Nvidia’s Hot Liquid Cooling Could Cut AI’s Water Bill, but Not Its Energy

Every prompt a user sends to an AI system draws on cooling water at a data center somewhere, and a United Nations University report published this month projects that, by 2030, the water footprint of the world’s AI data centers will equal the basic annual domestic needs of 1.3 billion people in Sub-Saharan Africa. Nvidia, the chipmaker whose GPUs power most of that build-out, says its newest Rubin generation can run its cooling liquid at 45 degrees Celsius, hot enough to operate without water-consuming chillers in many climates and, in the right conditions, to eliminate evaporative water cooling entirely. The technology is real and the savings are quantified, but the UN researchers behind the new figures warn that more efficient AI tends to invite more AI, and the wider climate ledger does not balance on its own.

The result is a crack in one of AI’s largest environmental costs, opened by a single engineering choice, sitting beside a structural problem that no cooling system can fix.

Why AI Data Centers Drink So Much Water

Cooling is the hidden line item in the AI bill. According to Nvidia, cooling has historically accounted for up to 40 percent of a data center’s electricity consumption, and traditional air-cooled facilities lean on chilled water and evaporation to keep processors within safe operating temperatures. A large data center can consume up to 5 million gallons of water per day for cooling, the UN University report found, and roughly 60 percent of total data center water use is indirect, embedded in the power plants that supply the electricity.

Those figures scale with the build-out. UNU researchers project global AI data center electricity demand will hit 945 terawatt-hours by 2030, nearly triple the combined annual electricity use of Pakistan, Bangladesh, and Nigeria, countries that together are home to more than 650 million people. The associated water footprint will reach 9.3 trillion litres, the report estimates, and the land footprint will exceed 14,500 square kilometres, roughly twice the Jakarta metropolitan area. The pressure is not evenly distributed. In Ireland, data centers accounted for 21 percent of total metered electricity in 2023, a level that pushed the national grid operator to pause new approvals around Dublin until 2028. In Querétaro, Mexico, and Uruguay, fast-tracked data center plans have collided with prolonged droughts.

The UN report also reframes a common assumption. Day-to-day inference, the act of answering a single user prompt, accounts for 80 to 90 percent of total AI energy demand, dwarfing the cost of training the models in the first place. ChatGPT alone processes an estimated 2.5 billion prompts a day, consuming roughly 383 GWh of electricity a year for a single product.

Hotter Than a Hot Tub: The 45°C Breakthrough

Nvidia’s answer is to stop fighting the heat and to circulate it instead. The company’s Rubin-generation AI infrastructure is the first to achieve 100 percent liquid cooling, with every chip and every networking component cooled entirely by liquid in a closed loop and no fans anywhere in the system. The coolant, a mix of 75 percent water and 25 percent propylene glycol, flows through cold plates that sit directly on the processors. The temperature ceiling is the headline. Nvidia’s newest AI servers can run their cooling liquid at up to 45 degrees Celsius, or 113 degrees Fahrenheit, hotter than a hot tub.

That number matters because the hotter the coolant can safely run, the less mechanical refrigeration is required. “With 45 degrees C, no water chillers are necessary for data centers,” Nvidia told reporters covering the announcement. “We are basically cooling this supercomputer with hot water.” Richard Whitmore, president and CEO of Schneider Electric’s Motivair cooling division, framed the shift in physical terms. “In the right geographic location, with the right system design, you don’t need any refrigeration equipment. You can just put big radiator coils outside and use the air temperature for all your cooling. It’s incredibly efficient.”

The reference design behind the system, called NVIDIA DSX, is the playbook Nvidia publishes for operators building AI factories. Ali Heydari, director of data center cooling and infrastructure at Nvidia, said the DSX design has “zero water consumption” outside of the roughly 1 percent of the year when some climates may still require chillers. Tom’s Hardware, covering a related Nvidia announcement last year, reported a 300x improvement in water efficiency for direct-to-chip cooling over conventional air-cooled servers.

What 100% Liquid Cooling Actually Changes

The closed loop is the structural shift. A traditional cooling-tower data center pulls in fresh water, evaporates a share of it into the air to dump heat, and discharges the rest. Nvidia’s design fills the loop once and runs it closed for the life of the facility, with no evaporation, no fresh intake, and no discharge tied to cooling. In favorable climates, the heat is rejected through outdoor dry coolers, essentially large radiator coils positioned outside the building, that the system only supplements with chillers a few days a year.

Industry estimates cited by Nvidia suggest that raising chiller plant temperatures by just one degree can cut cooling energy costs by about 4 percent. At full scale, the company estimates, a 50-megawatt hyperscale facility can save over $4 million annually in cooling-related energy and water costs by moving to a fully liquid-cooled architecture. The water saving is the larger of the two. A conventional cooling-tower system draws roughly 2.6 million gallons of water per megawatt per year. Nvidia’s 45-degree liquid-cooling architecture, in the right climate, drives that figure to near zero, which the company describes as up to a 100 percent reduction in facility cooling water use.

The Rubin platform also changes the physical plant. Because every component is liquid-cooled, rack density rises: a system that previously occupied six rack units now fits in two. Sealed front panels replace perforated bezels. The 85-decibel fan noise of a traditional server hall disappears, along with the choreographed hot-aisle, cold-aisle airflow that defines today’s data center floors. The waste heat is also a resource: it can be captured and routed to heat nearby commercial or residential buildings, an option Nvidia is actively marketing to operators.

The Numbers the Savings Claim Sits On

The headline figures are Nvidia’s own estimates, published in the company’s engineering blog. Independent confirmation of site-level water savings at 100 percent liquid-cooled scale is still thin, because the Rubin generation is not yet widely deployed. The most useful comparison is between the design’s two operating regimes.

The geography caveat is the load-bearing one. A site in the Scottish Highlands, where outdoor air is cool most of the year, can run dry coolers for most of the year. A site in Phoenix, Arizona, cannot, and the same architecture will need chillers on far more than 1 percent of its days. Nvidia’s published numbers assume the favorable climate case.

Why the Climate Math Still Doesn’t Add Up

The UN University report makes clear that cooling water is one slice of a larger pie. Even with zero cooling water use, an AI data center still draws electricity from the grid, and generating that electricity carries its own water and land footprints. The report finds that switching from coal to bioenergy, often marketed as a climate win, can cut electricity’s carbon footprint by 70 percent while increasing its water footprint more than thirty-fold and its land footprint a hundred-fold.

“What surprised us most is how often the choices that look greenest from a carbon perspective end up worse for water or for land,” said Dr. Miriam Aczel, the report’s lead author and a researcher at UNU-INWEH. Kaveh Madani, director of UNU-INWEH, made the same point more bluntly: “Low-carbon is not automatically low-water or low-land.” In a data center that has eliminated cooling water, the water footprint that remains is the one embedded in its power supply, and that number depends on where the power comes from.

There is also the cost of replacing the existing stock. Most of the world’s installed data center capacity is still air-cooled, and retrofitting it to the Rubin standard takes capital, downtime, and a workforce trained in liquid-cooling plumbing. Nvidia is selling the design to new AI factory builds, where the economics are cleanest, not to the legacy fleet.

The Rebound Problem Nobody’s Solving

Efficiency gains, the UN report argues, tend to inflate total demand. As models become cheaper to run, they get used more, and the per-prompt footprint shrinks while the absolute footprint grows. The same logic applies to cooling: a data center that no longer pays a water bill per megawatt can build another megawatt next door, and the world ends up with more data centers, not less water stress.

Madani called this out directly. “More efficient and affordable AI and energy mean more consumption of AI, making the overall footprint far bigger than what we save through efficiency gains.” The Jevons Paradox sits at the center of the UN report’s policy framework, and it is the reason the report’s recommendations are not limited to engineering fixes. It calls for governments to integrate AI infrastructure into water and land-use planning, for companies to disclose water and land footprints alongside carbon, and for permitting processes to reflect cumulative local impact, not just site-level efficiency.

Nvidia’s liquid cooling changes the slope of the water curve. It does not change the curve’s direction. The 1.3 billion-person water figure in the UN report is built on assumptions about compute growth, power mix, and cooling technology, and the 45-degree closed loop knocks down one of those inputs sharply. It does not address the other two. Whether the savings show up as a smaller water footprint or as more data centers depends on choices made by regulators, investors, and the operators who buy Nvidia’s chips. The hardware is ready. The policy frame is not.

For a related look at how the same Nvidia build-out is playing into cloud strategy, Apple’s move to put Private Cloud Compute on Google Cloud with Nvidia Blackwell GPUs shows the same architecture landing inside the privacy-focused AI services the technology was designed to enable. On the market side, Nvidia’s $5 trillion market cap, covered separately, now sits alongside a chip platform whose environmental claims are starting to be tested in real deployments.

Frequently Asked Questions

What is Nvidia’s new liquid cooling technology?

Nvidia’s Rubin-generation AI servers cool every chip and every networking component with a closed-loop liquid system, using a coolant of 75 percent water and 25 percent propylene glycol that runs at up to 45 degrees Celsius, hot enough to remove chillers in many climates. The architecture is described in Nvidia’s DSX AI factory reference design.

How much water can AI data centers save with liquid cooling?

Nvidia estimates that its 45-degree liquid-cooling architecture can reduce facility cooling water consumption from roughly 2.6 million gallons per megawatt per year for conventional cooling-tower systems to near zero in favorable climates, a figure the company describes as up to a 100 percent reduction.

What did the United Nations report say about AI and water?

A June 2026 report from the United Nations University Institute for Water, Environment and Health projects that, by 2030, the water footprint of AI data centers will equal the basic annual domestic needs of 1.3 billion people in Sub-Saharan Africa, alongside an electricity demand of 945 terawatt-hours.

Does liquid cooling fix AI’s climate problem?

It addresses cooling water use, which is one part of AI’s environmental footprint, but does not address the water and land embedded in the electricity supply, nor the rebound effect in which cheaper, more efficient AI invites more AI deployment, the report’s authors say.

Where can Nvidia’s 45°C liquid cooling run without chillers?

In climates with reliably cool outdoor air, such as the Scottish Highlands, the system can reject heat through outdoor dry coolers for most of the year. In hotter climates, including Phoenix, Arizona, chillers are still needed for a larger share of the year, though less than under conventional cooling.