NVIDIA is pushing back on the idea that the AI boom is draining local water supplies. The company points to a March 2026 estimate from the Manhattan Institute that data centers account for just 0.2% of U.S. freshwater use, and most of that is indirect — water consumed to generate the electricity they pull. The real fix, NVIDIA argues, is better cooling.
Why the water number matters
U.S. data centers used about 17.4 billion gallons of water directly in 2023, according to a Berkeley Lab report. That number excludes the far larger indirect share — another 211 billion gallons tied to power generation. By 2028, direct water use is projected to hit between 38 and 73 billion gallons. As AI scales, the indirect share climbs too. NVIDIA's argument is that the industry should focus on efficiency, not just total consumption.
What NVIDIA's cooling tech promises
NVIDIA's liquid-cooled systems operate at 45°C, which allows dry coolers to work in cool climates. The company says that cuts facility cooling water from 2.6 million gallons per megawatt per year to near zero. In 2025, NVIDIA claimed its Blackwell systems were 300 times more water efficient than traditional air cooling. That matters because cooling can eat up 40% of a data center's electricity, so the design also lowers power costs.
Ali Heydari, director of data center cooling and infrastructure at NVIDIA, describes the system as a closed loop that recirculates coolant instead of consuming fresh water. The company is betting that improved efficiency can decouple AI growth from water stress.
The limits of dry cooling
Dry coolers work best where it's cool. In hot, dry states — precisely where many data centers are being built — they lose effectiveness. That's a geographic constraint NVIDIA acknowledges. The company's case relies on the idea that data centers can be sited in cooler regions or use a mix of technologies. But the trend toward placing clusters near population centers and renewable energy sources often pushes them into warmer areas.
Memphis as a counterpoint
Elon Musk, who runs large NVIDIA-powered clusters through xAI, endorsed the company's push. But his own Memphis Colossus site tells a different story. That facility draws roughly 1.3 million gallons of drinking water per day from the local aquifer and ran dozens of gas turbines before securing permits. It's now the subject of a data center pollution lawsuit and community appeals. The contrast highlights the gap between what's possible in theory and what's happening on the ground.
The direct water share of data centers is set to grow, and the indirect share with it. NVIDIA's technology could reduce that footprint — but only if operators adopt it and regulators allow sites where it works. Whether dry coolers can keep pace with AI's thirst in the places where that thirst is growing fastest remains an open question.
