Nvidia is trying to change the way people talk about AI data centers—at least on one of the most emotionally charged topics: water.
For years, the public conversation around large-scale AI has been dominated by two numbers that are hard to ignore. First is energy: the electricity required to power racks of GPUs, the cooling systems that keep them from overheating, and the supporting infrastructure that makes the whole operation reliable. Second is water: not just the water used directly in cooling, but the broader reality that many regions rely on water-intensive cooling methods to keep power plants running and data centers stable.
Against that backdrop, Nvidia’s latest message is straightforward and provocative. In a blog post describing its “Rubin generation” reference design for a fully liquid-cooled data center, the company claims the approach can eliminate “pretty much all water usage,” while also cutting “massive amounts of power usage.” The pitch is that liquid cooling—done in a particular way, at scale, and integrated into the data center design—can reduce both the thermal burden and the need for water-heavy heat rejection.
But as with most technology announcements that touch the physical world, the interesting part isn’t only what Nvidia says it can achieve. It’s what the claim implies, what it doesn’t cover, and how it fits into the larger system constraints that determine whether an AI facility is sustainable in practice—not just in a lab or a marketing deck.
What Nvidia is claiming: liquid cooling as a system redesign, not a tweak
The key word in Nvidia’s framing is “reference design.” That matters because cooling isn’t simply a matter of swapping fans for pipes. A data center is a tightly coupled machine: the compute hardware generates heat; the cooling system removes it; the heat rejection method determines whether you need water; and the overall architecture affects efficiency, reliability, and even how quickly you can deploy new capacity.
In Nvidia’s view, the Rubin generation reference design pushes liquid cooling beyond incremental adoption. Instead of treating liquid cooling as an add-on for the hottest components, the design is presented as a more complete approach where liquid is used to move heat away from the compute more effectively. The company argues that this reduces the need for water-based cooling mechanisms that are common in many traditional setups, especially when facilities rely on evaporative cooling or other water-linked processes to handle peak loads.
The “runs hotter” phrasing that has circulated around the announcement is also part of the narrative shift. Higher operating temperatures can be a feature rather than a bug if the cooling method is designed to handle it safely and efficiently. In other words, Nvidia is positioning liquid cooling as enabling a different thermal envelope—one where the system can operate with less reliance on water-intensive heat rejection.
If the claim holds up across real deployments, it would be a meaningful change. Water use is often the limiting factor in where data centers can be built and how they can expand. Even when electricity is available, water constraints can slow projects or force operators into expensive mitigation strategies.
Why “pretty much all water usage” is a big deal—and why it’s also hard to verify from the outside
When a company says it can eliminate “pretty much all water usage,” it’s tempting to treat it as a near-binary outcome: either you use water or you don’t. In reality, water use in data centers exists on a spectrum. Some facilities use water only indirectly (for example, through power generation upstream). Others use water directly in cooling towers or evaporative systems. Still others use mostly air cooling but may still require water for certain components, maintenance cycles, or backup operations.
So the question becomes: what exactly does Nvidia mean by “water usage” in this context? Is it direct water consumption at the facility? Does it include water used for cleaning, humidification, or emergency modes? Does it assume a specific climate and a specific heat rejection strategy? And does it account for the full lifecycle of the facility, including construction?
Nvidia’s blog post, as summarized in coverage of the announcement, emphasizes the operational outcome: a liquid-cooled design that dramatically reduces water needs. That’s valuable information, but it doesn’t automatically answer the broader concerns that critics raise.
Those concerns tend to fall into three buckets:
1) Operational water use over time
Even if a design minimizes water during normal operation, what happens during peak demand, unusual weather, or equipment failures? Cooling systems are engineered for worst-case scenarios, and those scenarios can change the water profile.
2) Construction and supply chain impacts
Data centers are not just buildings; they’re massive infrastructure projects. Concrete, steel, electrical equipment, and specialized components all have embedded environmental footprints. Water use during construction can be significant depending on local practices and the scale of the project.
3) Electricity generation and upstream water
Even if a facility itself uses little water, the electricity it consumes may come from sources that use water for cooling at power plants. In many regions, that upstream water use is non-trivial. So “pretty much all water usage” at the facility level doesn’t necessarily mean “pretty much all water impact” overall.
This is why the claim is best understood as a targeted improvement in one part of the system. It’s a strong statement, but it’s not the final word on sustainability.
The power claim: reducing energy isn’t just about the GPUs
Nvidia also ties the design to power savings, saying it eliminates “massive amounts of power usage.” That’s another area where the details matter.
AI workloads are power-hungry, but the total facility power draw includes far more than the GPUs themselves. Cooling overhead can be a major contributor, especially when air cooling is inefficient at removing heat under high-density conditions. Fans, chillers, pumps, and heat exchangers all consume electricity. If liquid cooling can reduce the work required to move heat and maintain stable temperatures, it can improve overall efficiency.
However, there’s a subtlety that often gets lost in headlines: even if cooling becomes more efficient, AI demand tends to scale. Better efficiency can reduce the per-unit cost of computation, but it can also make it easier to justify building larger facilities. In other words, efficiency improvements can reduce intensity while total consumption still rises.
That doesn’t make the improvement meaningless—it just means the sustainability story depends on how the industry uses the technology. If liquid cooling enables more compute per watt without increasing water constraints, it could help operators meet growth targets in places where water is scarce. But if the industry responds by expanding capacity faster than efficiency gains, the net environmental impact could still be substantial.
The unique angle here is that Nvidia is addressing both constraints simultaneously: water and power. That combination is harder to achieve than focusing on one metric alone.
The trade-off nobody wants to ignore: cost and deployment complexity
One of the most practical questions raised by observers is cost. Liquid cooling can be more complex to build and operate than air cooling. It requires additional plumbing, leak detection, specialized components, and careful integration with rack design. It also changes maintenance procedures and potentially the failure modes operators must plan for.
Air cooling is comparatively straightforward: fans move air, and heat is rejected through air-side systems. Liquid cooling introduces a new layer of engineering: fluid dynamics, thermal interfaces, and the reliability of pumps and valves at scale.
Nvidia’s blog post, according to reporting, doesn’t emphasize the construction cost difference between a liquid-cooled facility and one that relies more heavily on air cooling. That omission matters because sustainability isn’t only about physics; it’s also about economics and incentives. If liquid-cooled designs are significantly more expensive to build, adoption may be slower, limited to certain customers or regions, or delayed until supply chains mature and costs drop.
There’s also the question of retrofits. Many existing data centers were built with air cooling assumptions. Retrofitting them for full liquid cooling can be difficult and disruptive. New builds are where the biggest gains are likely to appear, which means the water and power benefits may take time to show up at scale.
Still, the fact that Nvidia is presenting a reference design suggests it wants to standardize the approach. Standardization can reduce engineering uncertainty and speed deployment, which can indirectly reduce costs over time.
A “hotter” design: efficiency gains versus operational risk
The idea that the design “runs hotter” is intriguing because it challenges a common assumption: that cooler is always better. In data centers, lower temperatures can reduce stress on components, but they also require more aggressive cooling. Aggressive cooling can mean higher energy use and, depending on the heat rejection method, higher water use.
If liquid cooling allows higher inlet or operating temperatures while maintaining safe component performance, it can reduce the cooling system’s workload. That’s the efficiency logic.
But higher temperatures also increase the importance of monitoring and control. Thermal gradients, flow rates, and heat transfer performance must be consistent across racks and over time. Any variability can affect reliability. So the “hotter” claim is only as good as the system’s ability to manage heat precisely.
This is where reference designs can help: they can encode best practices for thermal management, redundancy, and control loops. But again, the proof will come from deployments, not from a conceptual design.
The bigger picture: water isn’t the only constraint—power availability is the bottleneck
Even if Nvidia’s design dramatically reduces water use, AI data centers still face a major constraint: power availability.
Large AI facilities require enormous electricity, and the process of securing that power—through grid upgrades, substations, and sometimes new generation—can take years. Coverage of the broader AI data center landscape has highlighted how power generation requirements and grid limitations can become a central challenge, sometimes even more immediate than cooling.
This creates a paradox. Cooling improvements can reduce the facility’s power draw, but the demand for AI compute continues to grow. Operators may still need to secure large power allocations to support the planned number of racks. In some cases, the cooling efficiency gains may help them fit within a given power budget, but they don’t eliminate the need for grid capacity.
So Nvidia’s water claim addresses one part of the sustainability debate