Nvidia's Cooling Advance Is Real. AI's Water Problem Is Not Solved

Nvidia this week introduced a warm-water cooling design for AI data centers that it says can eliminate nearly all water use inside the facility. That is a real engineering advance, and enterprises should welcome it. But one claim attached to the launch deserves pushback before it ends up in a sustainability report. According to TechCrunch, Nvidia chief sustainability officer Josh Parker told Axios that "the water consumption challenge for data centers is largely solved." It is not, and enterprise buyers of AI capacity should not repeat that line until their providers can show the full picture.

What did Nvidia actually announce?

The design, described by TechCrunch's Tim De Chant on June 22, uses a closed loop of coolant that is filled once and recirculated for the life of the facility. Coolant enters the racks at 45°C (113°F) and leaves at 55°C (131°F). At those temperatures, outside air can often carry the heat away through passive radiators, without evaporative cooling and sometimes without fans. In favorable climates, Nvidia says, that can mean a 100% cut in on-site water use.

That deserves credit. Evaporative cooling is where on-site water goes, so a design that removes most of it addresses the part of the footprint that neighbors and local water utilities see most directly. It also fits a broader engineering trend. As HPE Federal AI Chief Technologist Steve Heibein put it during a AI Connect expert webinar on AI and HPC convergence, "A thousand watt GPU needs a thousand watts of cooling." Liquid cooling is how the industry keeps up.

Why isn't AI's water problem solved?

Because most of the water tied to an AI data center is not used inside it. TechCrunch cited research showing that indirect water use, from generating electricity and manufacturing chips, can double or triple a facility's total water footprint. By the publication's estimate, Nvidia's system addresses roughly a quarter to a third of total consumption.

The power source drives much of the rest. TechCrunch, citing U.S. Geological Survey data, reported that U.S. fossil fuel power plants consume 2.7 billion gallons of water a day, mostly for evaporative cooling. It cited a University of California study putting natural gas plants at 1.17 liters per kilowatt-hour and coal plants at 2.2 liters. Wind and solar, by contrast, use about 0.01 and 0.03 liters per kilowatt-hour respectively, including manufacturing and panel cleaning.

The trend line makes this harder, not easier. TechCrunch noted that fossil fuel plants supply about half of today's data center power, according to the International Energy Agency, and that the IEA projects natural gas and coal will supply more than 40% of the new electricity needed for data center demand through 2030. A facility that uses no water on site but runs on gas-fired power is moving its water use upstream, not eliminating it.

Isn't on-site water the part that matters most to communities?

This is the strongest argument on the other side, and it has merit. Water drawn from a stressed local aquifer has a different impact than water evaporated at a power plant in another watershed. Local permitting fights are mostly about local water, and Nvidia's design helps there.

But enterprises are not only answering to the town next to the data center. They answer to investors, regulators and their own sustainability commitments, which increasingly ask about supply chain impacts, not just direct operations. A company that reports its AI workloads as water-free because its provider uses closed-loop cooling is reporting a partial number. If the power behind those workloads comes from gas or coal, the gap could be large.

What should enterprise buyers ask for?

The AI Connect community has spent years arguing that AI programs succeed when they measure what matters, not what is easy to measure. Water is no different. We think enterprise buyers of AI capacity, whether through hyperscalers, neoclouds or colocation, should ask for three things:

• Full-footprint water reporting. Ask providers to report water used on site and water associated with their power supply, by region, not just a facility-level water usage figure.
• Power source transparency. Water intensity depends heavily on the grid mix. Ask which generation sources supply the regions where your workloads run, and how that is expected to change.
• Comparable methods. TechCrunch said it asked Nvidia to clarify how it measures water use, and the article did not report a reply. Buyers should insist on clear, consistent methodology before they rely on any provider's claims.

There is also a lever enterprises control directly: where and how much they compute. Placing flexible workloads in regions with cleaner power, and cutting waste through right-sized models and better AI infrastructure cost discipline, reduces both the energy bill and the water footprint. The AI factory model AI Connect speakers have described is only efficient if the GPUs in it are well used.

The bottom line

Nvidia's warm-water design is good engineering and good news for communities near data centers. Declaring the water problem largely solved is premature. Enterprise leaders should welcome the progress, then ask their providers for the rest of the numbers. Until they get them, sustainability reports should describe AI's water footprint as a partially measured problem, not a closed one.

Frequently asked questions

Does Nvidia's new cooling eliminate data center water use?

It can nearly eliminate water used inside the facility in favorable climates, but not the water consumed by the power plants that supply it.

How much water do power plants use for data centers?

TechCrunch cited research putting natural gas plants at 1.17 liters per kWh and coal at 2.2 liters, versus about 0.01 for wind and 0.03 for solar.

What should enterprises ask AI providers about water?

Ask for on-site and power-related water use by region, the generation mix behind the workloads and a clear measurement method.

Opinion pieces reflect the views of The AI Enterprise Magazine Team.

Filed under: Data & AI Infrastructure, AI Governance & Trust