Nvidia unveiled its new 45°C cooling design for AI servers on Monday that it says can cut on-site data center water use to near zero. The system runs a 75-25 mix of water and propylene glycol at 45 degrees Celsius through a closed loop, letting the heat vent to outdoor air without evaporative cooling for most of the year. Three weeks earlier, a United Nations report had warned that AI’s water footprint could match the annual domestic needs of 1.3 billion people by the end of the decade.
Nvidia’s chief sustainability officer, Josh Parker, told Axios that “the water consumption challenge for data centers is largely solved.” TechCrunch reports that Nvidia’s on-site saving addresses only about a quarter to a third of AI’s total water footprint, with the rest hidden upstream in power generation and chip manufacturing. The United Nations report that put the 1.3 billion equivalent in print was published on 3 June 2026. Nvidia’s announcement came three weeks later, on Monday.
How Nvidia’s 45-Degree Coolant Works
The new design sits inside Nvidia’s Rubin generation of AI servers, which the company calls the first to run 100% on liquid cooling, with no fans anywhere in the system. Coolant flows from a single inlet through cold plates mounted directly on every chip and networking component, absorbs the heat, and exits at roughly 55°C. “Once the watts per chip crossed a certain level, liquid cooling became mandatory,” said Richard Whitmore, whose Motivair division at Schneider Electric has worked with Nvidia for nearly a decade.
Ali Heydari, Nvidia’s director of data center cooling and infrastructure, said in the announcement that the closed loop is filled once and runs for the life of the facility, with no evaporative cooling required except, in some climates, on roughly 1% of the year’s hottest days. Nvidia’s blog says the system cuts facility water use from roughly 2.6 million gallons per megawatt per year to near zero, a drop it attributes to the 45°C operating point. The temperature is unusually high for a data center: the industry standard has long been 30°C, which forces operators to run large chillers year-round. “The reason that they want to do this is that if you can cool the chips at a higher temperature, it becomes easier to vent that heat into the outside environment, because it’s a higher temperature supply, and heat flows downhill,” said Andrew A. Chien, a professor of computer science at the University of Chicago who directs the CERES Center for Unstoppable Computing. Chien called the new loop a step that “shows really what’s possible in terms of pushing up this liquid input temperature to 45°C.”
The savings can be significant at scale. A 50-megawatt hyperscale facility can save over $4 million a year in cooling-related energy and water costs by moving to liquid-cooled infrastructure, according to Nvidia. Cooling has historically consumed up to 40% of a data center’s electricity, and industry estimates cited by Nvidia say raising chiller temperatures by just 1°C cuts cooling energy costs by about 4%. The new design, the company says, lets operators skip the chiller stage for most of the year.
The UN’s Three-Week-Old Water Warning
Three weeks before Nvidia’s announcement, the United Nations University Institute for Water, Environment and Health published a report putting numbers on the AI water question. The report, “Environmental Cost of AI’s Energy Use”: Carbon, Water and Land Footprints, was released on 3 June 2026 from Richmond Hill, Ontario. Its core projection: data center electricity demand will nearly triple the combined annual electricity use of Pakistan, Bangladesh and Nigeria by 2030. The water that demand will draw, the report says, will equal the basic annual domestic water needs of 1.3 billion people in Sub-Saharan Africa.
- 945 TWh: projected 2030 global data center electricity demand, nearly 3% of projected world use
- 9.3 trillion liters: associated water footprint of 2030 data center electricity
- 14,500 km²: associated land footprint of 2030 data center electricity, about twice the Jakarta metropolitan area
The report’s authors argue these three footprints do not move in the same direction. “Switching from coal to bioenergy, for example, can on average cut the carbon footprint of electricity by 70 per cent, while increasing its water footprint more than thirty-fold and its land footprint a hundred-fold,” it says. “Low-carbon” is not automatically “low-water” or “low-land,” the report concludes, warning that evaluating AI sustainability through a single metric can hide trade-offs and shift environmental burdens onto regions already facing water or land stress.
The Water Hiding Outside the Walls
A cooling system that eliminates the on-site water draw still has to be fed by a power grid. That grid is, in most of the world, dominated by water-thirsty thermal generation. Fossil fuel power plants collectively generate about half of all data center power today, according to the International Energy Agency. The water those power plants consume is large, and uneven across sources.
- Natural gas: 1.17 liters per kWh
- Coal: 2.2 liters per kWh
- Hydropower: 6.8 liters lost to reservoir evaporation per kWh
- Solar: 0.03 liters per kWh
- Wind: 0.01 liters per kWh
Wind and solar are the smallest water users per kWh, TechCrunch reports, with figures that include the water needed to manufacture and clean solar panels. Natural gas, coal, and hydropower all sit much higher in the TechCrunch data. Hydropower dams do not consume water in the same direct way, but evaporation from their reservoirs amounts to 6.8 liters lost per kWh. Geothermal, a source some tech companies are starting to explore, varies widely, and some startups, like Fervo, have pledged to use mostly “degraded” water that would otherwise go unused. The per-kilowatt-hour spread is wide, and the choice of grid mix is the lever the cooling system cannot move.
The grid feeding most AI data centers is not moving quickly toward the low-water sources. TechCrunch reports, citing the IEA, that natural gas and coal are expected to provide more than 40% of the new electricity capacity needed to meet data center demand through 2030. Fossil fuel power plants in the U.S. alone consume 2.7 billion gallons of water per day, most of it for evaporative cooling, according to the U.S. Geological Survey, as cited by TechCrunch.
Nvidia’s solution addresses about a quarter to a third of AI data centers’ total water consumption, TechCrunch reports, because most of that water sits in coal, gas and hydro plant cooling, not in chip cooling. The UN report’s central number is the data center industry’s water footprint, and that number sits upstream in the grid, not on the chip. Hyperscale sites not on the grid often run their own natural gas turbines, Tom’s Hardware notes, removing the on-site cooling water draw but creating separate local pollution concerns. The IEA’s projection of 40%-plus fossil-fuel power through 2030 is what locks the upstream number in place.
When Efficiency Means More Consumption
Even setting the grid aside, there is a second problem Nvidia’s closed loop does not touch. The UN report invokes the rebound effect, the Jevons Paradox, warning that as AI systems become more efficient, they become cheaper to run and are used more.
A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient. But that is only a partial picture of the overall problem. 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 line is from Professor Kaveh Madani, who directs the UN University Institute for Water, Environment and Health and was named the 2026 Stockholm Water Prize Laureate. “If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean but that is solving one problem while creating other problems, often in places that didn’t ask for it,” said Dr. Miriam Aczel, the report’s lead author. Inference, the work AI does after training, accounts for 80 to 90% of total AI energy use, the report finds, which makes efficiency gains at the chip level a moving target. A typical conversational chat query is around 200 times more energy-intensive than basic text classification, the report says, and generating a single AI image can require around 1,450 times that baseline. ChatGPT alone processes around 2.5 billion prompts per day, roughly 383 GWh a year for a single product.
Microsoft said in August 2024 that its new water-free data center design would save more than 125 million liters of water per year per facility, and continued to add new capacity. Amazon, in a recent report, has also pushed to raise the operating temperature of its mostly air-cooled data centers, a smaller version of the same lever Nvidia is now pulling, The Verge reports. “AI workloads are not getting lighter. The compute demand driving data center construction is growing faster than almost any other category of infrastructure investment,” Nvidia’s blog notes.
Where Data Centers Are Straining Local Supplies
The pressure on local water shows up first in the places where data center capacity has grown fastest. The UN report documents several cases where AI infrastructure is now a dominant local load. In Ireland, data centers accounted for 21% of total metered electricity in 2023, exceeding all urban households combined, and the national grid operator has paused new approvals around Dublin until 2028. “If you map where data centres are getting built against where water stress is worst, you tend to see the same regions in some instances,” said Dr. Mir Matin, a co-author of the UN report who manages its Geospatial, Climate and Infrastructure Analytics Programme.
The same report points to Querétaro, Mexico, where expanding compute infrastructure is drawing on water supplies amid prolonged droughts, and to Uruguay, where plans for a water-intensive data center coincided with a 2023 drought that depleted Montevideo’s freshwater reserves, making tap water unsafe to drink. “The communities living near these sites are not necessarily the ones using the AI being run there,” Dr. Matin said, calling the imbalance an environmental justice issue. “That asymmetry is the issue. Without fixing it, we’ll just be repeating older patterns, where some places carry the costs and other places capture the benefits,” he said.
Why the New Cooling Won’t Reach Most Data Centers Yet
The closed loop is not a retrofit, and existing air-cooled facilities cannot simply swap in a hotter coolant. Nvidia’s reference design is built around the Rubin generation, the first to use 100% liquid cooling on every chip and networking component. The trade-off is cost. “The catch is that these systems are expensive,” Andrew Chien, who directs the CERES Center for Unstoppable Computing at the University of Chicago, told Fortune. Chien added that 45°C is “a direction that more people should be trying to get to, because it’ll reduce the total power consumption of these large data centers.”
The rollout will be slow. Tom’s Hardware reports that the new cooling addresses several of the issues that local governments raised earlier in 2026, when more than 75 data center projects were delayed, though those delays all preceded Nvidia’s announcement. The same outlet adds that the system is most effective in cooler climates, but should still work in warmer areas as long as ambient temperature stays below 113°F, with chillers only needing to come on during occasional temperature swings. “It’s super important to push it up, because in many cases it allows you to do that cooling, that exhausting of heat to the outside environment without running HVAC units, without running air conditioners,” Chien said.
Nvidia’s blog names Phoenix as one of the climates where the design faces its toughest test, alongside the Scottish Highlands as one of the easiest. The 45°C coolant can only be cooled by outdoor air when the ambient temperature is below 45°C, and beyond that, the chiller returns. The Verge reports that Nvidia’s blog does not publish a price for the system, and notes that operators not on the grid often run their own natural gas turbines, which removes the on-site cooling water draw but creates separate local pollution concerns.
“In the right geographic location, with the right system design, you don’t need any refrigeration equipment,” said Richard Whitmore, whose Motivair division at Schneider Electric has worked with Nvidia for nearly a decade. “You can just put big radiator coils outside and use the air temperature for all your cooling. It’s incredibly efficient,” he said. The up to a 100% reduction Nvidia cites is the upper bound and applies only where the outdoor air can handle the heat.
