The chipmaker says its next-gen liquid cooling systems achieve 300x water efficiency, but experts aren't ready to declare victory
Nvidia is making a sweeping claim: the water problem plaguing AI data centers is basically handled. A top executive at the company says its next generation of AI infrastructure can largely address the growing concerns around water consumption that have dogged the industry for years.
The tech behind the claim #
At the heart of Nvidia’s pitch is a shift from traditional air cooling to warm-water direct-to-chip liquid cooling. The company’s GB200 NVL72 rack-scale system, part of the Blackwell platform, uses full liquid cooling that Nvidia says achieves 300x water efficiency compared to conventional air-cooled architectures.
The upcoming Vera Rubin platform pushes this further. It’s designed to operate with inlet water temperatures up to 45°C, which is about 113°F. That’s warm enough that data centers running these systems wouldn’t need the massive industrial chillers that traditionally gulp down enormous volumes of water and electricity.
Data center cooling can account for 30% to 40% of a facility’s total energy consumption. Nvidia claims its Blackwell systems deliver energy efficiency gains of 25x to 30x over previous generations.
Instead of spraying water into cooling towers or running massive chillers, these systems pipe warm water directly to the chips that need cooling, then cycle that water back out. The water stays in a closed loop rather than evaporating into the atmosphere.
Microsoft has pursued a similar approach with its Fairwater AI data centers, which use closed-loop liquid cooling systems. The company claims those facilities consume water annually at a rate comparable to that of a single restaurant. For context, a typical hyperscale data center can consume millions of gallons of water per year.
Why experts aren’t popping champagne #
Warm-water cooling systems work best in specific conditions. In cooler climates, the warm water exiting the chips can be cooled using outside air alone, a process called free cooling, which eliminates the need for chillers entirely. But in hot, humid regions like Texas or Southeast Asia, where a huge number of data centers are being built, the math changes considerably.
The efficiency gains from warm-water systems mainly facilitate free cooling and lower overall water use, but they don’t guarantee elimination of water consumption in every deployment scenario.
Where crypto fits into the picture #
Cryptocurrency mining operations have increasingly been repurposing their facilities to support AI workloads. The infrastructure overlap is significant: both AI training and crypto mining require dense compute power, robust cooling, and reliable energy supplies. As Nvidia pushes high-density liquid cooling designs into the mainstream, mining operators are tapping into those same architectures to pivot toward AI hosting.
Nvidia’s announcements at events like its annual GTC conference have historically driven rallies in AI-themed crypto tokens. Projects like Bittensor (TAO), Fetch.ai (FET), and Render (RNDR), all of which are tied to decentralized compute and AI infrastructure, tend to react to Nvidia news as a proxy for broader AI demand.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our