Every viral post says AI is drinking our water. A bottle for every email. Whole towns going dry to cool a chatbot. The worry is understandable. The math behind it mostly is not.
And this one matters to everyone, not just the data-center crowd. So let us check the hype.
REALITY CHECK
The Claim
The new AI data centers going up everywhere will drain the local water. That is the fear behind most of the protests. Every new build, and every prompt running through it, adds to the thirst on the community's taps.
The Data
Start with the chips, because they changed the math. Nvidia's top AI racks, the GB200 and GB300, now draw around 120 kilowatts each. That is more heat than air can carry away. So they ship with direct liquid cooling built in, a sealed loop that runs coolant across the chips, filled once and recirculated. Nothing evaporates away like an old cooling tower. The market is following the hardware. Liquid cooling roughly doubled in 2025 to about 3 billion dollars, per Dell'Oro Group.
On top of that, the big operators are designing water out of new sites on purpose. Microsoft says every new data center design since August 2024 uses zero water for cooling, avoiding more than 125 million liters a year at each site. Oracle says its new AI sites use closed-loop cooling with no evaporation and effectively zero community water use. Meta's new one-gigawatt AI design, the first live later this year, uses a closed loop with dry coolers.
Be precise about what that means. These are the designs for new builds, and the first waterless Microsoft sites come online in late 2027. This is pledged and building, not finished everywhere. But the direction is clear, and it points down.
The Executive Verdict
The panic is aimed at the wrong cooling. The chips forced a change that happens to cut the water, and operators are now designing it out on top. The honest catch is timing and the old fleet. Judge a new build by how it is actually cooled, not by a headline about a plant from five years ago.
What They're Saying
The people building these sites are blunt about where the design is going.
“There is no evaporation, blowdown, or continuous makeup water requirement. Ongoing community water usage for cooling the data center is effectively zero.” Travis Grizzel, cloud infrastructure architect at Oracle (February 2026).
“Our next-generation datacenters consume zero water for cooling.” Steve Solomon, VP of Datacenter Infrastructure Engineering at Microsoft (December 2024). The loop is filled once at construction and recirculated.
Even the researcher whose work started the alarm is not saying stop. Shaolei Ren of UC Riverside put it plainly. Do not water your lawn at noon, and do not run your thirstiest AI at noon either. Run it smarter, at the right time and place.
What This Means for You
If you sit on a board, run a town, or just vote where you live, the useful move is not to feel guilty about your email. It is to ask two questions about any data center near you. How will it be cooled, and where will its power come from. A sealed liquid loop on clean power is a different animal than an evaporative tower on the town's drinking water.
Second, push for disclosure. Most operators are not required to report their water use, which is why the public numbers are estimates. The fix for a hidden number is to make it public, not to fear the technology.
After the data, where do you land on AI and water?
Signal vs. Noise
Three things worth your attention:
Water. Tucson said no, and the builder switched to air.
In August 2025 the Tucson city council rejected a large data center campus on city water, 7 to 0. The developer came back with an air-cooled plan. Mark's take: the water fights are winnable, and losing one teaches the whole industry to design the water out. Expect more air and closed-loop, faster.
Cooling. The chips are forcing the change anyway.
Nvidia's newest AI racks run too hot for air, so they ship with sealed liquid cooling by default. Mark's take: the move to low-water cooling is not a favor to anyone. It is physics. That is exactly why it will stick.
The outlier. xAI's Memphis site still runs on drinking water.
Colossus uses about 3 million gallons a day, most of it evaporated, with a recycled-water plant delayed into 2027. Mark's take: the real risk is a builder still cooling the old way in a city already short on water. Watch the specific site, not the whole sector.
Number of the Week
120 kilowatts. The heat from a single one of Nvidia's top AI server racks, too much for air to cool. It is why the newest AI data centers ship with sealed liquid loops that barely touch fresh water.
Source: Nvidia, 2024.
One Question Before You Go
What AI or cybersecurity claim do you want checked next? Reply and tell me. The next Hype Check might be yours.
Sources
Google, “Measuring the Environmental Impact of AI Inference” (2025)
Lawrence Berkeley National Laboratory, “United States Data Center Energy Usage Report” (2024)
Microsoft, “Sustainable by Design: Datacenters That Consume Zero Water for Cooling” (2024)
Oracle, “Closed-Loop Cooling in Oracle AI Data Centers” (2026)
Nvidia, GB200 NVL72 liquid-cooling design (2024); Dell'Oro Group, data center liquid cooling market (2026)
Mark Lynd is a 5x CEO, CIO, and CISO, 4x book author, and keynote speaker on AI and cybersecurity. He advises C-suites and executive teams on building efficient and resilient organizations. Learn more and book him at marklynd.com. He also publishes Hype Check Now, a newsletter on signal over noise in AI and cybersecurity, read by thousands of technology and security leaders worldwide.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.