At London Climate Week on June 22, the world's dominant AI chipmaker arrived with good news about its thirstiest product. Nvidia published a reference design for its next-generation data centers, the Rubin-era DSX, and said it had effectively eliminated their water use. "We have eliminated massive amounts of power usage and pretty much all water usage," its director of data-center cooling, Ali Heydari, told the conference. The headline wrote itself: Nvidia solves AI's water problem. The trouble is in where Nvidia drew the boundary line.
The engineering is real and clever. The DSX system cools every chip with liquid in a sealed loop, no fans anywhere, using a mix of water and propylene glycol, the same antifreeze chemistry in a car radiator. The trick is heat. The coolant runs in at up to 45 degrees Celsius, hotter than a hot tub, and leaves the servers at around 55. Because that fluid is already so warm, outdoor radiators can shed its heat passively, without the evaporative cooling towers that boil away fresh water by the millions of gallons. In most temperate climates, the design needs no water for cooling at all.
But cooling is the small part. Scientists at Lawrence Berkeley National Laboratory estimate that U.S. data centers consumed about 228 billion gallons of water in 2023. Of that, only 17 billion gallons went to cooling the servers. The other 211 billion, more than nine-tenths of the total, was consumed generating the electricity that powered them. That water evaporates at the power plant, far outside the walls of any data center, and a sealed cooling loop does nothing to bring it back. Nvidia's design tackles the 7 percent it can see and leaves the 93 percent untouched.
That gap is exactly why the public numbers are such a mess. A News9 review of the science found the viral water-per-query figures span a factor of two thousand: Google and Sam Altman put a typical query at roughly five drops, while a widely shared claim pegs an AI-written email at a full half-liter bottle. None is simply right, because they measure different things, some counting only the cooling water, others adding the power-generation water, others working from outdated hardware. The honest comparison the researchers keep reaching for is humbling: a single hamburger consumes more water than thousands of AI prompts, and the country's real water sinks remain growing food, watering lawns and taking showers.
Which leaves the actual story not as a fight over whether AI's thirst is catastrophic or trivial, but as a problem of disclosure. The companies building this infrastructure do not publish consistent figures on what it consumes, so estimates get to ricochet between five drops and a water bottle unchallenged. "The cleanest form of AI use is no use," said Kaveh Madani of the United Nations University, a deliberately provocative line aimed less at individual guilt than at the bait-and-switch of having generative features pushed into tools people never asked to make thirstier. A genuinely sealed cooling loop is worth having. Sold as the end of AI's water problem, it is mostly a way of looking at the smaller number.