PJM states, 2021 → 2026
—
AI DATA CENTERS · POWER, WATER AND BILLS · NOV 2026
Most of it is
not built yet.
We track — large AI and hyperscale campuses. — of their stated power is planned or under construction. So the cost to a community depends on decisions that are still open: who pays for the grid, how a campus cools, and what it must disclose.
TRACKED POWER CAPACITY
NAMEPLATE
—GW
Stated capacity, not current draw. Gross facility power where published, else IT load. A partially operational site carries its full target.
—tracked in the US
—of US capacity is planned or building
—of capacity is fully operational
—countries
of tracked capacity is a plan. A planned gigawatt is a request for power, not a load. The tariff, the water contract and the permit are still open.
02 · BILLS $9.33Bof PJM's $14.69B 2025/26 capacity auction came from data center load, by the market monitor's estimate. Rate design decides who pays it.
03 · WATER 500×is the spread in water per kWh between evaporative and closed-loop designs. The cooling plant decides the water, not the chip.
04 · DISCLOSURE 4 of 13operators publish assured annual energy and water figures. Only Google publishes water by site. Secrecy starts most water disputes.
01 / CAPACITY
Where the load
is coming.
Circle area shows stated power. Color shows build status. Purple and blue sites are not running yet. Select a site to see its source.
MARKER AREA
Smaller
Larger
— sites
Showing the complete inventory.
Power is nameplate capacity, not average draw. Read each site's note before you quote it.
FILTERED LEADERBOARD
Largest campuses by stated power
| Rank / campus | Status | IT load | Power | Compute (H100e) |
|---|
02 / WHO PAYS
Bills rose everywhere.
The link runs through the market.
Residential prices rose in every state where we track a site. The rise follows the regional power market more than local AI capacity. The measured data center cost is in the PJM capacity market, and rate design decides who carries it.
Other tracked states
—
Weak link to local AI capacity
Rank correlation between each state's price rise and the operating AI capacity we track there. State averages, with no controls for fuel or storm costs.
RESIDENTIAL PRICE RISE, 2021 → 2026 YTD (EIA)
States with a tracked site
Mostly in PJM Other US averageMW = tracked operating AI capacity · EIA Table 2.10 and EPM Table 5.6.B
US UTILITY TERRITORIES
Find your utility: prices, rate cases, water and named groups
—
Read before quoting: price trends are EIA state averages unless a docket gives the utility's own figure. 2026 is January to August. A listed group is a named, sourced complaint. A territory with no listed group can still have concerns. Grade A: price and water or rate checked against primary sources. Grade B: price checked, water from secondary sources or not published.
03 / WATER
Water is set
by design.
National totals are small, but one campus can take a large share of a town's water. Evaporative cooling saves electricity and uses water. Dry and closed-loop cooling do the reverse. The operator chooses, and a permit can ask.
Google's share of city water. The city released it only after a records lawsuit.
Data centers' share of average-day drinking water (Loudoun Water).
About 1.2B gallons a year of drinking-quality water (city utilities).
Of the system's peak capacity, if the reported 3.7 MGD AWS peak is correct. Not verified.
Lowest hyperscaler PUE
—
Lowest hyperscaler WUE
—
Evaporative cooling can cut compressor energy—but consume more water. There is no single winner without deciding which resource matters.
THE ENERGY–WATER FRONTIER
Lower is better on both axes
Design target / range midpoint—
Read before comparing: reporting years, site boundaries, and WUE definitions differ. Nebius's WUE is one Finland site. Crusoe's PUE is a design range, and Google's WUE is derived from consumption. PUE measures facility overhead, not useful compute per joule.
See all 13 operator disclosures
| Operator | Year | PUE | WUE L/kWh | Disclosure | Boundary |
|---|
04 / SITE EVIDENCE
Every campus,
scored on evidence.
The operator chart above compares six disclosures. This view asks the same question of every campus — how much cooling energy and how much water per megawatt. The 88 campuses we can score collapse onto a handful of positions, because both numbers are operator-level. That is the finding.
SITE-LEVEL PUE × WATER INTENSITY
Cooling energy and water are separate decisions
Bubble size = number of campuses at that position · click to open the largest—
What this shows — and what it does not. The points collapse onto operator-level rows. A campus's PUE is usually its operator's fleet figure, and its water figure is the operator's fleet WUE applied to site IT load. Neither is a meter reading. That clustering is the finding: at this level of disclosure, site efficiency is operator efficiency. The matrix is the part that is genuinely per-site. A liquid-cooled rack can still sit behind an evaporative plant, and an air-cooled hall can run on dry coolers.
All 442 sites · efficiency ledger
—
* Industry default PUE of 1.20 — no operator disclosure for this site. ≈ Water intensity built on a stated WUE estimate (xAI, 1.0 L/kWh), not an operator disclosure. H100e/MW is the site's best compute estimate divided by IT load. Power-based estimates sit at 500 by construction. Waterless designs stay at 0 here but are dropped from the chart, whose log axis cannot plot zero. The ledger's “PUE + water” filter therefore returns one more site than the chart.
05 / SCENARIO LAB
Same IT load.
Different resource bill.
Compare two disclosure profiles at full utilization for one year. This is a unit conversion, not a campus forecast.
VS
Annual electricity = IT MW × PUE × 8,760 hours. Water uses each operator's disclosed or derived WUE normalizer. It assumes constant load and excludes embodied resources and local grid emissions. Operators that miss either metric are not offered.
06 / HOW TO READ THE NUMBERS
Three measures.
Three different questions.
Power, water and compute are related, but none can stand in for the others. For local impacts, use power and water. Compute is a scale measure only.
How much power?
energy (MWh) = MW × hours × utilization
Stated capacity is the most a campus can draw. Average draw is lower. Meta's 2024 fleet drew about half of the capacity we track for it. PUE is facility energy ÷ IT energy, so a PUE of 1.10 adds 10% for cooling and power losses.
POWER · NAMEPLATE IS NOT DRAWHow much water?
WUE = water litres ÷ IT kWh
Withdrawal is the water a campus takes. Consumption is the water it does not return, mostly by evaporation. They are not interchangeable. Outside Google, our site water figures are a fleet average applied to the site, not a meter reading.
WATER · CHECK THE BASISHow much compute?
H100e = compute ÷ 1.979 PFLOPS
H100-equivalents compare sites with different chips. The value comes from announced GPU counts, third-party estimates, peak FLOPS or power. 68% of the total comes from power, so it adds no new information about resource use.
SCALE · NOT A RESOURCE MEASUREWhat this page does not track yet. Air emissions from on-site gas generation. Water used at the power plants that supply a campus. Total data center load per utility (we track large AI campuses only). Metered energy per site. Hourly carbon: "100% renewable" claims are annual, global matches.