Hot topics

Closed-Loop Water Cooling

What a 'closed-loop' cooling system actually is, how often the water inside it really gets replaced, what Oracle claims its AI data centers use — and why a New Mexico congresswoman's own numbers, from the state's own water-rights records, don't remotely match Oracle's.

Last updated September 29, 2026

Terms

Closed-loop cooling
A cooling system where the same water recirculates through sealed piping indefinitely, rather than being evaporated and replaced continuously the way a cooling tower works.
Dry cooler
A radiator-style heat exchanger that rejects heat into outside air using fans, not evaporation — what makes a closed loop genuinely non-evaporative, as opposed to a closed loop that still dumps its heat into an outdoor cooling tower.
Makeup water
Water added to replace what a system has lost — to evaporation in an open system, or to small leaks and maintenance events in a closed one.
Acre-foot
The volume of water covering one acre to a depth of one foot — 325,851 gallons. The standard unit for water-rights accounting in the arid Southwest, used below.

How a closed loop actually works

Our page on how data centers use water covers the conventional approach: evaporative cooling towers, which lose up to 85% of the water passing through them to the air. A closed loop is the alternative: water absorbs heat from the servers, carries it to a heat exchanger, and returns to do it again — the same water, indefinitely, rather than a constant fresh supply.

The detail that actually determines whether a "closed loop" uses meaningful water is what happens at that heat exchanger. If the loop dumps its heat into a dry cooler — fans blowing outside air across a radiator, no water involved — the system can run close to zero ongoing water use, which is the design Oracle describes. If instead the loop rejects its heat into an outdoor cooling tower, that outer loop is evaporating water constantly even though the inner loop touching the servers is sealed. A widely shared critique of "closed-loop" marketing makes exactly this point: "the closed loop isn't closed folks. It never was" — because in many real installations, a genuinely closed inner loop is paired with a very much open outer one. Whether that critique applies to any specific facility depends entirely on how its heat exchanger is built, which isn't always disclosed.

How often the water actually gets replaced

A closed loop isn't filled once and never touched again. During construction, installers flush the new piping with water to clear out debris and verify the system, and that flush water is disposed of as wastewater — a one-time event, not an ongoing one. Once the loop is filled and sealed for normal operation, industry guidance calls for testing the water's chemistry — pH, corrosion inhibitors, biological growth — at least quarterly, topping off small losses as needed rather than on any fixed replacement schedule. A full system flush and refill isn't done on a set calendar interval either; operators base it on what water testing and equipment inspections show, since scaling, corrosion, or contamination can force an early flush regardless of how "closed" the design is. In other words: closed-loop systems need far less water than evaporative ones, but "closed" doesn't mean "never opened."

What Oracle actually claims

In a Feb. 9, 2026 post, Oracle described its newer AI data centers — including sites in New Mexico, Michigan, Wisconsin, and Texas — as using "direct-to-chip, closed-loop, non-evaporative" cooling: heat removed at the chip, carried by a sealed loop, and rejected to outside air with "no evaporation, blowdown, or continuous makeup water requirement." Its most detailed figures are for Project Jupiter, the Doña Ana County, New Mexico campus it's building with OpenAI. Oracle has published the same basic numbers, restated slightly differently, across three separate follow-up posts between April and August 2026:

Initial fill, cooling loop (4 buildings)
~10M gal
Initial fill, Bloom Energy fuel cells
~960K gal
Ongoing maintenance, per year
~168K gal

Oracle's own figures, as published across its Feb., Apr., Jun., and Aug. 2026 blog posts; see sources.

Oracle totals these as roughly 13.5 million gallons over 15 years, which it describes as equivalent to "about nine average U.S. households per year," and separately says day-to-day cooling requires no added water at all — annual top-offs of roughly 0 to 1,000 gallons. It also says none of this touches the local public drinking-water supply: the water is non-potable industrial well water, purchased under contract from an existing New Mexico water-rights holder.

Does the data actually line up?

Two things are worth checking before taking the 13.5-million- gallon figure at face value: what's actually inside it, and what it leaves out.

What's inside the number. Of Oracle's own 13.5-million-gallon, 15-year total, about 11 million gallons — roughly 81% of the whole figure — is the one-time initial fill for the cooling loop and fuel cells, not ongoing use. Strip that out and the actual steady-state rate is closer to 168,000 gallons a year, or about 460 gallons a day. Both numbers are Oracle's own; blending a one-time fill into a 15-year total makes the headline figure look larger and more precise than the "practically nothing, day to day" framing Oracle also uses in the same posts. Neither framing is false, but they tell different stories, and Oracle has generally led with whichever one sounds smaller in context.

Oracle's figure

~168,000 gal/yr

Ongoing maintenance water for Project Jupiter's closed cooling loop, once the one-time initial fill is excluded — Oracle's own engineering estimate for the cooling system specifically.

The state's figure

~782M gal/yr

What Rep. Melanie Stansbury says New Mexico's own water-rights documents show the site could draw — up to 2,400 acre-feet a year, which she says is enough to supply about 30,000 homes.

That's a gap of more than 4,000 times between the two numbers, and it isn't a case of one side simply being wrong — they're almost certainly measuring different things. Oracle's figure is an engineering estimate for what the closed cooling loop itself needs. The figure Stansbury cited comes from the site's water-rights paperwork, which by its nature covers everything the project is legally permitted to draw — potentially including the fuel-cell plant at scale, future phases, landscaping, and ordinary facility use, plus whatever legal headroom a water-rights application typically requests above expected use. Stansbury has also flagged something more basic: as of her statement, "no permits have been filed for the project's long-term operational water needs," and more than 30 separate water-rights protests tied to the site remain unresolved, with no public hearings held yet. Whatever the cooling loop itself uses, the project's total legal claim on local water — the number that actually determines what's available to farms and households nearby — is a question the engineering figures alone don't answer.

There's a second, quieter tradeoff worth naming: a genuinely non-evaporative, dry-cooled design like the one Oracle describes avoids water use, but published research on data center cooling finds evaporative-assisted systems typically use 10–35% less electricity at peak demand than purely dry cooling, because evaporation can reject heat even when the air outside is too warm for a dry radiator alone. A facility that eliminates water use this way is very likely using more electricity to do it — and that electricity, generated somewhere else on the grid, usually has its own water footprint, the same "indirect" water use our page on data center water use covers in more detail. "Zero water at the building" isn't automatically the same as "zero water, full stop."

Case study: Project Jupiter's four water posts

Feb. 9, 2026
Oracle publishes its general closed-loop cooling explainer, introducing the non-evaporative design across its newer AI data centers.
Apr. 27, 2026
First Project Jupiter–specific follow-up: "Water Facts," addressing sourcing and confirming the project won't touch the local public drinking-water supply.
Jun. 12, 2026
NPR reports on local water-scarcity concerns in Doña Ana County, where groundwater levels have been falling and farmers are digging deeper wells.
Jun. 16, 2026
Second follow-up, "The Facts About Project Jupiter's Water Usage," publishing the 13.5-million-gallon/15-year figure and the fuel-cell power-plan change.
Aug. 11, 2026
Third follow-up aimed directly at "what Southern New Mexico residents should know," as local opposition continues.
Sept. 2026
Rep. Melanie Stansbury publicizes the state's 2,400-acre-foot figure and the 30-plus pending water-rights protests; regional outlets report the dispute as unresolved.

Four public posts from one company on the same project's water use, inside about six months, is itself informative — it suggests Oracle's own messaging hasn't settled the concern locally, whatever the merits of its engineering claims.

Nonpartisan, plainly

There are two separate questions tangled together here, and they deserve different answers. Whether a dry-cooled, non- evaporative closed loop can genuinely use very little water day to day is a real engineering question with a real answer, and nothing we found contradicts Oracle's basic mechanism or its 168,000-gallon annual maintenance estimate for the cooling loop itself — though it comes with a documented electricity tradeoff Oracle's posts don't emphasize. Whether the public can be confident about the project's total water footprint is a different question, and on that one, a sitting member of Congress citing the state's own water-rights records has put a number on the table — 2,400 acre-feet a year — that is orders of magnitude larger than anything Oracle has published, alongside more than 30 unresolved water-rights protests and no permits yet filed for long-term operational water. We're not in a position to adjudicate New Mexico water law. We can say plainly that a company's own favorable framing of its own engineering numbers is not the same thing as an independently verified account of a project's total water demand, and right now, for Project Jupiter, only the first of those exists.

See something wrong at a data center near you?

Noise, unpermitted construction, water or air complaints, or anything else worth documenting — Erin Brockovich's team runs a crowdsourced tracker where residents can report data center issues directly.

Report a data center issue ↗

Talking points

These are the questions we think you should ask those who are running for office and will represent you.

  1. Should a data center operator be required to publish independently verified water-use figures, rather than self-reported estimates, before a permit is approved?
  2. Should a project's total permitted water rights be made public and easy to compare against the operator's own marketing claims, in plain language, before construction begins?
  3. Should pending water-rights protests have to be resolved before a large industrial water user like a data center is allowed to begin construction?
  4. Should the tradeoff between a facility's water use and its electricity use be disclosed together, rather than water savings being promoted without mentioning the added power demand?

Read more

WONKRZ STORE

Upgrade Your Gear Today

Discover our latest collection of products and exclusive deals. This store supports various initiatives to keep people informed and engaged in our democracy regardless of political party.

Shop Now →