ANTM

AI Data Center Cooling: Water, Energy and What a Zero-Water Claim Actually Means

Cooling decides how much water and electricity a data center uses. Reading Microsoft's and Google's own pages shows what a zero-water claim covers, what it trades away and what the sources do not say.

Editorial desk

Published 5 min read

A bundle of fine teal lines sweeping in from the left into a grid of teal squares with glowing nodes, with a red line and three small coral droplets to the right, on a near-black background
Illustration generated with AI (FLUX.1 [schnell] (Black Forest Labs) via Cloudflare Workers AI, Apache 2.0). Prompt and direction by ANTM.

Every watt a data center draws ends up as heat, and something has to carry that heat away. The choice of cooling system decides how much electricity the building uses on top of its servers, how much water it consumes, and how much of that water is used on site or at a power plant. As AI racks get denser, operators are moving to liquid cooling, and some now advertise "zero-water" designs. This article explains the main cooling approaches, what Microsoft and Google say about the trade-offs, and how to read a zero-water claim. For the electricity side of the picture, see what US official data shows about data center electricity use; for the hardware, see GPU vs TPU accelerators compared.

Why data centers need cooling at all

Microsoft puts it simply: servers produce heat and "must stay at the right temperature, which requires cooling."[3] Google says the same in its own words, that data centers generate heat and need to be cooled to operate, and describes its decisions as "balancing the trade-offs between using energy and water for cooling."[5]

That trade-off is the whole subject. Water can carry heat away cheaply because evaporating it absorbs a lot of energy, so evaporative cooling needs less electricity. Removing the water means replacing evaporation with mechanical cooling, which draws power. Neither company claims a free option.

The main approaches, as the sources describe them

ApproachHow it works, per the sourceWaterEnergy note from the sourceSource
Outside-air cooling with evaporative assistOutside air cools the building; water is added only when it is hot. Microsoft says this has been used across its fleet since 2008 and uses water only when outside temperatures exceed 85°F (29.4°C)Used only in hot periods; water cycles 2 to 5 times and a portion evaporatesMicrosoft says outside air alone can cool many data centers below that temperatureMicrosoft [2][3]
Cooling towersEvaporate water through the year, according to Microsoft's definitionsEvaporative, year-roundNot quantified in the pages we readMicrosoft [2]
Water-cooled, using sourced or recycled waterGoogle says options include air cooling, water cooling, refrigerants or a combination, chosen per siteVaries; Google reports reclaimed or non-potable water at more than 25% of campuses (2022)Google: water-cooled data centers use about 10% less energy than many air-cooled onesGoogle [4]
Closed-loop, chip-level liquid coolingWater recirculates through a closed loop with no evaporation; Microsoft's 2024 AI designFilled once; Microsoft says zero water consumed for cooling during operationMicrosoft says PUE rises, offset in part by warmer water temperatures and economizing chillers; it calls the energy increase nominal and gives no figureMicrosoft [1]

Two details are worth keeping. First, Google's 10% figure compares water-cooled sites with "many" air-cooled ones in a 2022 post, not a measured fleet average today.[4] Second, Microsoft's page for the 2024 design uses the term "chip-level cooling," not direct-to-chip; the June 2026 post describes the same design as recirculating, closed-loop, direct-to-chip cooling.[1][2]

What the numbers say, and what they do not

Microsoft's June 2026 post says its average water usage effectiveness (WUE) fell from 2.3 to 0.27 litres per kilowatt-hour in 2025, and that about 90% of its 2025 owned fleet operates on "low- to zero-water" cooling systems.[2] The post does not say how much of that is specifically closed-loop. The December 2024 post states a different figure, a "FY 2024 global average withdrawal WUE of 0.30 L/kWh", and uses it to estimate that the new design avoids more than 125 million litres of water per year per data center.[1] We do not know whether the 0.30 and 0.27 figures use the same boundaries, so we do not compare them directly. The 125 million litre figure is an estimate that multiplies a fleet average by the data center's load, not a measurement at one site.

Google's most recent per-site figures that we found are from 2021: an average data center used about 450,000 gallons per day, and the global fleet about 4.3 billion gallons.[4] They are five years old. Google's current water page gives no consumption totals; it says Google reports annual water use for its data center locations elsewhere, which we did not open, and it describes a goal to replenish more freshwater than it consumes across data centers and offices by 2030, which is a target, not a result.[5]

How AI hardware changes the picture

NVIDIA describes the GB200 NVL72 as connecting 36 Grace CPUs and 72 Blackwell GPUs in "a rack-scale, liquid-cooled design," and says liquid cooling increases compute density and reduces floor space.[6] The same page says liquid-cooled racks reduce water consumption and carbon footprint, but gives no figure, and it states no rack power draw. Our reading is that dense racks are the reason liquid cooling is spreading, because moving that much heat with air alone requires far more space and airflow. That is an interpretation. The sources we opened do not quantify it.

Five checks for a zero-water claim

  1. What is the boundary? Microsoft's claim covers "cooling during operations." Water is still used for restrooms and kitchens, and the loop needs an initial fill.[1]
  2. Is it a design or a fleet? The 2024 design applies to new builds. Microsoft says its pilots in Phoenix and Mt. Pleasant, Wisconsin come in 2026, new sites online from late 2027, and that its existing fleet still mixes air-cooled and water-cooled systems.[1]
  3. What happened to the energy? Microsoft says PUE increases. Ask for the PUE and the annual energy change, not only the water saved.
  4. Where does the electricity come from? Power plants also use water. None of the pages we read quantifies this, so a site-level water figure may not capture the whole picture. Treat that as an open question.
  5. Is the number measured or modelled? Withdrawal versus consumption, fleet average versus site, and the year all change the figure. Check the footnote.

Limitations of this article

We read company pages only. We could not open the Lawrence Berkeley National Laboratory data center report, which the US Department of Energy cited in its December 2024 announcement, so no independent estimate of data center water use appears here. We ran no measurements. Microsoft's local water-use page carries no date, so we list the retrieval date, 1 November 2026. We will add a dated update when an independent source is opened and verified.

Frequently asked questions

Do data centers use water for cooling?
Some do. Microsoft says its older designs use evaporative cooling when outside temperatures exceed 85°F (29.4°C), and that water cycles through the system 2 to 5 times with a portion evaporating.[3] Its 2024 AI design uses a closed loop with no evaporation.[1]
What does zero-water cooling mean?
In Microsoft's wording it means zero water consumed for cooling during operation. Water is still used for restrooms and kitchens, the loop is filled initially, and the claim concerns the new design rather than the whole fleet.[1][2]
Is air cooling or water cooling more energy efficient?
It depends on the site. Google says water-cooled data centers use about 10% less energy than many air-cooled ones and that there is no one-size-fits-all solution.[4] Microsoft says replacing evaporative cooling with mechanical cooling increases its PUE.[1]
Why do AI chips need liquid cooling?
NVIDIA describes the GB200 NVL72 as a rack-scale, liquid-cooled design with 72 GPUs in one NVLink domain and says liquid cooling increases compute density. The page we read gives no rack power figure.[6]

The ANTM newsletter

The signal, not the noise.

Sourced AI coverage in your inbox. Double opt-in, unsubscribe in one click.

Referenced sources

  1. 1.
    Sustainable by design: Next-generation datacenters consume zero water for cooling(opens in a new tab)

    Microsoft Cloud Blog (Steve Solomon)CompanyPublished Dec 9, 2024Accessed Oct 3, 2026

  2. 2.
    Inside Microsoft's two-decade push to cut water intensity while scaling for growth(opens in a new tab)

    The Official Microsoft Blog (Judy Priest, Steve Solomon)CompanyPublished Jun 24, 2026Accessed Oct 3, 2026

  3. 3.
    Understanding water use at Microsoft datacenters(opens in a new tab)

    Microsoft LocalCompanyPublished Nov 1, 2026Accessed Oct 3, 2026

  4. 4.
    Our commitment to climate-conscious data center cooling(opens in a new tab)

    Google (Urs Hölzle)CompanyPublished Nov 21, 2022Accessed Oct 3, 2026

  5. 5.
    Advancing responsible water use at our data centers(opens in a new tab)

    GoogleCompanyPublished Nov 1, 2026Accessed Oct 3, 2026

  6. 6.
    NVIDIA GB200 NVL72(opens in a new tab)

    NVIDIACompanyPublished Nov 1, 2026Accessed Oct 3, 2026

ANTM Editorial

Editorial desk

The editorial desk at AI's Next Top Model. Every article is sourced to primary documents and approved by an editor before publication. See the editorial policy for how we work.