.png)
The Truth About Nvidia’s New Cooling Systems
- 2 days ago
- 4 min read
Written by Logan Dorton
Edited by Tanner Drant and Francesca Howard
In late June, Nvidia announced a new warm-water cooling system for its AI data centers. It said, through Chief Sustainability Officer Josh Parker, that the water challenge for its data centers is now “largely solved.” That is quite the claim.
Currently, a single large data center can consume up to five million gallons of water a day, about what a town of tens of thousands uses, according to the Environmental and Energy Study Institute. Communities near data centers have spent the last two years fighting new construction on exactly those grounds. If Nvidia’s system does what the company says it can do, one of the most politically and ethically charged fights in the AI industry gets a lot easier.
The engineering behind the announcement is valid. What the announcement leaves out is where the water actually goes.
Nvidia’s new water cooling system itself is a closed loop. A coolant made of 75 percent water and 25 percent propylene glycol is added once and recycled for the life of the facility. It enters server racks at 45 degrees Celsius (113 degrees Fahrenheit)and exits at 55 degrees Celsius. That warm coolant is then run through outdoor dry coolers, essentially large radiators and cooled back down through contact with ambient air before circulating back to the chips. No evaporation. No cooling towers. No new water gets added after the initial fill.
This design targets the two largest water consumers in a traditional data center simultaneously. The first is evaporative cooling, which accounts for most of the daily water loss in a conventional facility. The second is the fleet of chillers that support conventional cooling systems, which not only consume power but often rely on water-intensive equipment to keep coolant temperatures low enough for the chips. Because Nvidia's coolant runs so much hotter than the water a chiller would produce and is hot enough to shed its heat to outdoor air, they can be turned off for most of the year. Chillers account for close to 40 percent of a data center’s electricity consumption, so the new system’s electricity savings compound the water savings. In favorable climates, Nvidia says the on-site water reduction can reach 100 percent.
That number is not just marketing; it is a legitimate engineering achievement, and the largest companies in the sector have already placed their orders. Amazon, Microsoft, Google, Meta, and Oracle have all reserved production of Nvidia’s Rubin chips, which contain the new cooling system, through 2027 and into 2028. If the transition goes as planned, community and regulatory opposition to new data center construction will lose one of its most powerful arguments, as water withdrawal permits are among the few remaining leverage points that local governments have over hyperscaler expansion. A near-zero on-site water footprint significantly changes the political dynamic, and Nvidia knows it.
A problem arises when you consider what Nvidia is actually measuring. According to TechCrunch, the company treats the data center as the unit of focus. Everything inside the building’s walls is counted; everything outside them is not. The truth is that data centers do not power themselves; rather, the electricity that runs them comes from power plants. These plants themselves often require a substantial amount of water to run. Generating a single kilowatt-hour of electricity from natural gas requires about 1.17 liters of water. From coal, it is roughly 2.2 liters. Hydroelectric power loses about 6.8 liters per kilowatt-hour to evaporation from reservoirs. Fossil-fuel plants supply roughly half of the electricity used by data centers today, and that share continues to grow as AI demand skyrockets.
The result, as the MIT Sloan Management Review points out, is a sustainability paradox. Nvidia’s cooling breakthrough may zero out one category of the overarching problem, but it leaves a much larger area untouched. Water is still being consumed, just by power plants rather than the centers themselves. The on-site number improves dramatically, but the system-wide number improves much less, and, depending on where a data center draws its electricity from, it may not improve at all.
That is the distinction Nvidia’s language elides. The engineering claim of near-zero on-site water use is true and defensible. However, the rhetorical claim that the water issue is “largely solved” depends entirely on where people place the challenge.
The distinction between the on-site and off-site water usage improvements matters because the phrase “largely solved” only applies to the former. A company applying to build a new facility in a drought-affected region can now point to Nvidia’s numbers and argue that concerns about local water withdrawal are outdated, which is technically correct on the withdrawal question but misleading on the consumption question. Whether regulators can tell the difference will determine how much political cover this breakthrough really provides.
The real question the industry is not yet answering is how impactful the cooling savings will be when paired with the power source. If the electricity powering the data center comes mostly from natural gas or coal, the water saved would only account for 20-30% of the total water usage, on and offsite, because 70-80% of water usage towards AI comes from power plants. If the mix moves toward solar and wind, which use almost no water, then the on-site savings actually reduce AI’s overall water footprint.
Nvidia has indeed built something worth celebrating, but also built something needing careful oversight. The engineering is a genuine step in the right direction, but the mega-company’s framing focuses on where the numbers look best. In the end, the question becomes whether AI’s water problem will actually be smaller in 5 years or just harder to see.
Sources:
