Inside Microsoft's two-decade push to cut water intensity while scaling for growth

What this is

Microsoft's own account of how its datacenter water intensity fell, published 2026-06-24. It sits behind the metric page: same company, same programme, but this is the narrative with the mechanism in it. Read it for the design, not for the numbers — the numbers are better sourced from the efficiency page.

Key claims

• WUE fell from 2.3 L/kWh to 0.27 L/kWh by 2025 — "nearly 90%" better than the first generation of designs. The 0.27 endpoint reconciles with the FY25 global WUE on the metrics page.
• 2030 target: a 40% improvement in datacenter water-use intensity, against a 2022 baseline; 25% achieved as of 2025. That is a little over half the target delivered in three of the eight years.
• FY25 was Microsoft's first water-positive fiscal year — "replenishing more water than it withdrew" across global operations. Replenishment is an offset accounting concept, not a reduction in withdrawal at any given site.
• Alternative (recycled / reused / non-potable) water share by site: Quincy, WA 74% · Singapore 99% · San Antonio, TX 79%.

The mechanism — the durable part

The 2024 datacenter design uses a chip-level closed-loop, direct-to-chip cooling system: water is recirculated through cold plates for "precise zonal temperature control without water evaporation," so the design "consumes zero water for cooling during operations."

The physics is what makes this durable while the percentages date. Evaporative cooling rejects heat by vaporising water — cheap in energy, expensive in water, and it is the reason a hot-climate site can be the water-constrained one. A closed loop rejects the same heat to air, via dry coolers or chillers. The heat does not disappear; it moves onto the electrical side of the ledger. So a zero-water design is a water-for-power substitution, and the interesting quantity is the exchange rate.

Microsoft does not state that exchange rate anywhere on this page. No power figure, no PUE delta, no kW-per-MW-of-IT penalty for the closed-loop design. That absence is worth carrying: it means the zero-water claim cannot be netted against the facility-overhead number, and it is consistent with the efficiency page's global PUE ticking 1.16 → 1.17 across the same period.

Why it matters here

• Water is a siting constraint the power calls assume away. A campus can clear an interconnect queue and still fail on a water permit; the closed-loop design is the lever that removes water from the siting equation — at an unstated power cost, in a market where power is the binding constraint.
• No live call rests on water. This is on the rung for the mechanism and for brand-side credibility, not to move a number under a published position. Recorded as such rather than dressed up.

Limitations

Corporate blog: selected numbers, no methodology, no scope boundary stated (unlike the metrics page, which at least names its fleet filter). "Water positive" is replenishment accounting and depends on a separate FY24 Environmental Data Fact Sheet not read here. The site-level alternative-water shares are three chosen sites out of a fleet of many. The zero-water design is described as the 2024 design — nothing on the page says what share of the operating fleet uses it, which is the number that would make it material.

Source: Inside Microsoft's two-decade push to cut water intensity while scaling for growth, Microsoft, 2026-06-24