Nvidia announced at London Climate Week that its Rubin-generation AI infrastructure achieves 100% liquid cooling, cutting data center water consumption from roughly 2.6 million gallons per megawatt annually to near zero.
The most persistent knock against AI infrastructure buildout has been water. Every megawatt of AI compute running through conventional cooling towers drinks roughly 2.6 million gallons a year, and as hyperscalers race to deploy ever-larger clusters, that number compounds fast. On Monday at London Climate Week, Nvidia's chief sustainability officer Josh Parker made a pointed claim: "The water consumption challenge for data centers is largely solved." The vehicle for that claim is the Rubin-generation reference design, the first AI infrastructure platform Nvidia has built with 100% liquid cooling across every chip and every networking component, with no fans anywhere in the closed loop.
The physics behind it are straightforward, even if the engineering is not. Coolant, a 75% water and 25% propylene glycol mix, flows through cold plates that sit directly on the processors, pulling heat out at source. The key variable is the supply temperature: Rubin's system runs liquid at up to 45 degrees Celsius, or 113 degrees Fahrenheit. That sounds hot, but it's precisely the point. At that temperature, outdoor dry coolers, essentially large passive radiator coils positioned outside the building, can reject the heat without mechanical chillers or evaporative towers in most climates. The same coolant recirculates in a closed loop, so no new water is consumed. Gizmodo and Fortune both covered the announcement Monday; Axios confirmed the core claims the same day.
For a 50-megawatt hyperscale facility, Nvidia's own figures put the annual savings at over $4 million in cooling-related energy and water costs. That's not a rounding error in a capital budget, and it arrives at a moment when hyperscaler spending is accelerating, not tapering. According to Dell'Oro Group, AI infrastructure capex rose sharply in Q1 2026, with Google Cloud up 63%, AWS up 28%, and Microsoft committed to $190 billion for the calendar year. When the Rubin platform ramps in the second half of 2026, the efficiency argument will be embedded in the hardware from the start, not retrofitted.
Colocation operators and data center developers will feel this differently than the hyperscalers. The Rubin reference design effectively sets a new baseline for what modern AI infrastructure looks like, and any facility designed around air cooling or evaporative towers now carries an obsolescence question. Supermicro has already announced manufacturing support for Nvidia's Vera Rubin NVL72 and HGX Rubin NVL8 platforms with expanded rack-scale liquid-cooling capacity, which tells you something about where the supply chain is moving. For data center REITs and the wave of AI infrastructure startups competing on efficiency metrics rather than raw compute density, the shift matters: the conversation between facility operators and prospective tenants is no longer just about power density per rack but about water infrastructure, local climate, and whether a site can actually run chiller-less.
That geography point is where TechCrunch's coverage offered the sharpest caveat on Monday. A data center in the Scottish Highlands and one in Phoenix, Arizona, face different realities. When outdoor temperatures approach 115 degrees Fahrenheit, as they sometimes do in Nevada and Arizona where many new builds are concentrated, additional cooling resources are still needed. Nvidia's system works cleanly in temperate climates with reliably cool outdoor air. In hotter regions, it reduces water use significantly but doesn't eliminate the need for backup chilling. TechCrunch also noted that water consumption outside the facility, in electricity generation and chip manufacturing, can double or triple the total water footprint of an AI operation. Parker's claim that the challenge is "largely solved" applies to on-site facility water. The upstream picture is a different calculation.
Don't dismiss the announcement on those grounds, though. Nvidia isn't claiming to have solved the full-lifecycle water equation of AI, and pretending the company said that would be a misread. What it has done is close off the loudest argument regulators and local governments have used to push back against data center permits: that AI clusters are reckless consumers of scarce water. In a temperate climate, the Rubin design makes that argument difficult to sustain. The HVAC market noticed immediately. Data Center Dynamics reported that the Rubin cooling reveal triggered a drop in HVAC-related shares the same day, a market telling you something about where the structural shift in infrastructure capex is landing.
The broader implication for anyone watching the AI buildout is that the efficiency race is now multi-dimensional. For the first two years of the current infrastructure cycle, competition ran almost entirely on compute density: who could pack the most FLOPs per rack and procure the most Nvidia hardware. That phase isn't over, but a second competitive axis is emerging around operational costs, power procurement, and water governance. Rubin raises the floor for what responsible AI infrastructure looks like, and it does so not through policy but through design. The next wave of colocation and hyperscale facilities won't get to treat water as an afterthought.
Also read: China's all-CPU LineShine supercomputer reaches exascale without a single Nvidia or AMD chip • Masayoshi Son says Earth wins the AI compute race after dismissing Musk's orbital data center vision • SpaceX sheds $620 billion in market value within days of its record IPO as investors punish the Cursor deal
