On June 22, 2026, at London Climate Week, Nvidia took the stage to unveil a new data center reference design that could dramatically reshape the water footprint of AI infrastructure. The Vera Rubin DSX—a purpose-built architecture for next-generation AI server racks—integrates a closed-loop liquid cooling system capable of slashing on-site water consumption to near zero, provided the local climate cooperates. It's a bold technical step, but it stops well short of a comprehensive solution to the water challenges that plague the AI industry.

The timing wasn't accidental. As global regulators, environmental groups, and local communities increasingly scrutinize the resource appetite of hyperscale data centers, Nvidia aimed to position itself as a leader in sustainable computing. The DSX design, named after the astronomer who unlocked the secrets of dark matter, attempts to address one of the most visible metrics of data center sustainability: the millions of gallons of water that traditional cooling towers evaporate each day.

Inside the Vera Rubin DSX: A Closed-Loop Revolution

The Vera Rubin DSX isn't a product you can buy off the shelf. It's a reference architecture—a blueprint that data center operators can follow when building facilities tuned for Nvidia's highest-end AI accelerators, including the forthcoming Rubin platform. The core innovation lies in its cooling subsystem. Rather than relying on evaporative cooling towers that consume vast quantities of water, the DSX deploys a secondary liquid loop that transfers heat from the rack-level cold plates to an outdoor heat rejection unit.

In favorable climates—those with low ambient temperatures and moderate humidity—the system can reject all waste heat directly to the atmosphere using dry coolers, much like a giant car radiator. This eliminates on-site water use entirely for cooling. In hotter or more humid locales, the design allows for supplemental adiabatic cooling, which pre-cools the air entering the heat exchanger by evaporating a small amount of water on its surface. Even then, water consumption is a fraction of what traditional cooling towers demand.

Nvidia claims the design can support rack power densities well above 100 kilowatts, which is rapidly becoming the norm for dense AI clusters. A single Vera Rubin DSX pod, comprising multiple racks, could cool upwards of a megawatt of IT load without a single drop of water leaving the facility as evaporation loss, provided the ambient conditions hold.

Why Water Cooling Matters for AI

To grasp the significance of the DSX, you need to understand how thirsty modern AI infrastructure has become. Training a single large language model can consume tens of millions of gallons of water over its lifecycle—not just for cooling the chips, but also embedded in the electricity generation that powers them. Data centers in the U.S. alone used an estimated 1.7 billion gallons of water per day in 2025, a figure that has only climbed as AI workloads multiply.

Water scarcity is no longer a distant concern. Communities from Arizona to Ireland have pushed back against new data center projects, forcing operators to commit to water-positive pledges or face outright moratoriums. Google, Microsoft, and Amazon have all touted water replenishment programs, yet their total water footprints continue to swell as AI product rollouts accelerate. Nvidia, as the dominant force in AI hardware, has felt the pressure to decouple chip performance from water intensity.

The Vera Rubin DSX targets the most visible part of the problem: cooling water consumption at the data center site. By switching to a closed liquid loop and dry heat rejection, it addresses the "scope 2" water used within facility boundaries. For a hyperscale campus that might otherwise consume millions of gallons per day, the reduction can be staggering—potentially over 90% in optimal climates. Nvidia framed the announcement as a direct response to operator demand for more sustainable infrastructure.

The Catch: Not a Complete Water Fix

But there's a reason the London Climate Week presentation included careful caveats. Nvidia's own press materials and the DSX technical briefs acknowledge that on-site cooling water is only one slice of the water pie. The phrase "near-zero on-site water cooling" leaves ample room for other water impacts to remain.

First, the electricity that powers the DSX racks doesn't appear out of thin air. Thermal power plants—coal, gas, nuclear—withdraw enormous volumes of water for cooling. Hydropower facilities evaporate water from reservoirs. Even solar and wind have indirect water footprints through manufacturing and maintenance. Nvidia's reference design does nothing to reduce the water intensity of the grid that feeds it. If a DSX data center plugs into a coal-heavy grid, the plant cooling water consumption per megawatt-hour remains unchanged.

Second, the manufacture of AI accelerators is notoriously resource-intensive. Semiconductor fabrication facilities (fabs) require ultra-pure water in staggering quantities to wash silicon wafers. A single large chip like Nvidia's Rubin could embody thousands of gallons of upstream water. The DSX design does not mitigate this embedded water, nor does it address the water used in mining and refining the rare minerals that go into advanced packaging and HBM memory stacks.

Third, the "in favorable climates" asterisk matters enormously. Many of the world's largest data center hubs—Ashburn, Virginia; Phoenix, Arizona; Singapore—endure hot, humid summers where dry cooling alone can't keep up without either a massive efficiency penalty or a small but real water assist. In such conditions, the DSX might still consume significant water through adiabatic pre-cooling, though far less than a cooling tower. Nvidia did not provide specific water usage effectiveness (WUE) targets for various climate zones, leaving operators to guess how much water they'd actually save in practice.

Community and Analyst Response

Early reactions from data center engineers and sustainability analysts have been cautiously optimistic but skeptical of Nvidia's framing. Forum discussions broke down the physics: A closed loop with dry coolers can absolutely eliminate on-site cooling water if the system is designed with high approach temperatures—meaning the coolant loop runs hot enough that even warm outdoor air can still absorb heat. That tradeoff often means running chips at higher temperatures, which can reduce performance or long-term reliability. Nvidia has not disclosed the allowable chip junction temperatures for Rubin in the DSX configuration.

Others pointed out that water consumed off-site for power generation often dwarfs on-site cooling water. A data center with zero cooling water but a PUE (power usage effectiveness) of 1.0 would still responsible for roughly 0.5–2 liters of water per kilowatt-hour of electricity from a typical grid, depending on the generation mix. That indirect consumption can exceed the avoided on-site evaporative loss, making the overall water savings less impressive than the headline suggests.

Environmental advocates welcomed the move but called for Nvidia to publish full water footprint disclosures, including scope 3 (supply chain) impacts. "Near-zero on-site cooling is a great milestone," said one London-based policy analyst. "But if you're drinking a bottle of water while telling everyone you're water-free, the optics don't align. We need transparency across the full value chain."

The Bigger Picture: Nvidia's Sustainability Narrative

The Vera Rubin DSX announcement fits into a multi-year sustainability push that has seen Nvidia emphasize power efficiency more than water. Previous keynote slides boasted about the energy proportionality of Blackwell GPUs and the efficiency of NVLink fabrics, but water rarely received top billing. London Climate Week suggests that's changing, likely due to upcoming EU data center efficiency regulations that will mandate water usage reporting and, eventually, binding reduction targets.

Nvidia may also be laying groundwork for future chip generations that can operate at higher temperatures without throttling, which would make dry cooling more universally viable. Rumors suggest that the Rubin architecture includes a thermal tolerance mode that allows the GPU die to reach 95°C continuously, up from the 85°C typical of Hopper. If true, that would enable liquid-to-air heat rejection in climates as warm as 40°C, greatly expanding the "favorable climates" zone. However, Nvidia did not confirm such specifications at the event.

What's Missing: Water Stewardship Beyond the Data Center Floor

To truly address the "AI water fix," experts argue, Nvidia and its customers must embrace a holistic water stewardship strategy that includes:

  • Location-based siting: Placing data centers in regions with abundant, low-competition water resources or temperate climates where dry cooling works year-round.
  • Grid decarbonization: Pairing facilities with on-site renewable generation or power purchase agreements that displace water-intensive thermal electricity.
  • Chip lifecycle water accounting: Publishing detailed water footprints for each product generation, from wafer to final assembly, and working with suppliers to reduce those impacts.
  • Water reuse and circularity: Capturing all effluent streams—even those small adiabatic mists—and treating them for re-use, aiming for genuine water neutrality.

Nvidia's DSX reference design touches only the first of these by making dry cooling more practical. It's a vital piece, but not the whole puzzle.

The Road Ahead for AI Data Centers

Despite the limitations, the Vera Rubin DSX could accelerate an industry shift away from evaporative cooling, which has been the default for decades because it's cheap and effective. Hyperscalers like Microsoft and Meta have already invested in "zero-water" campuses that use air-cooled or closed-loop systems, but those designs often cap out at lower rack densities. The DSX promises to bring similar water savings to the ultra-dense AI pods that represent the fastest-growing segment of data center capacity.

If widely adopted, the design could save billions of gallons of water annually across the global AI fleet. That's a meaningful environmental victory, even if it doesn't tell the complete story. The real test will be whether Nvidia's partners actually build to the reference specification, or whether project economics lead them back to water-intensive cooling towers when land, power, or climate constraints bite.

In the coming months, we'll likely see the first DSX-based deployments announced by major cloud providers and Nvidia's DGX cloud customers. Their real-world water performance numbers will either validate Nvidia's projections or expose the hidden water costs that the London announcement glossed over. Until then, the Vera Rubin DSX stands as an ambitious blueprint—one that takes a significant step toward solving AI's water problem, but still leaves the faucet dripping.