Microsoft reveals zero-water cooling for AI datacenters

AI chips run blistering hot, and datacenters have historically guzzled massive amounts of water to cool them down. In water-stressed areas, that has made big tech's server expansion a major headache for local communities.

Microsoft wants to change that narrative. The company just revealed its next-generation datacenter architecture designed specifically for heavy AI workloads, claiming zero water consumption for cooling. Across a single facility, the new setup will avoid burning through more than 125 million liters of water every year.

Trapping heat in a closed loop

Traditional datacenters often rely on evaporative cooling to save electricity. It works like sweating: heat dissipates as water evaporates away, but fresh water supplies get drained in the process.

Microsoft's new design ditches open evaporation in favor of chip-level liquid cooling inside a sealed, closed loop.

Water is filled into the system just once during construction. From there, it endlessly circulates between the chips and mechanical chillers to carry away heat, requiring zero fresh water top-ups. Aside from routine administrative use like sinks and restrooms, the cooling cycle itself evaporates nothing.

That directly attacks Water Usage Effectiveness (WUE), the industry metric for water efficiency. In fiscal year 2024, Microsoft logged an average WUE of 0.30 L/kWh, a 39% drop from 2021 and down 80% since its first-generation facilities. For sites using the zero-water setup, cooling WUE is expected to plunge near zero.

There is, of course, an engineering trade-off. Swapping evaporative cooling for mechanical chillers usually drives power usage up. Microsoft says it mitigates the energy penalty by letting chip-level solutions run at warmer temperatures, pairing them with high-efficiency economizing chillers. The company expects only a "nominal increase" in fleetwide annual energy consumption.

Turning standard now, online by late 2027

This is not a far-off research demo. Starting in August 2024, Microsoft made this next-generation design the default blueprint for every new datacenter it plans to build.

Actual rollouts will roll out in phases. Pilot projects will fire up in 2026 across Phoenix, Arizona, and Mt. Pleasant, Wisconsin. Fully operational sites built from scratch under the new standard are slated to come online in late 2027.

For existing sites, Microsoft continues mixing air and water cooling while sourcing reclaimed water in drought-prone regions like Texas, Washington, California, and Singapore.

Why it matters

The AI infrastructure race is no longer just about buying GPUs; it is an outright battle for local resources. Datacenters routinely draw fierce public pushback over local watershed depletion, especially in desert hubs. Microsoft's pledge to eliminate cooling water evaporation gives it a massive PR and zoning advantage when applying for permits.

We think shifting straight to closed-loop liquid cooling at the chip level was inevitable anyway. Future AI silicon packs thermal loads that open-air and basic cooling setups simply cannot handle. Committing to zero evaporation now forces hardware supply chains to adapt early.

The real test will be power. In an era where power grids are already squeezed to their limits by AI, trading water savings for even a slight bump in electricity consumption carries real risks. How cleanly Microsoft manages that delicate trade-off will likely dictate whether rival hyperscalers are forced to copy the playbook.