Nvidia says it can cut data center water use. The AI boom has a bigger problem

Nvidia announced that its new Vera Rubin AI servers can sharply reduce data center water consumption by operating at higher coolant temperatures, potentially cutting on-site water use to near zero in many climates. However, experts caution that data centers will still drive large water use through electricity generation and require enormous power, meaning better cooling alone won't end criticism of AI's resource consumption.

Data centers, in case you haven’t heard, have a water problem https://www.wired.com/story/data-center-operators-fix-water-use-problems/ . As AI https://www.fastcompany.com/section/artificial-intelligence companies race to build the massive computing facilities needed to train and run powerful models, the water required to cool those facilities has become a flashpoint for communities, utilities, and environmental critics. According to reporting by the Environmental and Energy Study Institute http://www.apple.com/ EESI , a large data center can use up to 5 million gallons of water per day, roughly as much as a town with tens of thousands of residents. That concern has become part of a broader backlash over artificial intelligence infrastructure, including rising electricity demand, pressure on local grids, and the possibility that nearby ratepayers could end up shouldering some of the cost. That is why Nvidia’s announcement last week https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/ landed as a welcome, if limited, answer to one of the AI boom’s more stubborn infrastructure questions: Can the industry keep building bigger and more powerful data centers without consuming vast amounts of water to cool them? The company says its newest AI servers, built around its Vera Rubin platform https://www.nvidia.com/en-us/data-center/technologies/rubin/ named for the pioneering astronomer , can sharply reduce on-site water use by letting their cooling systems run hotter than before. The servers are cooled by a circulating fluid that enters the system at 45 degrees Celsius, or roughly 113 degrees Fahrenheit, hotter than a typical hot tub. The coolant can rise to 55 degrees Celsius, or about 131 degrees Fahrenheit, before being cooled back down outdoors by dry coolers, which work like large radiator coils that transfer heat outside the data center. Then the same fluid circulates back past the chips in a closed loop. The fluid, a mixture of water and propylene glycol, runs through structures called cooling plates that sit atop the computing chips, including Nvidia’s Rubin GPUs and associated Vera CPUs, to absorb their excess heat. The significance of the design is not simply that Nvidia is using liquid cooling. It is that the servers and cooling system can operate at relatively high temperatures, which means that in most outdoor temperatures in most climates, the fluid can cool back down to a reusable level without the need to evaporate water to carry heat away. According to Nvidia, in many climates, the system can cut data center water consumption for onsite cooling to close to zero. “The 45-degree intake temperature—that is really the newest innovation that’s really transformative,” says Josh Parker, Nvidia’s head of sustainability. The question is how much of AI’s water problem better cooling can actually solve. Experts say Nvidia’s technology is genuinely innovative and does have the potential to limit the amount of water, and potentially electrical power, used for cooling. But they also caution that data centers will still indirectly drive large amounts of water use through electricity generation, and will continue to require enormous amounts of power. In other words, better cooling technology alone is unlikely to end criticism of the AI industry’s resource consumption. For decades, servers in data centers were cooled like computers in homes and offices: by circulating cool air past hot chips and other components to absorb the excess heat they produce as they run. That air, in turn, is often cooled through processes involving the evaporation of water. That’s not unique to data centers. Evaporative cooling is used in all sorts of buildings, from factories to office towers. But in a high-powered data center packed with power-intensive processors, the water use can add up quickly. “Evaporation is a very effective heat removal mechanism,” says Aaron Wemhoff, a professor of mechanical engineering at Villanova University who has studied data center water usage. “This is why we sweat, because the evaporation of the sweat droplets off our skin is pretty much nature’s best way of keeping us cool.” Traditionally, air cooling was simpler and cheaper than liquid cooling, which requires piping fluid to a data center’s worth of chips, even though cooling fluids like the water-propylene glycol mixture are more efficient at absorbing heat than air. But with the Vera Rubin platform, which packs computing power into tight spaces to optimize the cross-chip networking needed for AI and otherwise boost efficiency, liquid coolant and its greater capacity for transporting heat became a practical necessity. Nvidia’s previous line of server technology had liquid cooling as an optional feature, Parker says. But it is the only cooling option available for the Vera Rubin platform, which Nvidia announced in May https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory it was “ramping into full production” for use in the AI data centers sometimes known as AI factories. “Really, the state-of-the-art AI factories require it to be able to operate at kind of the limits of the AI frontier,” says Parker. Because the chips and cooling system can tolerate those high temperatures, evaporative cooling is generally unnecessary, Parker says. In the vast majority of U.S. metro areas, so-called passive cooling without the need for evaporative cooling or other active refrigeration is sufficient 99% of the year. Even in sun-baked Phoenix, he says, that number drops to only about 88% of the year. Additionally, modern liquid cooling technology cuts the portion of a data center’s electrical power use that’s devoted to server temperature control, Parker says. “Overall cooling infrastructure typically represents about 5%–10% of total facility power, substantially lower than in traditional air-cooled facilities because moving liquid is far more energy-efficient than moving large volumes of air,” he writes in an email. Still, water use by AI servers is not limited to what happens on-site. Like data centers, many types of power plants naturally get hot as they burn fuel like gas or coal or generate heat from nuclear fission. Those facilities also typically use evaporative processes within their cooling towers to maintain proper temperatures. “If it’s a coal-fired plant, or a nuclear plant, or a natural gas plant, they can consume a lot of water,” says Eric Masanet, a professor at the University of California, Santa Barbara’s Bren School of Environmental Science & Management. “But if it’s a solar, photovoltaic power system, or wind power, those consume very little.” Depending on the mix of grid power sources, and in some cases on-site generating systems https://www.bloomenergy.com/blog/onsite-power-new-geographies-will-dominate-data-center-market-by-2028/ , data centers can still drive significant water use even if they are not evaporating water onsite. According to a 2024 Department of Energy report https://escholarship.org/uc/item/32d6m0d1 cited by EESI, data center indirect water consumption from electricity use in 2023 reached about 211 billion gallons—as much water as the population of New York City uses in seven months. And while AI servers continue to get more efficient, AI companies show no sign of reaching the limits of their demand for either computing or electrical power. Nvidia says its Vera Rubin platform can in some cases deliver up to 10 times https://nvdam.widen.net/s/7hztspzswk/gpu-architecture-datasheet-vera-rubin-nvidia-us-5198950-web the AI processing power per megawatt as its previous Grace Blackwell platform. But a more efficient server or cooling system can reduce the resource burden of a given workload while also making it easier for companies to run much larger workloads. Additionally, operators of data centers and the communities where they’re located still need to build and plan for the facilities’ peak power and water demands, says Shaolei Ren, an associate professor of electrical and computer engineering at the University of California, Riverside, who has written about AI resource use. If a data center anticipates using public water for evaporative cooling on the hottest days, utilities need to plan for that potential spike in demand and its effects on the larger water system. “Peak water is the infrastructure challenge for the local water systems,” says Ren. Exactly how Nvidia’s advances in cooling technology will influence data center planning and placement in the long term remains to be seen. Data center operators could focus more heavily on the broad swaths of the country where outside temperatures generally make it easy to get coolant back down to 45 degrees Celsius. As Parker points out, AI companies and other users could also temporarily shift computing workloads to cooler locations on hot days, or limit utilization until temperatures cool down. But for now, AI companies appear more likely to put every watt of power they can into crunching training data and responding to user queries. They also face increasing political constraints on where data centers can be built, which may make optimizing for outside climate less of an option. That means Nvidia’s hotter-running cooling system could make AI data centers less dependent on water at the site level. It may also make the next generation of AI infrastructure more efficient than the one before it. But it is unlikely to end the broader debate over how much water, electricity, and public infrastructure the AI boom should be allowed to consume.