# AI Data Centers: Power, Water and Who Pays A source-linked inventory of **442 major AI and hyperscale data center campuses** in 71 countries. Companion studies cover US electricity prices, water and operator efficiency. Every row names its source. **Live dashboard: ** **For environmental and community groups, start with [`docs/community_brief.md`](docs/community_brief.md).** ## The thesis **Most AI data center capacity is not built yet. So the main community question is about decisions that are still open, not about damage that is already done.** - **Most of the load is a plan.** We track 123 GW of stated capacity. Planned and under-construction sites hold 70% of it, and operational sites hold 17%. In the US the figures are 68 GW, 66% and 15%. - **Electricity costs reach households through the regional market.** PJM's market monitor estimates that data center load added $9.33 billion to the $14.69 billion 2025/26 capacity auction. Rate design decides who pays that cost: minimum-take tariffs, direct assignment of transmission, and bring-your-own generation. - **Water is local, and the design sets it.** At the same IT load, the reported and design water intensities differ by a factor of 500 between evaporative and closed-loop designs. One evaporative campus (Google Council Bluffs) consumed more water in 2025 than Meta's whole global fleet in 2024. - **Disclosure is the bottleneck.** Only 4 of 13 operators publish assured annual energy and water figures. Only Google publishes water by site. Secrecy, more than volume, starts most of the water disputes in our data. The data does not support "every data center raises your bill". It also does not support "data centers lower your bill". The brief gives the evidence, the caveats and the gaps, for example air emissions from on-site gas, which we do not track. **About the compute figures.** The inventory also estimates compute in H100-equivalents (H100e), so that sites with different chips can be compared. H100e measures scale, not resource use. For power, water and cost, use the MW, water and price columns. ## Files | File | Description | |---|---| | `docs/community_brief.md` | **Start here.** The thesis and the evidence on power, water and bills, for community readers | | `data/published/data_centers.csv` | Final dataset, one row per campus, sorted by best H100e estimate | | `data/published/data_centers.json` | Same data as JSON (typed numbers) | | `index.html` | Self-contained dashboard: capacity map, who-pays view (state prices, PJM estimate, utility finder), water and efficiency views, scenario lab. Open directly in a browser. | | `data/canonical/data_centers.csv` | Canonical source of truth (edit this, then rebuild) | | `data/raw/data_centers_raw.csv` | Original research capture (kept for provenance) | | `pipeline/build.py` | Computes H100e columns and regenerates outputs: `python pipeline/build.py` | | `pipeline/frontend/` | Maintainable HTML template, CSS, and JavaScript embedded into `index.html` by the build | | `pipeline/country_table.py` | Regenerates the country-level summary table | | `data/us_utility_impacts.csv` | 55 US utility territories: prices, rate cases, water, complaint groups | | `docs/us_water_energy.md` | US local energy/water impacts report | | `docs/us_impacts_verification.md` | Verification log for the US impacts study | | `data/operator_efficiency.csv` | PUE / WUE / water-per-MW-year for 13 operators | | `docs/operator_efficiency.md` | Operator efficiency analysis and comparability caveats | | `data/gpu_platforms.csv` | 41 accelerator platforms: TDP, rack cooling, heat-to-liquid share, sources | | `data/gpu_cooling.csv` | Per-site accelerator match, heat estimate, rack cooling, heat rejection, water | | `data/gpu_cooling.json` | Same, typed JSON | | `pipeline/build_gpu_cooling.py` | Generates the per-site cooling dataset from the platform table | | `docs/gpu_cooling.md` | Accelerator/cooling analysis: why GPU vendor does not determine water | | `DATA_DICTIONARY.md` | Column definitions for all five published CSVs | | `AGENTS.md` | Operating brief for AI agents: role, conventions, git and credential policy | | `PROGRESS.xml` | Append-only log of decisions, incidents, verifications and changes | Companion studies: - `data/us_utility_impacts.csv` + `docs/us_water_energy.md`: US local energy/water impacts, with `docs/us_impacts_verification.md`. - `data/operator_efficiency.csv` + `docs/operator_efficiency.md`: PUE/WUE comparison across Google, Meta, Microsoft, AWS, xAI and the neoclouds. - `data/gpu_cooling.csv` + `docs/gpu_cooling.md`: per-site accelerator and cooling evidence, with rack cooling separated from facility heat rejection. ## Interactive dashboard `index.html` follows the thesis, in this order: - **Overview:** stated power by build status (world and US), and four finding cards: power, bills, water and disclosure. - **Capacity map:** search and filter 442 campuses. Marker size and the leaderboard use stated power by default. Compute (H100e) is an option. Each site shows its status, its note and its source, and you can export the filtered rows. - **Who pays:** the residential price rise for each state with a tracked site (EIA), colored by PJM membership. Each bar shows the tracked operating AI capacity in that state. A card gives the PJM market monitor's capacity-auction estimate. A finder for the 55 utility territories: prices, rate cases, named groups, water and sources, with the verification grade. - **Water:** documented shares of local supply, then the operator PUE/WUE chart and the 13 operator disclosures. - **Site evidence:** cooling energy and water per MW for every campus. A coverage strip states how many sites can answer it. Of 442 sites, 306 have a power denominator, 127 an operator-disclosed PUE, 92 a water intensity, and 87 both. Those 87 collapse onto **six distinct positions**, because both numbers are operator-level. A rack-cooling × heat-rejection matrix covers all 442 sites. The page marks the industry-default PUE of 1.20 with `*`. - **Scenario lab:** annual facility electricity and water for two operator profiles at a chosen IT load. These are unit conversions at full utilization, not campus forecasts. - **Method:** power, water and compute, and the list of what the page does not track yet. Edit the sources in `pipeline/frontend/`, then run `python pipeline/build.py` to regenerate the self-contained page. The build embeds the site, operator, cooling, utility and state price values at build time. The map tiles, Leaflet, and web fonts need an internet connection. ## H100e methodology An **H100e** = the compute of one NVIDIA H100 SXM. Operators do not publish H100e, so every value here is an estimate. The build computes four estimates. `h100e_best` uses the most direct available signal: 1. **`h100e_gpu` — from announced GPU counts** (most direct). Conversion factors (dense FP8 compute ratio vs H100's 1,979 TFLOPS): - H100 / H200 / H800 / unspecified "GPUs" = **1.0** - B200 / GB200 = **2.25**. GB300/B300 = **2.5** - A100 = **0.315**. Trainium2 = **0.34**. Ascend 910B ≈ **0.35**. L40S = **0.19** - AMD MI355X = **2.5** (5.0 PF dense OCP-FP8 per AMD ÷ 1.979 = 2.54, rounded like B300) - **Mixed generations.** If the site publishes a count per model, we add up count × factor for each model. If it does not, and the models have different factors, `h100e_gpu` stays blank and `h100e_best` uses the next method. If only part of the fleet has a published count, we convert that part and the note calls the value a floor. 2. **`h100e_third_party`** — estimates from Epoch AI's data center directory, recorded during research. We use these for several Stargate sites, China sites, DayOne and others. A third-party H100e always goes in this column, also when the third party derived it from GPU counts. 3. **`h100e_flops`** — announced cluster peak PFLOPS (dense FP8-equivalent) ÷ 1.979. 4. **`h100e_power`** — power-based estimate. Assume 60% of IT load powers H100-class accelerator systems at ~1.2 kW per GPU, which gives **500 H100e per MW of IT load**. If we know only gross facility power, IT load = MW ÷ 1.2 (PUE 1.2), which gives **≈417 H100e per MW**. `h100e_method` says which estimate `h100e_best` came from: `gpu_count` > `third_party` > `flops` > `power`. ## Columns `name, operator, city, state_province, country, latitude, longitude, power_capacity_mw, it_load_mw, announced_gpu_count, gpu_types, cluster_peak_pflops, h100e_gpu, h100e_third_party, h100e_flops, h100e_power, h100e_best, h100e_method, status, status_note, source_url, location_shared_with` - `power_capacity_mw` = gross facility/utility power. `it_load_mw` = IT load where reported. - `location_shared_with` = the number of other sites at the same point (to ~1 km). 41 sites share a point with at least one other site, for example seven in Riyadh. Those coordinates are probably a city centre, not the campus. The map spreads such sites on a small ring and marks them as approximate. - Blank = not published or not found. **We invented no values**. Where a figure is a rumor or analyst estimate, `status_note` says so. - `cluster_peak_pflops` appears only where officially announced. Most are blank, because few operators publish it. The Deutsche Telekom 0.5 EFLOPS claim is inconsistent with its GPU count, and we keep it as announced (see note). ## Status classification Each site has a `status` column (also color-coded on the map): | Status | Meaning | Count | |---|---|---| | `operational` | Fully operating. The campus may still have announced expansions | 199 | | `partially_operational` | First phase(s) live, but the campus is still ramping to its target capacity | 35 | | `under_construction` | Construction underway. No live AI capacity yet, or first energization imminent | 117 | | `planned` | Announced or leased, construction not yet started (incl. some with uncertain status) | 91 | So **only 199 of 442 sites (45%) are fully operational**. Gigawatt-scale projects still under construction dominate the field. Judged by H100e, the split is starker. About **81%** of the total tracked H100e sits in sites that are not yet fully operational. 70% sits in sites that are only planned or under construction. Fully operational sites hold 11.6M H100e (19%). The biggest single estimates — Fermi America, Wonder Valley, Stargate Saline, Port Washington, Hyperion, Stargate UAE — are all planned or under construction. (Before the Nov 2026 review this paragraph said "roughly two-thirds". That was wrong: the share was 81% then too.) Chinese sites mostly use Ascend or heterogeneous chips. Their announced "X万P" figures are FP16/mixed-precision and recorded in notes. We do not convert them into `cluster_peak_pflops`. ## Caveats - **Scope**: major AI/hyperscale campuses only (roughly 100 MW+ or named 10k+ GPU clusters). We exclude ordinary colo/enterprise data centers and national-lab supercomputers (Frontier, El Capitan…). The exceptions are systems that are notable AI systems, for example JUPITER/JAIF and the EuroHPC AI factories. - **Estimates span very different confidence levels**. An Epoch AI projection for a 2028 build-out and a measured IT load are not equally reliable. Read `status_note` before you quote any number. - Power-based H100e is an upper-bound-style estimate. It credits the whole campus to H100-class accelerators, which overstates sites that host general cloud workloads. **It supplies 41.6M of the 61.4M total (68%).** Most of that is nameplate capacity at sites that are not built yet. - We compiled the data via web research from operator press releases, regulator filings, Epoch AI, DataCenterDynamics, and press coverage (sources in `source_url`). The data is current as of **November 2026**. The field moves fast, and many "under construction" figures are targets. - GPU conversion factors are dense-FP8 compute ratios. They ignore memory bandwidth, FP4, and interconnect differences that matter a lot in practice. ## Headline numbers (as of compilation) - **442 campuses**, 349 with a quantifiable H100e estimate - **~61.4M H100e** total across tracked sites. This is not a count of installed compute: - 68% (41.6M) is power-based, 26% (15.7M) third-party, 7% (4.1M) from GPU counts. - 70% (43.1M) is at sites that are only planned or under construction. Fully operational sites hold 11.6M (19%). - The two largest figures, Fermi America's ~4.6M and Wonder Valley's ~3.1M H100e power-based estimates, are speculative proposals without named tenants. - Largest tracked sites: Stargate Saline Township MI (~2.9M), Stargate Port Washington WI (~2.6M) and Meta Hyperion LA (~2.1M power-est). Also Stargate UAE (~2.0M), xAI Colossus 2 (~1.1M) and Stargate Abilene TX (~1.0M). - Nov 2026 review: the total was ~64.0M. A correction pass (`pipeline/verification/verify_apply5.py`) applied the mixed-generation rule and moved Epoch estimates to the third-party column. It also removed an unverified 1,289 MW figure for Meta Prineville. Meta Prometheus changed from 2.25M (1M-GPU target × GB200 factor) to 347,500 (695 MW IT). See `PROGRESS.xml` C-entry for every row. ## US local impacts (water & energy prices) Companion research on the 121 US sites: **`data/us_utility_impacts.csv`** (55 rows, one per utility territory). Each row holds the serving utility and type, a 5-year energy price trend, rate events and named community groups that complain about bills. It also holds documented water consumption, water complaints, sources and a verification grade. See also **`docs/us_water_energy.md`** (summary, headline findings, and the water big-drinker table). We re-based all price series on EIA annual data (Table 2.10) and the Jan–Aug 2026 monthly table. We checked the load-bearing water and rate claims against primary sources. See **`docs/us_impacts_verification.md`** for the correction table and residual caveats. Key takeaways: PJM capacity costs drove the sharpest 2021→2026 bill increases (PA +53.6%, NJ +47.3%, OH +47.3%, IL +45.9%, VA +41.7%, US average +33.2%). The 2025-26 period brought a wave of data-center rate classes (Dominion GS-5, AEP Ohio DCT, Duke, PGE Oregon +29.7% on DCs). Water-disclosure secrecy is the most common trigger for complaints (The Dalles, Chesterfield "Peanut", Eagle Mountain confidentiality clause, Racine lawsuit). This research found and fixed three dataset errors. Cipher Barber Lake is in Colorado City TX, Nscale is in the town of Madison, Rockingham County NC, and Boydton is co-op territory, not AEP. ## Operator efficiency (PUE / WUE) Companion comparison of Google, Meta, Microsoft, AWS, xAI and the neoclouds: **`data/operator_efficiency.csv`** (PUE, WUE, water per MW-year, low-carbon share, disclosure grade, sources) and **`docs/operator_efficiency.md`** (analysis). Headline: Meta is best on both axes among the big four (PUE 1.08, WUE 0.19 L/kWh). AWS uses least water (0.12). Google has the best PUE of the largest fleet (1.09) but is 5–8× more water-intensive per kWh (1.02, derived). Microsoft sits between (1.17 / 0.27). Energy and water trade off: evaporative cooling cuts PUE and raises water. Nebius and Crusoe post the best water numbers on cold-climate liquid-cooled designs. CoreWeave, Applied Digital, Lambda, Galaxy and xAI publish no usable PUE/WUE. ## Accelerators, cooling and water Companion study of what hardware each site runs and what that implies for cooling. **`data/gpu_platforms.csv`** (41 accelerator platforms with TDP, rack system, rack cooling, heat-to-liquid share and sources), **`data/gpu_cooling.csv`** / **`.json`** (one row per site) and **`docs/gpu_cooling.md`** (analysis). `pipeline/build_gpu_cooling.py` generates them after `pipeline/build.py`. Headline: **the GPU vendor does not determine water use.** Rack cooling (air, direct-to-chip liquid, immersion) and facility heat rejection (evaporative towers, dry coolers, hybrid) are separate decisions. Google's liquid-cooled TPUs still reject heat through cooling towers. NVIDIA's liquid racks can run on dry coolers, but only if the operator builds them. Of 442 sites, 177 name a platform, and 52 have enough evidence for an accelerator-heat estimate (36 single TDP, 16 ranged). 93 carry a water figure, and every one is an operator-fleet WUE scaling, not a site measurement. Two of those (xAI) rest on a stated 1.0 L/kWh estimate rather than a disclosure. Hardware confidence is B 39 / C 65 / D 338, with no A. No site combines an official TDP, a published count and site-specific cooling evidence. The dashboard's **site efficiency** view turns that coverage into the headline it implies. The page plots a campus only if it has both an *operator-disclosed* PUE and a water intensity. The industry-default 1.20 that fills 315 rows does not count as evidence. That leaves 88 campuses, and they land on six distinct positions (Google 26, Meta 20, Microsoft 19, AWS 14, Nebius 5, Crusoe 4). The PUE is the operator's fleet figure, and the water is the operator's fleet WUE applied to site IT load. Bubble area is the number of campuses at that position. The rack-cooling × heat-rejection matrix is the part that stays genuinely per-site. 36 sites are evaporative with no rack evidence, and 39 are direct-to-chip with unknown rejection. Only 34 of 442 name both a rack cooling type and a rejection type. ## Country-level summary | Country | Sites | MW (known) | H100e (known) | |---|---:|---:|---:| | United States | 121 | 68,148 | 35,895,726 | | China | 29 | 1,926 | 836,420 | | Malaysia | 25 | 4,481 | 2,624,634 | | India | 25 | 7,560 | 3,002,471 | | Mexico | 23 | 1,828 | 838,280 | | Canada | 20 | 10,404 | 4,341,931 | | Japan | 16 | 2,094 | 880,405 | | South Korea | 13 | 913 | 441,789 | | Saudi Arabia | 12 | 2,616 | 1,152,050 | | South Africa | 11 | 887 | 374,834 | | Brazil | 10 | 3,282 | 1,297,416 | | Australia | 9 | 2,550 | 1,173,667 | | United Kingdom | 9 | 990 | 519,734 | | United Arab Emirates | 8 | 1,360 | 2,171,001 | | Germany | 7 | 260 | 148,435 | | France | 6 | 2,430 | 1,073,833 | | Italy | 5 | 707 | 318,208 | | Sweden | 5 | 820 | 335,067 | | Poland | 5 | 43 | 33,650 | | Israel | 4 | 196 | 89,382 | | Finland | 3 | 985 | 428,167 | | Denmark | 3 | 810 | 357,500 | | Spain | 3 | 400 | 160,000 | | Ireland | 3 | 675 | 301,250 | | Netherlands | 3 | 350 | 151,800 | | Norway | 3 | 272 | 146,606 | | Turkey | 3 | 300 | 125,003 | | Chile | 3 | 270 | 125,000 | | Nigeria | 3 | 177 | 88,500 | | Kazakhstan | 2 | 300 | 126,011 | | Taiwan | 2 | 400 | 150,000 | | Indonesia | 2 | 190 | 132,386 | | Bangladesh | 2 | 200 | 83,333 | | Singapore | 2 | 206 | 103,000 | | Vietnam | 2 | 144 | 61,333 | | Thailand | 2 | 147 | 65,167 | | Kenya | 2 | 30 | 11,250 | | Pakistan | 2 | 10 | 4,042 | | Uruguay | 2 | - | - | | Argentina | 1 | 500 | 208,333 | | Trinidad and Tobago | 1 | 450 | 187,500 | | Oman | 1 | 400 | 166,667 | | Jordan | 1 | 400 | 166,667 | | Belgium | 1 | 300 | 130,000 | | New Zealand | 1 | 280 | 116,667 | | Costa Rica | 1 | 200 | 83,333 | | Paraguay | 1 | 106 | 44,167 | | Armenia | 1 | 100 | 41,667 | | Portugal | 1 | 1,200 | 31,500 | | Philippines | 1 | 50 | 25,000 | | Guatemala | 1 | 50 | 20,833 | | Egypt | 1 | 20 | 10,000 | | Morocco | 1 | 16 | 8,000 | | Iceland | 1 | 5 | 5,821 | | Slovenia | 1 | - | 5,053 | | Uzbekistan | 1 | 12 | 5,000 | | Romania | 1 | - | 2,527 | | Czechia | 1 | - | 430 | | Serbia | 1 | - | 58 | | Greece | 1 | - | - | | Switzerland | 1 | - | - | | Qatar | 1 | - | - | | Bahrain | 1 | - | - | | El Salvador | 1 | - | - | | Ethiopia | 1 | - | - | | Bulgaria | 1 | - | - | | Lithuania | 1 | - | - | | Hungary | 1 | - | - | | Ukraine | 1 | - | - | | Azerbaijan | 1 | - | - | | Peru | 1 | - | - | MW = sum of known gross power, with IT load where gross data is missing. It mixes announced, operating and speculative capacity, so treat it as a scale indicator, not a current draw. H100e totals are the sum of per-site best estimates. ## Coverage history The counts in this section are the totals after each research pass. Later passes added more sites, so the current counts are in the country table above. The first research pass was thin on Asia and empty on Latin America. Later deep-dives expanded China (24 sites, the "Eastern Data, Western Computing" intelligent computing centers). Malaysia grew to 22 sites, because Johor is one of the fastest-growing AI DC markets. They also expanded Japan (15), South Korea (12), Europe (+26, incl. EuroHPC AI factories and gigafactory candidates), the Middle East (+19), and Latin America (+15). Most of the Latin America entries are an announced pipeline. No operating 50MW+ AI-dedicated campus or 10k+ GPU cluster exists in the region yet. The largest GPU deployment is Mexico's planned 14,480-GPU Coatlicue supercomputer. A final North America pass added Canada (+14). It includes Meta's first Canadian DC, the 1GW Sturgeon campus, Canada's sovereign AI factories from TELUS/Bell/Calcul Québec, and the speculative 7.5GW Wonder Valley proposal. It also added deeper Mexico (+17, the Querétaro corridor: AWS/Google/Oracle/Alibaba regions, ODATA QR01/02/04, Ascenty, Bleeding Edge's LatAm-first Blackwell B300 neocloud). It added the first Central America/Caribbean entries (+4, all small or speculative. The region has no AI-scale campus yet). Three final passes added Africa (+15: Teraco Johannesburg/Cape Town, Nigeria's Kasi and Rack Centre, Egypt's sovereign AI DC, Morocco's Nexus. No published GPU counts exist anywhere on the continent). They added Eastern Europe, the Caucasus and Central Asia (+18). That covers EuroHPC AI factories in Czechia/Romania/Bulgaria/Lithuania/Slovenia/Hungary, Kazakhstan's Alem H200 supercomputer and 100k-GPU Data Center Valley, and Armenia's Firebird. They also added India/South Asia (+21). That covers the IndiaAI national GPU pool, Yotta's Blackwell Ultra buildout, L&T Vyoma's 10k-B300 factory for Together AI, Google/Adani Vizag and other gigawatt campuses. Pakistan and Bangladesh get their first entries. A US re-sweep (+63) then filled the biggest remaining hole. It added Google's TPU-campus fleet (Columbus/New Albany/Lancaster OH, Bristow VA, Omaha/Lincoln/Papillion NE, Mesa AZ, Midlothian TX, Kansas City, Storey County NV). Several of those sites carry 100k-330k H100e Epoch estimates. It also added Meta's remaining sites (Rosemount, Kuna, Cheyenne, Eagle Mountain, Los Lunas, Montgomery, Temple, Bowling Green, El Paso 1GW). It added Microsoft (Pecos 2GW, Goodyear, Project Osmium) and AWS (Berwick PA Rainier leg, Richmond County NC, Georgia, Arizona). It added QTS Richmond (~562MW). It also added the crypto-to-AI conversion wave (Hut 8 Beacon Point 1GW for Nvidia, Riot, Cipher/AWS, Bitdeer, MARA, Core Scientific, Bitfarms, Applied Digital PF3). It also added neoclouds (Nscale Ward County, Vultr's 24k-MI355X Springfield supercluster, CloudBurst, and Fermi America's proposed 11GW Amarillo HyperGrid). That last site is now the single largest power figure in the dataset, and it is still speculative. A final country-by-country sweep (+63) closed the remaining regional gaps. Australia went from 1 to 9 sites. The additions are AirTrunk SYD2/SYD3/MEL1/MEL2, CDC Marsden Park at 720MW (the Southern Hemisphere's largest), NEXTDC M3, Firmus's 18,400-GPU GB300 cluster and AWS Sydney. Singapore went from 0 to 2 (Nxera Tuas, AirTrunk), and Indonesia 0 to 2 (Oracle and PDG in Batam). Thailand went 0 to 2, the Philippines 0 to 1 (VITRO), and New Zealand 0 to 1 (Datagrid's 1GW Invercargill project). Denmark went from 0 to 3 (Meta Odense ~475MW, Google Fredericia, CA Skive), and Belgium 0 to 1 (Google Saint-Ghislain). Ireland went from 1 to 3 (Meta Clonee, AWS Cruiserath), and the Netherlands 1 to 3 (Google Eemshaven, Microsoft Middenmeer). Spain gained Microsoft Aragon, France SoftBank Bouchain/Eclairion, Italy Trino/Milan, and Sweden atNorth Sollefteå. It also added Canada's Saskatchewan/New Brunswick wave, Brazil's RT-One/Ada/Tecto buildout, Chile's Cerrillos and Ascenty SCL04/05, and Peru's first entry. Also the Saudi Microsoft region, Oman/Jordan 400MW proposals, Israel's Ofek, Kenya's iXAfrica, and India's STT Palava/PDG MU1/AdaniConneX Noida/CtrlS Future City additions. ## Regenerating ```bash python pipeline/build.py # canonical CSV -> published CSV/JSON + index.html python pipeline/build_gpu_cooling.py # published CSV + gpu_platforms.csv -> gpu_cooling.csv/json ```