A Data Center and a Neighborhood

Take the same piece of ground twice. Build a hyperscale data center on it, then build houses on it, and compare what each one draws from the grid and the water main. The power answer is not close. The water answer is closer than the argument about it suggests, and it is getting closer every year.

Blue is the data center. Oxblood is the neighborhood. Every figure below is computed from the four inputs, and the formulas are stated at the bottom so the arithmetic can be checked rather than trusted. The companion reading is Dark Fiber With a Power Bill, which argues that power, not capital, is the binding constraint on the build-out.

Inputs

100 acres The same land is used for both.
250 MW 2026 hyperscale runs 100 to 500 MW. Campuses are now planned in gigawatts.
0.45 L/kWh US average 2023. Legacy evaporative 1.8, best in class 0.20, closed loop near zero.
3.5 homes per acre Typical US suburban subdivision, gross of roads and setbacks.

Held fixed: PUE 1.20, load factor 0.90, 2.5 people per household, 10,364 kWh per household per year (EIA 2024), 300 gallons per household per day (EPA).

Electrical load

724x

Water drawn

7.3x

Equals, by electricity

228,000

Residents housed

0 vs 875

Nobody lives in the first one. That is the land-use trade being made.

The full numbers

MeasureData centerNeighborhoodRatio

What the comparison shows

At the default settings the data center draws roughly seven hundred times the electrical load of the housing that would fit on the same ground, and somewhere between three and thirty times the water depending entirely on how it is cooled. Those two facts are not symmetrical, and the asymmetry is the point.

Water is a choice. The slider spans a factor of ninety because the industry actually spans a factor of ninety: legacy evaporative cooling near 1.8 liters per kilowatt hour, the current American average near 0.45, Meta reporting 0.20 on recent buildings, and closed-loop and air-cooled designs approaching zero. New builds are migrating fast, largely because the political cost of visible water use now exceeds the energy penalty of not using it. Set the slider to 0.02 and the data center uses less water than the houses.

Power is not a choice. There is no design that removes the load, because the load is the product. A site can improve its PUE from 1.5 to 1.1 and save a fifth of the overhead, and the IT draw underneath it does not move at all. Move the cooling slider all the way down and the electricity ratio does not change by a single percent.

Which means the public argument is mostly aimed at the smaller of the two numbers. Water is local, visible, and engineerable. Power is systemic, invisible at the fence line, and structural. A community that negotiates hard on gallons and waves through megawatts has won the argument it could see.

Formulas

facility load MW  =  IT load × PUE
annual energy kWh  =  facility MW × 1000 × 8760 × load factor
water gal/day  =  annual kWh × WUE ÷ 3.785 ÷ 365

homes  =  acres × density
neighborhood MW  =  homes × 10,364 ÷ 8760 ÷ 1000
neighborhood gal/day  =  homes × 300

Sources and limits

Footprint anchor: Microsoft operates a data center of about one million square feet on roughly eighty-five acres outside Atlanta, and has built sites as large as three hundred and fifteen acres. Hyperscale is conventionally defined from forty megawatts upward; recent transactions include three hundred megawatts of critical IT load, and campuses are now planned in gigawatts. Household electricity is the EIA figure of 10,364 kWh a year for 2024. Household water is the EPA figure of more than three hundred gallons a day for the average family. Water use effectiveness figures are as reported by operators and by industry surveys.

Two limits worth stating. Reported water use effectiveness is self-reported and covers on-site consumption only, so it excludes the water consumed off site in generating the electricity; on a thermoelectric grid that indirect figure can exceed the direct one, which means both columns here understate water, and the data center column understates it more. And the housing comparison assumes new suburban development at the stated density; at townhouse or apartment densities the neighborhood column rises sharply and every ratio falls, which is why the density slider is there rather than baked in.