How Gemini recycles its cooling water

When a query reaches Google’s infrastructure, it is processed in a data centre that may be running a fundamentally different cooling system to the evaporative towers covered in the rest of this series. Rather than spraying water into open air and losing it to evaporation, Google has deployed closed-loop cooling across a growing share of its fleet, a system where the same water circulates repeatedly, absorbing heat and then releasing it to the outside air through a sealed heat exchanger, before returning to the servers to do it again.

The distinction matters enormously. In a conventional open-loop evaporative cooling tower, 70–80% of the water drawn evaporates into the atmosphere with every pass^[1]FWPCOAView source →. It is gone. The closed-loop alternative loses almost none, the water is sealed inside the system, and only a small top-up is needed for minor leakage. That difference, between an open pipe and a sealed circuit, is arguably the single largest lever available to reduce AI’s operational water footprint.

The mechanics of cycling

In a conventional cooling tower, the number of times water circulates before being discharged is called the “cycles of concentration.” As water evaporates, the dissolved minerals, silica and salts it carried become increasingly concentrated. Left unchecked, they form scale on pipe walls and corrode equipment, so operators periodically drain a portion, the “blowdown”, and refill it with fresh water. Standard towers run at 3–10 cycles before blowdown is required^[2]Li, Yang, Islam, RenView source →, depending on source water quality.

Google’s approach is to push that number higher through chemical treatment and water-quality monitoring, stretching the useful life of each litre before it is retired. In a true closed-loop system, sealed direct-to-chip liquid cooling, the concept of cycles largely disappears: the same fluid circulates indefinitely, never exposed to open air. Microsoft’s newest data centres use exactly this design, filling the system once during construction^[3]TechRadarView source → and circulating water between servers and chillers with no fresh supply thereafter, avoiding more than 125 million litres a year^[3]TechRadarView source → versus an evaporative equivalent.

Where Google actually stands

Google’s position is more nuanced than a simple “closed-loop” story. As of 2025, roughly two-thirds of its data centres still use evaporative cooling^[4]AxiosView source →; the rest use air cooling or recycled, non-conventional water. Google’s argument, a legitimate one, is that evaporative cooling uses significantly less electricity than air cooling, producing fewer emissions where the grid is clean but water isn’t scarce. The optimum depends on local conditions: what matters more here, electricity or freshwater?

Google has committed to assessing watershed health before each build and defaulting to air cooling or recycled water in water-stressed regions^[5]GoogleView source →. In Douglas County, Georgia, it cools a campus with treated municipal wastewater that would otherwise be discharged into the Chattahoochee River. In India, a new facility uses air cooling. In Chile, a $200 million data centre was paused after an environmental court ruling, and the location-first lesson appears to have landed.

What Google has not done is commit to a universal closed-loop standard across its fleet, a step Microsoft has taken for new builds. Instead its strategy is to replenish more than it consumes: more than 7 billion gallons replenished in 2025^[6]9to5GoogleView source → through 165 stewardship projects across 97 watersheds, targeting 120% replenishment by 2030.

Why this is the biggest lever

Water efficiency is not, ultimately, about water. It is about infrastructure decisions made at the point of construction. Unlike carbon, which can be offset or reduced through grid decarbonisation after a facility is built, WUE is almost entirely locked in by the cooling system chosen during initial design^[11]EquinixView source →. A data centre built with evaporative towers carries a high WUE for its entire 20–30 year life. A closed-loop or air-cooled facility starts near zero and stays there.

The right question isn’t “will you offset your water use,” but “will you need to use it in the first place.”

When a hyperscaler announces it will be “water positive by 2030” through replenishment, it is mostly committing to fund watershed restoration elsewhere, not to stop withdrawing at source. Replenishment does not return water to the specific aquifer, river or community drawn upon. The Querétaro data centre and the Querétaro community draw from the same regional water table; planting trees in Nebraska does not refill it.

The closed-loop question is therefore the right one to ask of any new AI infrastructure: not whether the water will be offset, but whether it needs to be used at all.