A company-wide water figure, offset by credits purchased in a different watershed, tells a chief information officer almost nothing about whether the building running their workload can keep operating.

Replenishment credits are doing the same work in water reporting that renewable energy certificates have done in carbon reporting for a decade, and the comparison is not flattering. A hyperscaler that consumes 8.1 billion gallons across its facilities in a year, and funds watershed projects that return roughly 4.5 billion gallons somewhere else, publishes a figure that reads as 64 percent replenished and climbing toward a 120 percent target by 2030. That accounting is internally consistent. It is also close to useless for the one question a technology buyer should ask, which is whether the specific building running their workload will still have water when a drought year and a disclosure law arrive at the same time.

The distinction is not academic. Water, unlike electricity, does not move across a national market. A gallon returned to a river in the Pacific Northwest does not lower the draw on an aquifer in central Arizona. Christopher Mims made the point plainly in the Wall Street Journal on July 6, using Michigan as the example: a company can produce a report showing the same amount of water it consumed, but if a data center runs dry in a water-stressed county, an abundance of water in another state cannot help it. Carbon offsetting at least operates on a single shared atmosphere. Water offsetting does not have that physics on its side.

Every operator discloses differently, and the differences are the story

Line up the four largest operators and no two report on the same basis. Meta published water figures for the sites it owns, and left out the ones it leases or has under construction. Google reported for owned and leased sites, and left out facilities run by third parties. Microsoft reported a company total with no site-level breakdown. Amazon, until this spring, reported no aggregate total at all, offering instead a ratio of liters per kilowatt-hour, a measure of efficiency that a community sitting next to a growing campus cannot convert into gallons drawn from its own supply.

Amazon did begin publishing an annual total this year, and reported that its withdrawal fell about 2 percent from 2024 to 2025 at the sites it owns and operates, even as building count grew. That is a real disclosure and worth acknowledging. It still does not answer the siting question, because a global withdrawal that ticks down while capacity climbs can coexist with a single county where the draw is rising fast enough to matter locally. The company average moves in one direction. The building next to the reservoir can move in the other.

This inconsistency is not an oversight that a better template would fix.

It is the product of a reporting regime designed around corporate goals rather than local exposure. A water-positive-by-2030 pledge is a portfolio target. It is met by netting consumption in stressed basins against replenishment in unstressed ones, and against efficiency gains averaged across a global fleet. The pledge can be entirely on track while the single facility a buyer depends on sits in exactly the county where the next moratorium bill gets written.

Cooling architecture is where the real exposure lives

Two facilities running identical chips can have opposite water profiles depending on how they reject heat. Evaporative cooling is cheap on power and hungry on water, because it throws heat away by turning water into vapor that leaves the site for good. Closed-loop and direct-to-chip liquid cooling recirculate the same fluid and consume far less, at a higher energy cost. Nvidia has said its latest rack designs solve the water problem with a closed-loop system, and Microsoft has committed to closed-loop tech for new builds starting in 2027. Those are the right direction. They are also years out, and they do not retrofit the evaporatively cooled campuses already drawing from stressed basins today.

For a buyer, the cooling choice is a proxy the sustainability report will not hand over. A campus that leans on evaporative cooling in a low-water region carries a regulatory and continuity exposure that a closed-loop campus in the same region does not. Neither number shows up in a water-positive percentage. Both show up in a permit hearing.

State legislatures have already found the gap. Utah now requires water use disclosure and reporting. Minnesota requires non-potable water where feasible. At least 16 states carried active water-related data center bills into their 2026 sessions, and the trend line points at facility-level reporting, not corporate averages. When those rules land on a specific building, the company-wide net figure a provider published becomes irrelevant to the county deciding whether to renew a withdrawal permit.

The buyer's problem here is continuity, not conscience. A hyperscaler that concentrated capacity in a water-stressed region to chase cheap land and cheap power made a bet that the water would keep flowing and the disclosure rules would stay loose. Both halves of that bet are now under pressure. The technology leader who signed a multi-year capacity commitment inherited the bet without pricing it, because the number they were shown was designed to keep it invisible.