Nvidia Says AI Data Centres Can Run on Almost No Water, but There Is a Catch

Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters.

Nvidia reckons it has answered the loudest complaint aimed at AI data centres, which is how much water they swallow. The fix is real, the marketing is generous, and the bigger water bill sits somewhere the new design never touches.

Let's get into it.

ChatGPT is insanely powerful.

But most people waste 90% of its potential by using it like Google.

These 1,000+ proven ChatGPT prompts fix that and help you work 10X faster.

Sign up for Superhuman AI and get:

Claim your free prompts

Water has become the most effective weapon for anyone trying to block a new data centre. Residents near proposed sites in Arizona, Georgia and Spain have turned cooling water into a planning fight, and the figures give them plenty to work with.

The United Nations warned in June 2026 that AI-related water use could match the yearly needs of 1.3 billion people by the end of the decade, and a Pew Research Center survey found that most Americans who know what a data centre is already consider it bad for the local environment.

So when Nvidia published a new cooling design on 21 June 2026 and said it could take cooling water close to zero, it was aiming straight at the objection that stalls projects fastest.

The Hot Tub Trick

The design sits inside Nvidia's Vera Rubin DSX reference architecture, the blueprint the company first unveiled on 16 March 2026 and expanded with new cooling software on 1 June 2026. The whole design turns on running everything hotter than the industry used to think was safe.

Traditional servers need cold air and cold water to stay alive, with most sites settling around 30 degrees Celsius. Nvidia's newest systems run their coolant at up to 45 degrees Celsius, which is 113 degrees Fahrenheit, hotter than the water in a domestic hot tub. The coolant itself is roughly three-quarters water and one-quarter propylene glycol, the same chemistry that fills a car radiator. It gets poured in once and then circulates in a sealed loop for the life of the building, so nothing evaporates and nothing needs topping up.

Because the liquid carries heat away while staying hot, the building can shed that heat through outdoor radiators known as dry coolers rather than the evaporative cooling towers that drink millions of gallons.

Ali Heydari, Nvidia's director of data centre cooling and infrastructure, put it plainly when he said the design removes almost all of the power and water that older sites burn through.

The Numbers Behind the Claim

The savings Nvidia quotes are large enough to explain the attention. A conventional cooling-tower setup uses roughly 2.6 million gallons of water per megawatt every year, and the closed loop can bring that figure down to almost nothing in the right climate, which is where the up to one hundred per cent headline comes from.

The company estimates that a fifty-megawatt site could save more than four million dollars a year once energy and water are added together. Some sustainability analysts go further and describe water-use efficiency improving by more than three hundred times against older air-cooled halls. On paper, a heavy local water user turns into something close to water neutral for cooling.

Where the Claim Gets Shaky

The phrase doing the heavy lifting in all of this is up to. A data centre in the Scottish Highlands and one in Phoenix, Arizona face very different weather, and Nvidia admits as much.

In hot regions the outside air sometimes climbs past the point where dry coolers can cope on their own, so backup chillers still switch on for a handful of days each year, and those chillers still want water. The company also published no price for the equipment, and a spokesperson said suppliers will set their own, which leaves smaller operators guessing whether the maths works for them at all.

Andrew Chien, a computer science professor at the University of Chicago who runs the CERES Center for Unstoppable Computing, welcomed the move while calling truly zero water use unrealistic. He said the real prize is pushing the input temperature up to 45 degrees Celsius, because warmer coolant lets a site vent heat to the outdoors without running air conditioning, though he warned that the systems are expensive.

Nvidia's chief sustainability officer Josh Parker went the other way and told an audience before London Climate Week that the water problem for data centres is largely solved, which is a generous reading of a design that still depends on the right postcode.

The Third That Actually Gets Fixed

This is the part the headline quietly skips over. Cooling the chips on site accounts for only about a quarter to a third of the water an AI system consumes across its whole life.

The larger share is hidden upstream in the power plants that feed the building. A natural gas plant uses roughly 1.17 litres of water for every kilowatt-hour it generates, and a coal plant closer to 2.2 litres, so a data centre running a bone-dry cooling loop on a fossil-fuel grid is still soaking up water somewhere else.

More water disappears further back still, in the fabrication of the chips themselves, where wafer cleaning and chemical processing run on enormous volumes that no cooling design can recover. Removing the water inside the walls is real progress, and it leaves most of the footprint exactly where it was.

Why It Still Matters Anyway

None of this makes the announcement empty. Microsoft said in 2024 that its newest data centres would stop using water for cooling and save more than 125 million litres a year at each site, so Nvidia is pushing an industry shift along rather than inventing one, and its scale means the design will reach far more builders than any single operator could.

Cheaper cooling also feeds straight into cheaper AI, because every watt and every gallon a company strips out of its infrastructure is a cost it can take out of the price it charges for tokens. The honest framing is the most useful one. Nvidia has solved a real and visible problem, the local water fight, while the harder problem of the grid behind it stays precisely where it was.

The Bottom Line

Nvidia's hot-water loop is a genuine engineering win, and for any site in a cool climate it can turn a major water drain into almost nothing. But the up to one hundred per cent number describes a best case rather than a promise, and it only ever covered the third of AI's water that sits inside the building.

The other two-thirds lives in the power station and the chip factory, and no clever coolant touches that. Treat this as a real step that the marketing has dressed two sizes too big.

Category: Claims Analysis

"Take this sustainability or product claim [paste the claim] and pull it apart for me. Identify the exact metric being measured, the conditions required for the best-case number to hold, what the claim quietly leaves out of scope, and the single question I should ask the vendor to expose the gap. Keep it to one short paragraph per point."

Every big AI announcement now arrives with a number that sounds final and a footnote that quietly shrinks it. The skill worth building is reading the footnote first, because that is usually where the real story sits. Nvidia did something genuinely clever here, and it still pays to ask which third of the problem the clever bit actually touches. Hit reply, I read every response.

See you in the next one.

— Vivek

P.S. Know someone who argues about data centres and water at dinner parties? Forward this to the most opinionated environmentalist you know. They can subscribe at https://savvymonk.beehiiv.com/