Global AI Compute Footprint
Operational and announced AI compute worldwide, with the US footprint read against state tier classifications and ISO/RTO boundaries.
Abstract
As of 2026-04-23, published AI compute sites with coordinates span 346 facilities across 37 countries: 29 compute-thresholded frontier training campuses and 317 broader GPU clusters (hyperscaler regions, sovereign builds, academic installations, regional sites). Frontier capacity remains heavily concentrated: the compute-thresholded list appears in 3 countries when the US is included, while the broader cluster base extends across the rest of the map.
The US carries 129 of those sites, ~57% of tracked nameplate MW, and 5 of the 11 gigawatt-class campuses globally. The US footprint splits across all four state tiers, with roughly 76% of US MW landing in T1 or T2 (favorable or workable) states and 24% in T3 or Avoid. By ISO/RTO, announced capacity concentrates in PJM, MISO, and ERCOT — three of the regions with the largest published 2030 generation deficits.
The public map understates China. MIIT figures put Chinese intelligent computing capacity at ~1,590 EFLOPS at end-2025, organised under eight national computing hubs with cumulative program investment above $33 billion. The US–China frontier-compute gap has narrowed from roughly 10× in 2018 to about 3× in 2025 on public FLOPS estimates, with the Chinese buildout bounded mostly by Ascend production and HBM supply rather than by site availability.
The global footprint at a glance
The 29 frontier training campuses are the compute-thresholded subset running at the leading edge of cluster scale; the 317 broader GPU clusters span hyperscaler regions, sovereign builds, academic installations, and regional sites. 11 sites on the map carry announced nameplate at or above 1 GW. Sources are cited at the end of the brief.
US concentration by state tier
T1 states combine a fast interconnection queue with a clear path to behind-the-meter or co-located generation. Avoid states combine long queues with high blended rates. The four tiers sit on a 7-dimension composite of queue efficiency, permitting speed, BTM pathway, transmission headroom, resource access, saturation, and cost.
US sites by state tier
State tiers from the Power Investor state-score framework (7-dimension composite).
| Tier | Sites | MW | % of US MW | GW-class |
|---|---|---|---|---|
T1 — Favorable Fast queue, low blended rate, clear BTM path | 31 | 5.7 GW | 28% | 2 |
T2 — Mixed Moderate queue and cost; workable with patience | 33 | 10.1 GW | 49% | 2 |
T3 — Constrained Slower queue or higher cost; narrow project set | 29 | 3.5 GW | 17% | 1 |
Avoid High cost + long queue; structural disadvantage | 36 | 1.5 GW | 7% | 0 |
| US total (resolved) | 129 | 20.8 GW | 100% | 5 |
Two things stand out. First, the gigawatt-class subset is not evenly distributed across tiers — 4 of the 5 US gigawatt-class sites sit in T1 or T2 states. Announced megaproject capacity is gravitating to the states where interconnection and permitting risk is lowest. Second, a meaningful tail of announced nameplate still sits in T3 and Avoid states — capacity on the map that will have to navigate queue timelines of six years or more and blended rates above 12¢/kWh before it energizes.
By nameplate MW, the top US states absorbing announced capacity are TX (T1), PA (T2), WI (T3), TN (T2), GA (T2).
US concentration by ISO, read against the 2030 deficit
Each US site is bucketed by ISO/RTO. The 2030 deficit figure is the published gap between projected regional demand and available firm supply in that year.
US sites by ISO / RTO
2030 deficit estimates are from the Power Investor supply-demand research where published.
| ISO / RTO | Sites | MW | % of US MW | GW-class |
|---|---|---|---|---|
ERCOT 2030 deficit est. 9 GW | 25 | 5.2 GW | 25% | 2 |
SPP 2030 deficit est. 3 GW | 12 | 1.7 GW | 8% | 0 |
MISO 2030 deficit est. 5 GW | 22 | 5.3 GW | 26% | 2 |
PJM 2030 deficit est. 14 GW | 26 | 6.8 GW | 33% | 1 |
SERC 2030 deficit est. 6 GW | 8 | 1.2 GW | 6% | 0 |
WECC (non-ISO West) | 9 | 323 MW | 2% | 0 |
CAISO | 13 | 74 MW | 0% | 0 |
NYISO | 1 | 0 MW | 0% | 0 |
ISO-NE | 0 | 0 MW | 0% | 0 |
Unassigned / cross-ISO | 13 | 129 MW | 1% | 0 |
PJM carries 26 sites and 33% of announced US MW on the map, against a published 2030 deficit of 14 GW. ERCOT carries 25 sites and 25% against a 9 GW deficit. SERC and MISO together carry 30 sites. SPP — the ISO with the fastest queue and the lowest published deficit — carries 12 sites, a smaller share of announced MW than the site economics would suggest. The asymmetry between announced concentration and published deficit is the structural piece: the ISOs already short of supply through 2030 are the ISOs absorbing the largest share of announced capacity, which is the path that compounds, not resolves, the deficit.
Outside the US: sovereign and regional patterns
The non-US map carries 217 sites across 36 countries. Frontier-scale capacity outside North America is narrow in the public record: 2 countries carry at least one compute-thresholded frontier site with published coordinates. The broader cluster base spreads much wider and breaks into recognisable patterns — state-backed campuses in the Gulf, sovereign-aligned builds in India, Korea and Japan, academic and regional clusters across Europe, and emerging deployments in Brazil and Southeast Asia.
Public site-level data on Chinese AI compute is limited, so the actual capacity of Chinese frontier compute is considerably higher than what is listed below. The aggregate picture from MIIT disclosures and outside analysis is internally consistent: MIIT reports 725 EFLOPS of intelligent computing capacity at end-2024 and ~1,590 EFLOPS at end-2025, growing above 40% annually, with 215.5 EFLOPS of that sitting inside the eight national hubs under the “East Data West Compute” (东数西算) program — Beijing-Tianjin-Hebei, Yangtze Delta, Greater Bay Area, Chengdu-Chongqing, Inner Mongolia, Guizhou, Gansu, and Ningxia. Cumulative investment in the program reached ~¥239B ($33B) by mid-2024.
Facility-level evidence is thinner but consistent with a buildout well above the public catalogue. Huawei’s Gui’an campus in Guizhou is rated at 200 MW across 101.4 hectares; Alibaba’s Zhangbei complex in Hebei is estimated by Epoch AI satellite analysis at 200–500 MW against 12 EFLOPS of advertised AI capacity; ByteDance is building a ~150 MW campus in Datong (Shanxi) and has separately reserved a ~500 MW footprint with VNet. The operational-vs-announced gap is material though: CSIS and SemiAnalysis estimate Huawei shipped roughly 200k Ascend 910B units in 2024 against a 400k 910B / 910C target for 2025, bounded by HBM supply and SMIC 7nm wafer throughput. Epoch AI places the US–China frontier-compute gap at roughly 3× in 2025, down from ~10× in 2018. The 1,590 EFLOPS headline is best read as nameplate mixed-precision capacity across heterogeneous silicon, not as a one-for-one comparison to US hyperscaler H100 / B200 fleets. A future brief will take the Chinese pipeline in more depth.
| Country | Sites | MW | GW-class |
|---|---|---|---|
| Brazil | 11 | 4,811 | 1 |
| South Korea | 9 | 3,110 | 1 |
| Saudi Arabia | 11 | 2,382 | 1 |
| France | 19 | 1,979 | 1 |
| Canada | 7 | 1,406 | 1 |
| India | 12 | 1,087 | 1 |
| China | 2 | 203 | 0 |
| United Arab Emirates | 6 | 132 | 0 |
| Finland | 5 | 110 | 0 |
| United Kingdom | 9 | 102 | 0 |
| Japan | 32 | 77 | 0 |
| Singapore | 6 | 67 | 0 |
Execution risk across the announced pipeline
Every project on the map carries execution risk the coordinates do not convey — transformer lead times of three to four years, gas turbine delivery backlogs, interconnection-queue position, local permitting and zoning friction, water availability, community engagement posture, and equipment supply chains that are already tight. The sites here have at least published coordinates and a visible development path. The larger universe of AI-compute announcements — site letters of intent, unoptioned parcels, pre-interconnection speculative plays, early press releases without a named site — carries materially higher execution risk and does not appear here. A future brief will take that announced-but-unmapped pipeline and look at what fraction is likely to energize on the timelines that have been stated.
Related research
- Energy Price & Queue Efficiency by Utility Territory — 90+ US utility territories scored on blended energy cost and interconnection speed.
- US Power Supply-Demand Deficit Forecast — bottoms-up generation pipeline model through 2035, including the state tier framework and regional 2030 deficit figures used here.
Sources
- Epoch AI, Notable AI Data Centers. Compute-thresholded frontier training campuses. epoch.ai/data/data-centers
- Epoch AI, GPU Clusters. Broader base of hyperscaler, sovereign, academic, and regional GPU installations. epoch.ai/data/gpu-clusters
- Natural Earth 1:110m Land. Coastlines for the base map, via the
world-atlaspackage (public domain). naturalearthdata.com - US state and ISO/RTO boundaries. HIFLD-sourced state polygons and ISO/RTO regional boundaries.
- China compute capacity. MIIT; NCSTI on the eight national hubs; National Data Bureau via SCMP; CSIS and SemiAnalysis on Ascend production + Hopper inventory; Epoch AI on the US–China frontier-compute gap; Huawei Gui’an; ByteDance Datong via DCD.