NVIDIA's latest AI chip platform, called Rubin, cools itself entirely with liquid instead of the air-and-liquid mix that data centers have used for years. That shift has real consequences for how companies build and run the massive facilities that power AI services.

Here is the key detail: Rubin's cooling system works at 45°C (about 113°F), which is much warmer than the chilled water that older systems required. To put that in perspective, think of it like this: a typical data center used to need to actively refrigerate its cooling water down to about 70°F, burning expensive energy to keep that water cold all year. Rubin can use water that is closer to hot-tub temperature, which means facilities can often skip the refrigeration step entirely and just let natural cooling towers handle the heat—saving enormous amounts of money and electricity.

A Completely New Approach

For the past ten years or so, data center cooling has been a compromise. The main AI chips themselves got dunked in cold liquid to stay cool, but all the supporting hardware around them—the networking chips, power supplies, and cables—stayed cooled by regular air blowing through fans. This created a weak link: no matter how well you cooled the main chips, the air-cooled parts prevented you from packing equipment as tightly together as you wanted.

Rubin changes that. Every component in the rack—the processors, the network switches, the power distribution boards, all of it—now gets cooled with liquid. According to NVIDIA's announcement, this unified approach eliminates the thermal bottleneck that has limited how dense data centers could be.

It is worth putting the progression in context. NVIDIA's previous generation, Blackwell, already cut water use by 300 times compared with older chips and squeezed out 12% more computing power while actually using less energy. Rubin starts from that baseline and goes further.

Water and Energy: The Full Picture

When NVIDIA talks about cutting water use, they mean the liquid coolant pumped through pipes inside the data center building. That is real and significant. But it is not the whole story, because data centers also consume huge amounts of water indirectly when power plants generate the electricity they use—and that depends on where the power comes from and what fuel the local grid relies on. As TechCrunch explained on 22 June 2026, Rubin solves the on-site cooling side but does not change the electricity side of the equation.

So when you hear sustainability claims about Rubin, it is fair to note that they apply to direct water usage in the facility itself, not the complete environmental footprint.

When Rubin Arrives and What It Means

Rubin went into production in January 2026, when NVIDIA announced it at a major tech conference. In June, they published the full technical details confirming the all-liquid design and 45°C operating temperature. Data center teams that are planning to buy Rubin equipment now have clear specifications to work with.

For the companies running these facilities, the shift is substantial. A facility that can run Rubin at 45°C can skip mechanical chillers during much of the year in cooler climates, which cuts power consumption and the need for expensive cooling infrastructure. That is huge for operating costs.

But there is a catch. If you want to use Rubin, you cannot half-do it with your older facilities. The platform requires complete liquid cooling infrastructure from the start. Companies with older buildings that still use mostly air cooling now face a choice: fully commit to building out liquid infrastructure, or stick with previous-generation chips. There is no halfway option.

The big picture here is that NVIDIA is treating cooling as central to chip design rather than as something to patch in later. As AI chips get more powerful and burn more electricity, that priority makes sense. The real question now is whether the companies that build and operate data centers can keep up with these infrastructure changes as the technology keeps evolving.