Nvidia recently announced a warm-water liquid cooling system that, according to one of its executives, eliminates "pretty much all water usage" inside a data center. The company's chief sustainability officer went even further, telling reporters that the water consumption problem for data centers is "largely solved."

Let's pump the brakes on that one.

The system itself is legitimately clever. Coolant enters server racks at 45°C (113°F) — toasty for a human, perfectly fine for a GPU — absorbs heat from the chips, and exits at 55°C (131°F). At that temperature, passive radiators can shed the heat into outdoor air without evaporative cooling towers, fans, or chillers in many climates. The coolant runs in a closed loop, filled once and recirculated indefinitely. On-site water consumption: effectively zero. That part checks out.

But here's where the press release quietly stops telling you things.

The Boundary Problem

Nvidia's water accounting draws a neat box around the data center building itself. Anything happening inside that box — cooling systems, humidity control, all of it — gets measured and optimized. Anything outside the box? Not Nvidia's problem, apparently.

This is a classic case of optimizing the metric you control while conveniently ignoring the ones you don't. And the numbers outside that box are not small.

Water consumption in electricity generation and chip fabrication can double or triple the total water footprint of a data center facility. Which means Nvidia's genuinely impressive cooling solution addresses somewhere between a quarter and a third of AI's actual water problem. The rest is happening upstream, out of frame, and off the sustainability report.

Where the Water Actually Goes

Power plants are thirsty. Fossil fuel plants — which currently supply roughly half of all data center electricity, per IEA figures — are among the largest water consumers in the United States, burning through 2.7 billion gallons per day nationally. Natural gas plants consume about 1.17 liters of water per kilowatt-hour generated. Coal is worse at 2.2 liters per kilowatt-hour. Most of this is evaporative cooling, the same basic mechanism Nvidia just eliminated from inside the data center — just located somewhere else, at massive scale.

Hydropower looks cleaner on paper but isn't off the hook either. Reservoir evaporation works out to roughly 6.8 liters lost per kilowatt-hour — higher than fossil fuels, actually. Geothermal is a mixed bag depending on the specific technology, though some newer enhanced geothermal players are engineering around the worst of it.

Wind and solar are the genuinely good news here: roughly 0.01 and 0.03 liters per kilowatt-hour respectively, including manufacturing and panel cleaning. Those are numbers small enough to essentially round to zero.

The Inconvenient Trajectory

Here's the catch: the IEA projects that natural gas and coal will still supply more than 40% of the new electricity needed to meet data center demand through 2030. Tech companies aren't just passively accepting this — many are actively building or contracting for new gas capacity to feed their AI buildouts. The economics of massive, always-on compute loads make intermittent renewables complicated, and dispatchable gas plants solve that problem quickly, if messily.

So we have a situation where Nvidia solves the cooling water problem inside the facility, which is real and worth doing, while the industry simultaneously locks in years of fossil-fuel-powered electricity that carries its own substantial water footprint. The net improvement to AI's total water consumption is... considerably less dramatic than "largely solved."

Credit Where It's Due, Skepticism Where It's Earned

None of this is to say Nvidia's cooling system is worthless — it isn't. Eliminating on-site water consumption matters, especially in water-stressed regions where data centers are competing with agriculture and municipal supplies for the same aquifer. The engineering is real. The efficiency gains from ditching fans and chillers are real. If you're building a data center, this is the kind of system worth taking seriously.

But "the water consumption challenge is largely solved" is a sustainability claim that only survives scrutiny if you've agreed in advance to ignore most of the water consumption. That's not solving a problem. That's redrawing the boundary until the numbers look better.

The actual solution involves what's happening on the grid — which is a harder, slower, more expensive problem than swapping out your cooling architecture. Nvidia can't fix that alone, and to be fair, no one expects them to. But they probably shouldn't be taking victory laps on it either.