# Power, water and bills: what the data shows A brief for environmental and community groups. It uses the datasets in this repository and names each source. Current as of November 2026. --- ## 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.** The evidence shows two real ways that these campuses can put costs on a community: 1. **Electricity.** New data center load raises shared grid costs in regional power markets. The clearest measurement is in PJM, the grid that serves 13 states from Illinois to Virginia. 2. **Water.** Campuses that cool by evaporation can take a large share of a small town's water. Campuses with dry or closed-loop cooling take very little. Each impact depends on a choice that a utility commission, a city or a county can still set: - **Who pays** for new generation and transmission. - **How the campus cools**, and where its water comes from. - **What the operator must disclose**, site by site, every year. The data does not support "every data center raises your bill". It also does not support "data centers lower your bill". It does support this: the cost risk is real at the regional level, and rate design can assign it. The water risk is local and depends on the design, and most operators do not publish it. --- ## 1. Most of the load is a plan We track 442 campuses with 123 GW of stated power capacity. That figure is nameplate capacity, not the power the sites use today. | Status | World GW | Share | US GW | Share | |---|---:|---:|---:|---:| | Operational | 20.4 | 17% | 10.5 | 15% | | Partially operational | 16.5 | 13% | 12.5 | 18% | | Under construction | 41.0 | 33% | 23.5 | 35% | | Planned | 45.5 | 37% | 21.7 | 32% | | **Total** | **123.4** | | **68.1** | | GW = gross facility power where published, else IT load. A partially operational site carries its full target, not its live load. What this means for a community: - **70% of tracked capacity is planned or under construction.** In the US the share is 66%. For most sites, the tariff, the water contract and the permit are still under negotiation or open to challenge. - **Texas holds half of the tracked US capacity (33.2 GW), and 83% of it is planned or under construction.** One proposal, Fermi America in Amarillo, is 11 GW with no named tenant. - **A planned gigawatt is a request for power, not a load.** Some of these projects will not be built. Utilities and grid operators must still plan for the requests, and that planning has a cost. - **Nameplate is more than average draw.** Meta reports 18.06 TWh of data center electricity for 2024. That is an average of 2.06 GW. We track about 3.97 GW of Meta IT capacity, and some of it is not built yet. So the average draw was about half of the capacity on paper. For scale: the US planned and under-construction capacity alone is 45.2 GW. At full load for one year, that is about 396 TWh. Lawrence Berkeley National Laboratory (LBNL) estimates that all US data centers used 176 TWh in 2023 (4.4% of US electricity). LBNL projects 325–580 TWh by 2028. Full load is an upper bound. Real use will be lower, because some projects will stop and no campus runs at 100%. --- ## 2. Electricity: the cost goes through the regional market **Finding.** In PJM, the cost of data center load reaches every customer through the capacity market. The size of the effect is large. - PJM's Independent Market Monitor estimates that data center load added **$9.33 billion** to the 2025/26 capacity auction. The auction total was $14.69 billion. Without data center load, the monitor estimates $5.35 billion. The model includes 7.9 GW of existing and forecast data center load ([IMM, Part G, 3 Jun 2025](https://www.monitoringanalytics.com/reports/reports/2025/IMM_Analysis_of_the_20252026_RPM_Base_Residual_Auction_Part_G_20250603_Revised.pdf)). - The monitor's conclusion: "data center load growth is the primary reason" for the high capacity prices. - Capacity costs go into retail bills through each state's supply rates. New Jersey's 17–20% supply-rate step in June 2025 followed that auction. **Caveat.** The monitor's figure is a model result: the auction with and without data center load. It is not an accounting of bills. It covers all data centers in PJM, not only the AI campuses in this dataset. **What the state data shows.** Residential prices rose in every state where we track a site. The US average rose 33.2% from 2021 to the 2026 year to date (EIA). The rise follows the market structure more than local AI capacity: - The six states in our table that sit mostly in PJM (PA, NJ, OH, IL, VA, WV) rose by **44% on average**. The other 26 tracked states rose by **26%**. - Across the 32 states, the price rise has **only a weak relation** to the operating AI capacity that we track in each state. The rank correlation is 0.15, which is too weak to mean much for 32 states. - Iowa has 1.2 GW of tracked operating capacity. It had the smallest rise in the table (+14.3%). New York has 0.75 GW, and its price rose by 50.2%. **Caveat.** These are state averages, without controls for gas prices, storm costs or other causes. Our inventory covers large AI campuses only, not every data center. Virginia has the largest data center market in the world, but we track only 2.2 GW there. So the state comparison cannot prove that data centers raise bills. It also cannot prove that they do not. **The decision that matters: rate design.** From 2025 to 2026, most large utilities filed a data center rate class. Examples from our utility table: - Dominion Virginia GS-5: a class for loads of 25 MW or more, from 2027. It sets an 85% minimum take on transmission and distribution, and 60% on generation. - AEP Ohio: data centers pay for at least 85% of contracted capacity, for up to 12 years. - Wisconsin: the PSC made data centers pay 100% of new generation costs. - PGE Oregon: a new class with +29.7% for data centers and −1.3% for homes. Consumer advocates in Virginia, Ohio, Illinois, Wisconsin, Minnesota, Oregon and Pennsylvania argue that these terms are not strong enough. The full list is in `data/us_utility_impacts.csv`. **Utility claims that data centers lower bills.** Some utilities make this claim, for example MidAmerican Iowa, MDU North Dakota and Entergy. Treat it as a claim. In Mississippi and Arkansas, the contracts that support it are trade secrets under state law. --- ## 3. Water: local, set by design, mostly not disclosed **Finding.** Data center water use is small at national scale, but it can be a large share of a local supply. The cooling design sets the size, and the design is a choice. Measured or reported shares of local supply: | Place | Share | Basis | |---|---|---| | The Dalles, Oregon | ~40% of city water | City records, released after a lawsuit | | Loudoun County, Virginia | ~9% of average-day drinking water | Loudoun Water | | Columbus area, Ohio | ~3% of city supply (~1.2B gal/yr) | Columbus utilities | | Ridgeland, Mississippi | 82% of peak capacity, if the reported 3.7 MGD peak is correct | Reported, not verified | Volumes at single sites can be large. Google's Council Bluffs, Iowa campus consumed 1,346 million gallons in 2025. That is about the use of 12,000 US households, at the EPA figure of about 300 gallons per family per day. Meta's whole global fleet consumed about 786 million gallons in 2024, which is less than that one site. **Design sets the water.** At the same IT load, the reported water intensity differs by a factor of 500: | Operator | Water per kWh of IT energy | Water for 1 MW for one year | Cooling | |---|---:|---:|---| | Google | 1.02 L (consumption, derived) | 2.36M gal | Evaporative at large campuses | | Microsoft | 0.27 L | 0.63M gal | Mixed | | Meta | 0.19 L (withdrawal) | 0.44M gal | Mixed | | AWS | 0.12 L | 0.28M gal | Mixed | | Crusoe | ~0.002 L (design) | ~0.005M gal | Closed loop | Evaporative cooling uses less electricity and more water. Dry cooling uses more electricity and less water. So a water-stressed place and a grid-stressed place can rationally ask for different designs. **Liquid-cooled chips do not mean a waterless campus.** The rack can use liquid, and the plant can still reject that heat through cooling towers. Google's liquid-cooled TPU campuses do this. Ask about the plant, not the chip. **Caveat.** Only Google publishes water by site. For all other operators, our site water figures are the fleet average applied to the site. No site in our dataset has a metered site figure for water per unit of IT energy. For 32 of 61 US campuses that are planned or under construction, the cooling water dependency is unknown. --- ## 4. Disclosure is the bottleneck - Of 13 operators, **4 publish assured annual figures** for energy and water: Google, Meta, Microsoft and AWS. xAI, CoreWeave, Applied Digital, Lambda and Galaxy publish no usable figures. - **Secrecy, more than volume, starts most water disputes in our data.** Examples: The Dalles (records lawsuit), Eagle Mountain, Utah (the city contract keeps use confidential), Chesterfield, Virginia, and Racine, Wisconsin. - **Disclosure laws work.** Nebraska LB1010 and Utah HB76 require reporting. Mississippi SB 2001 goes the other way: it makes the contracts trade secrets. --- ## 5. What this dataset cannot tell you yet We state these gaps so that no reader takes the data for more than it is. - **Air emissions.** We do not track on-site gas generation, for example the xAI turbines in Memphis or the gas plants planned for Meta Hyperion. For many neighbours, this is the main health issue. - **Water at the power plant.** Thermal power plants use water too. Our water figures cover the campus only. - **Total data center load per utility.** We track large AI campuses, not every data center. The utility's own load forecast is the better source. - **Actual energy use per site.** We have capacity, not metered energy (MWh). - **Hourly carbon.** "100% renewable" claims are annual and global matches. They do not mean that the power at a given campus is renewable at every hour. --- ## 6. Questions to ask about a proposed campus These questions follow from the findings. Each one asks for a number that the operator or the utility has. 1. **Load.** How many MW does the contract request, and on what schedule? 2. **Who pays.** Does the tariff have a minimum take, an exit fee and a collateral requirement? Who pays for the new transmission? 3. **On-site generation.** What fuel, how many MW, and under which air permit? 4. **Cooling.** Is the plant evaporative, dry or hybrid? What is the design water use per MW, on the peak day and over a year? 5. **Water source.** Drinking water, reclaimed water or groundwater? What happens to the campus supply in a drought? 6. **Reporting.** Will the operator publish each year, for this site, the energy use (MWh), the water withdrawn and the water consumed? --- ## Sources and data - Site capacity and status: `data/published/data_centers.csv`. Method and caveats: `README.md`. - Utility prices, rate cases, water and named groups: `data/us_utility_impacts.csv`, `docs/us_water_energy.md`. Checks: `docs/us_impacts_verification.md`. - Operator energy and water efficiency: `data/operator_efficiency.csv`, `docs/operator_efficiency.md`. - Cooling design by site: `data/gpu_cooling.csv`, `docs/gpu_cooling.md`. - PJM Independent Market Monitor, *Analysis of the 2025/2026 RPM Base Residual Auction, Part G* (3 Jun 2025, revised). - LBNL, *2024 United States Data Center Energy Usage Report* (Dec 2024): https://eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report - EIA, *Electric Power Annual* Table 2.10 and *Electric Power Monthly* Table 5.6.B (state residential prices). - US EPA WaterSense, household water use (~300 gal per family per day). The state comparison in section 2 uses `pipeline/verification/eia_table.txt` and the operating capacity per state from the site dataset. The script is `pipeline/verification/price_vs_capacity.py`.