Starmind and Water: How Much Cooling Water Gets Saved?

Why Ground Data Centers Drink So Much Water

Servers turn nearly all their electricity into heat. The dominant way to reject that heat cheaply is evaporative cooling — literally boiling water into the sky. The industry measures this with Water Usage Effectiveness (WUE): liters of water consumed per kilowatt-hour of IT energy.

The 2026 landscape: the legacy industry average sits around 1.8 L/kWh; Amazon's June 2026 disclosure cited a current industry average of 0.84 L/kWh, with its own fleet at 0.12–0.19; Microsoft reports 0.30. Peer-reviewed analysis of hyperscaler filings puts the weighted average at 0.59 L/kWh.

Those small numbers hide huge volumes. A single Google hyperscale campus averages 550,000 gallons per day; Google's Council Bluffs, Iowa site alone consumed 1 billion gallons in 2024. U.S. data center water consumption reached nearly one trillion liters per year by 2025, driven by AI expansion.

How Starmind Cools With Zero Water

In orbit, evaporative cooling is impossible and unnecessary. Each first-generation AI1 satellite pumps waste heat from its 120 kW compute core (150 kW peak) into a 110 m² deployable liquid radiator that faces the 3-kelvin void. Heat leaves as infrared radiation — governed by the Stefan–Boltzmann law, which rewards running radiators hot: doubling radiator temperature sheds heat 16x faster.

The coolant loop is closed and sealed at launch. After that, the satellite consumes exactly zero liters for its entire operational life. No cooling towers, no makeup water, no blowdown, no aquifer — the vacuum is the cooling tower.

The Math: Calculating Starmind's Water Savings

Here's the calculation the rest of the internet skips. Full-scale Starmind: 1 million satellites × 120 kW average compute = 120 GW of continuous IT capacity. Running year-round, that's 120 GW × 8,760 hours = ~1,051 TWh of IT energy per year — energy that would otherwise be processed (and cooled) on the ground.

Multiply by the WUE a ground data center would have paid, and you get annual direct cooling water avoided:

The base case — 880 billion liters per year — equals roughly 350,000 Olympic swimming pools, or the annual household water use of about 6.5 million Americans. Put differently: full-scale Starmind avoids approximately the entire current direct water footprint of the U.S. data center industry, every year.

The milestones matter too. At SpaceX's projected buildout rate of 100 GW of new orbital compute per year, each single year of Starmind deployment permanently displaces ~735 billion liters of future annual ground water demand (base case). And the very first tranche — 10,000 satellites, about 1.2 GW — already offsets ~8.8 billion liters a year, roughly Amazon's entire 2025 global data center water consumption.

The Hidden Multiplier: Indirect Water

Direct cooling is only half the story. Power plants consume water too — thermoelectric generation (gas, coal, nuclear) uses roughly 0.5–3 liters per kWh generated. A federal analysis pegged U.S. data centers' indirect footprint at 211 billion gallons in 2023 against 176 TWh consumed — about 1.2 gallons (4.5 liters) per kWh.

Starmind draws its power from onboard solar, so it skips this entirely. Applied to 1,051 TWh/year, the avoided indirect water is on the order of 4–5 trillion liters annually — several times larger than the direct savings, and conservative, since a ground facility's grid draw exceeds its IT load (PUE of 1.2–1.4) while grids are only partially decarbonized.

The Honest Counterargument: Earth Is Getting Less Thirsty

A credible savings estimate must account for where ground cooling is headed. Three trends compress the gap:

• Zero-water designs. Since August 2024, all new Microsoft data centers use closed-loop, chip-level liquid cooling that consumes no fresh water after the initial fill — avoiding 125+ million liters per site per year, with the first sites online in late 2027.
• Liquid cooling adoption. Direct-to-chip systems cut direct water use by 70–90% across retrofits, and hyperscaler WUE has fallen 39–80% over the past decade.
• The trade-off nobody mentions: zero-water cooling raises electricity draw (chillers replace evaporation), which raises indirect water. Ground engineering can move water consumption from the cooling tower to the power plant — it cannot delete it. Orbit deletes both.

Bottom line for the model: against 2030's best ground fleets, Starmind's direct savings shrink toward the 0.19–0.30 WUE scenarios. Its indirect savings — the trillions of liters embedded in electricity — remain untouched by any terrestrial cooling innovation.

What Space Doesn't Save

• Chip manufacturing water. Semiconductor fabs consume massive ultrapure water regardless of where the chip ends up — Starmind's satellites are built (and their D3 chips fabbed at Terafab) on Earth, in Texas.
• Launch operations. Starship pads use water deluge systems and methane production has a water footprint. Per satellite-year of operation, this is small — but it isn't zero.
• Ground stations. Downlink infrastructure remains terrestrial, though its compute (and cooling) load is trivial next to a data center campus.

FAQ

Bottom Line

“How much cooling water does Starmind save?” finally has a number: on the order of 880 billion liters per year of direct cooling water at full scale — the size of America's entire current data center water bill — and multiples of that once you count the water hidden inside grid electricity. Even under the most generous assumptions about terrestrial cooling progress, moving AI compute to orbit is the only architecture that takes water out of the equation completely.

The variables to watch: Starship's launch cadence (which sets the deployment rate) and ground zero-water adoption (which sets the baseline). We track both — see our companion pieces on why SpaceX is leaving Earth's grid behind and who builds Starmind's hardware for the full picture.