AI, in real life
Move a control.
Make a big number click.
One campus. Thousands of homes.
As much electricity as
Compare the same year of electricity use on both sides.
Find kWh on your electricity bill. Use energy, not dollars.
The real number: 2,326,875 MWh campus electricity per year ÷ 10,791 kWh per home-year × 1,000.
This is a volume comparison with household electricity purchases, not a forecast of your bill. Your input stays in this browser unless you choose to share the comparison.
The campus is Colossus, Memphis (Boxtown), Tennessee. These are annual estimates at recorded operating capacity, distributed evenly over time. They cover the whole campus, including non-AI uses. Actual operation varies.
Default home: 10,791 kWh/year, from EIA’s 2022 US average purchases. A backyard pool is defined here as 20,000 US gallons. Cars use the EPA’s 4.6 tonnes CO₂/year comparison; campus estimates are CO₂e.
Tap to see what changes.
About 19.1 gallons return.
About 80.9 are consumed, mainly through evaporation.
Per 100 withdrawn, based on Google Papillion’s FY2025 report. Returned water is discharged, not necessarily returned to the same source or fit to drink.
See the reported volumes →0homes’ yearly electricity purchases
No operating capacity recorded for Hyperion. The announced project is a separate scenario.
See this project’s assumptions →Save a project. Compare it. See changes in our records when you return.
Build your watchlist Free · saved on this deviceEnter a US ZIP code. We’ll show you the closest places in our guide.
No account needed. Your ZIP gives an approximate starting point.
Just exploring?
Bigger circles mean more estimated water use or pollution. Select a circle for a short summary, or use the table to browse by name. These estimates include the power plants supplying each site.
Cooling towers evaporate water on site. Power stations evaporate more, somewhere else, to make the electricity — that half almost never appears in a company's own reporting.
Distant power plants versus turbines running on the site itself. On-site combustion is a small share of the total and a large share of what the nearest neighbourhood actually breathes.
No operator publishes per-site water or air emissions. Everything here is modelled from disclosed capacity rather than copied from a report that does not exist. The whole model is five lines:
facility_MWh = IT_MW × 8760 × load_factor × PUE water_basin = facility_kWh × WUE (evaporated on site) water_power = facility_kWh × grid_water (evaporated generating it) NOₓ SO₂ PM = grid_MWh × grid_rate + onsite_MWh × turbine_rate health_$ = Σ tonnes × damage_value
Water consumed is not water withdrawn. A cooling tower takes water in and evaporates most of it; what evaporates leaves the basin. That is the number here — the water the watershed does not get back — and it is far smaller than the withdrawal figures utilities publish.
Seventeen sites publish a real annual water figure. For those, WUE is back-solved so the model reproduces the published number exactly, and the site is marked fitted to disclosure. Fitting them showed the remaining engineering estimates ran roughly 1.7× high, so they were corrected by that factor.
Air pollution here means criteria pollutants, not carbon. Nitrogen oxides, sulphur dioxide and fine particulates are what damage lungs, and they are what an air permit actually regulates. CO₂e is reported alongside, but it is a climate number, not an air-quality one. Emission rates come from the generation mix of the grid a site draws on. Where a campus runs its own gas turbines, that share is charged at an uncontrolled simple-cycle rate of 0.45 kg NOₓ/MWh — roughly fifteen times a modern combined-cycle plant with selective catalytic reduction, which is the entire substance of the permit fights in Memphis and Abilene.
Health damages are a range, not a point. The upper figure uses EPA benefit-per-ton values of $18,000 / $48,000 / $420,000 per tonne of NOₓ, SO₂ and PM2.5 at a ~$12 million value of a statistical life. The lower figure uses the midpoint of EASIUR’s national averages, two to four times smaller. Implied deaths divide the EPA dollar figure by that VSL. Both are national averages; real damages scale with how many people live downwind.
Diesel backup is now a modelled term where the permit was retrieved. Engineering estimate: standby MW × 50 test hours × 30% load × 8 kg NOₓ/MWh (mid of Tier 2 and uncontrolled AP-42). Where only a genset count exists, each unit is treated as 2.5 MW. Campuses without a retrieved inventory still exclude diesel, so those on-site figures remain floors. Permitted NOₓ caps are shown separately; many sit 1–4 tons under the 100-ton Title V major-source line by excluding emergency hours.
The “fully built” scenario is announced capacity, not a queue. ERCOT’s large-load queue has been quoted above 400 GW; Grid Strategies says utility forecasts are overstated by roughly 25 GW through duplicate requests and 96–150% load factors. Sites flagged speculative stay in the announced column and are not treated as demand.
This is per-site criteria pollution, not a company carbon score. Climate Power’s similarly named tracker scores eight companies on grid carbon intensity. Those are different numbers.
| Grid | NOx | SO₂ | PM2.5 | Water | CO₂e |
|---|
Pollutant rates in kg per MWh generated; water in litres per kWh; CO₂e in kg per kWh. Hydropower reservoir evaporation is included in the water column, which is why the Pacific Northwest and Norway look water-expensive despite being nearly carbon-free.
A little context helps
Data centres use water for cooling. The power plants supplying them can use water too. Each site’s summary separates the two.
We estimate pollution from the energy a site uses. That helps compare projects, but doesn’t tell you the air quality at your home.