NVIDIA says AI data centers can cut water use with warmer liquid cooling

Why it matters

NVIDIA is trying to turn one of AI infrastructure's biggest permitting risks into a hardware-sales argument: newer GPU systems may use less water, but they also enable more buildout.

NVIDIA used a Monday post on X to push back on the claim that AI data centers are becoming an unmanageable water burden, pointing to a Manhattan Institute argument that data centers account for roughly 0.2 percent of U.S. freshwater consumption and to a new generation of liquid-cooled AI systems that can run warmer than earlier infrastructure.

The timing was not accidental. At London Climate Week, NVIDIA also put a sharper version of the case into the market: Josh Parker, NVIDIA's chief sustainability officer, told Axios that "the water consumption challenge for data centers is largely solved." The claim is broader than a cooling update. It is NVIDIA's answer to a political and permitting problem that has started to sit directly in the path of AI infrastructure growth.

Parker is the right messenger for that job. He joined NVIDIA in August 2023, after the ChatGPT shock turned GPUs from a back-office accelerator market into the main constraint on AI deployment, according to a separate Axios profile. NVIDIA co-founder and CEO Jensen Huang has spent three decades turning specialized chips into the computing substrate for each new wave of software. Parker's job is to keep the environmental story from becoming the limit on that strategy.

https://x.com/nvidia/status/2069147938098483586

The cooling claim

The technical claim centers on NVIDIA's latest rack-scale systems. NVIDIA says the GB300 NVL72 is a fully liquid-cooled architecture that integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single platform. Axios reported that NVIDIA's coolant, a recirculated mixture that includes water and propylene glycol, can run at 113 degrees Fahrenheit. Because the liquid can operate at a higher temperature, data centers may be able to rely less on chillers, which consume energy and can drive water demand depending on facility design and climate.

That is the strongest version of NVIDIA's case. Direct-to-chip liquid cooling attacks the problem where NVIDIA has leverage: the heat density of AI racks. If the cooling loop can move heat efficiently without forcing facility operators to chill large volumes of water or air, the water used inside the data center can fall sharply for new builds designed around those systems.

But NVIDIA is also making a market argument. The X post cited a Manhattan Institute essay that frames the water debate as an allocation problem rather than a volume problem. The essay says data centers use water directly for cooling and indirectly through the electricity they consume, and it cites estimates that data centers account for roughly 0.2 percent of U.S. freshwater consumption, most of it indirect through power generation.

That figure is useful context, but it is not the whole operating reality for data center developers. National shares can look small while local constraints are decisive. A facility that is modest in a national denominator can still matter if it lands in a drought-prone county, on a constrained municipal system, or in a region where utilities are already being asked to support several gigawatts of new AI load.

The constraint is local

The federal baseline shows why both sides of the debate can be directionally right. The Department of Energy said a Lawrence Berkeley National Laboratory report found that U.S. data centers consumed about 4.4 percent of total U.S. electricity in 2023 and could rise to 6.7 percent to 12 percent by 2028. The same DOE release said data center electricity use climbed from 58 TWh in 2014 to 176 TWh in 2023, with a projected increase to 325 TWh to 580 TWh by 2028.

Water follows that power buildout in two ways. Cooling water is the visible piece, because it is tied to a facility and a community water system. Indirect water use is less visible, because it depends on the power plants serving the load. A Congressional Research Service FAQ, citing the LBNL report, says cooling accounts for much of the non-IT power draw in data centers and notes that a 100-megawatt U.S. data center could directly consume roughly the same amount of water as 2,600 households, depending on cooling strategy.

Researchers are increasingly focused less on annual gallons and more on peak capacity. A March 2026 paper, "Small Bottle, Big Pipe", estimated that if 2024 water-use intensity persists, U.S. data centers could require 697 million to 1,451 million gallons per day of new water capacity through 2030. Under a more optimistic scenario with 10 percent annual reductions in water-use intensity, the estimated requirement falls to 227 million to 604 million gallons per day. The point is not that every data center will drain a region. It is that water infrastructure is built for peaks, not national averages.

That is where NVIDIA's argument becomes both more credible and more self-interested. Warmer liquid cooling can reduce the facility-level water burden of the AI systems NVIDIA sells. It can also make higher-density AI factories easier to permit and cheaper to operate, expanding the market for those same systems. Efficiency improvements do not automatically reduce total impact if they also accelerate deployment.

A reputational fight over AI infrastructure

NVIDIA is not alone in trying to reset the water narrative. Microsoft, Google and Amazon have all been pushing versions of the same message: new facilities can use closed loops, reclaimed water, improved cooling controls or lower-water designs. The World Resources Institute, taking the other side of the ledger, has warned that mid-sized data centers can use up to 300,000 gallons of water a day, while large facilities can consume as much as 5 million gallons daily, and that many new U.S. facilities are being built in water-stressed areas.

That is the real audience for NVIDIA's Monday post. The chipmaker is not merely correcting a statistic. It is trying to persuade utilities, regulators and local governments that the next wave of AI infrastructure can be built without repeating the water footprint of older cooling designs.

The open question is adoption. NVIDIA's GB300-class systems can change the economics of new AI factories, but they do not instantly retrofit the installed base, and they do not erase the electricity-water link in regions where power generation itself consumes water. The water backlash is not going away because a rack runs hotter. But NVIDIA is making clear where it wants the argument to land: not on whether AI data centers should be built, but on which cooling architecture gets approved when they are.