranded-asset risk if forecast demand does not appear.
Power usage effectiveness measures total data centre energy against energy delivered to computing equipment; China uses it as policy, bringing the average from 1.54 to 1.48 under a national action plan with a 1.25 target for large data centres by 2025.
What actually happened, in order, with sources.
Where this topic’s events took place, as far as our sources establish it. Events with no single location — a standards publication, say — and events we have not yet attributed are both counted as unattributed rather than omitted.
Dec 12, 2025
On 12 December 2025 Ireland's Commission for Regulation of Utilities required new data centre grid connections to meet at least 80% of annual demand with additional Irish renewable generation — explicitly excluding capacity already supported under REFIT, RESS or ORESS — and to provide matching onsite or nearby generation or storage participating in the wholesale market.
Ireland's regulator states data centres rose from 5% of national electricity demand in 2015 to 22% in 2024, with 9.4 TWh projected for 2025 rising to 14.6 TWh and 31% of national demand by 2034.
Jun 2025
Texas enacted Senate Bill 6 in June 2025 covering cost sharing, interconnection transparency and emergency operations for large loads, including a $100,000 minimum transmission screening study fee; ERCOT projects peak summer demand of 145 GW by 2031 against 85 GW in 2024, with about 32 GW of it from data centres.
Jul 2024
Around sixty Northern Virginia data centres disconnected from the grid simultaneously in July 2024, creating a roughly 1,500 MW power surplus that required emergency adjustments to prevent cascading outages.
2022
China launched the east data west computing programme in 2022, directing non-time-sensitive processing of eastern data to western provinces where large solar and wind farms are sited, while time-sensitive services remain in the east.
All 8 sources behind this page — what we read, what we tried to read and could not, and what we looked at and set aside, with the reason in view for each. A concept shared with another collection brings its own references with it, so some entries here were surfaced for a neighbouring topic rather than this one.
Original publisher links. Files open on the publisher’s site; we do not host copies. A linked document is not an additional source or an independent verification.
Belfer Center for Science and International Affairs, Harvard Kennedy School · Published 2026-02-10
Carbon Brief · Published 2025-09-15
Lawrence Berkeley National Laboratory · Published 2025-01-15
arXiv (Chen, Wang, Colacelli, Lee, Xie) · Published 2025-09-08
Carbon Brief · Published 2025-04-16
Commission for Regulation of Utilities (Ireland) · Published 2025-12-12
We attempted these and were refused or served nothing. Nothing on this page rests on them; they are published so the gaps are checkable rather than invisible.
What this page does and does not claim.
Fourth packet through the generic ingestion pipeline (August 2026). Six sources were directly retrieved and two are recorded as references only — both of them IEA pages, which returned HTTP 403. That is the packet's most consequential gap: the IEA's Energy and AI report is the reference most other sources in this field cite, and its figures appear here only at one remove through Carbon Brief, attributed that way rather than presented as read at source. Geographically this collection covers the United States (a Department of Energy national laboratory and Harvard's Belfer Center), Ireland (the energy regulator's own binding decision — the clearest instance anywhere of a state acting on this), China (via Carbon Brief), and global proportion. The China figures are deliberately left unreconciled: published estimates differ by more than a factor of two and no source retrieved this session resolves them, so the spread is asserted as the finding. Two of the six sources are non-peer-reviewed or synthesising preprints and are marked as such; none is ever the primary source for an assertion. Not yet editor-reviewed; every assertion reads as reported.
Source check, 2026-09-17. Numeric-presence checks passed for 13 assertions using available source text, which may be cached. This is not verification of their meaning. What this check does and does not prove →
This page is also available as structured data: /api/v1/topics/powering-ai