If you’ve used ChatGPT or Claude, you know it types an answer back, like a very smart research assistant. But a newer kind of AI does more than just talk. These advanced AI systems don’t just generate text; they do things: analyze complex data, optimize supply chains, or even design new materials.
These powerful AI systems need incredibly powerful computers to run. Think rooms full of specialized chips, called AI servers, working non-stop. And here’s the thing: those chips generate a tremendous amount of heat.
Keeping these “AI factories” cool has become a massive, expensive problem. It’s not just a technical headache; it’s a major drain on budgets and a huge environmental concern. Until recently, cooling alone could eat up to 40% of a data center’s total electricity.
Why we thought “colder is better” for AI servers
Honestly, this confused me for years. Most people, myself included, have this picture of a data center: a room so cold you need a jacket, like a walk-in freezer. Decades ago, if a data center didn’t feel like that, you’d assume something was wrong. So, for a long time, the industry pushed vast amounts of chilled air through these facilities to keep everything frosty.
But that instinct was incomplete.
The silicon chips inside computers can actually handle much warmer environments than we typically provide. The real challenge isn’t keeping the room cold. It’s getting the heat off the chips themselves, directly and efficiently.
The old way, using massive air conditioners and huge fans to blast cold air everywhere, was incredibly wasteful. It’s like trying to cool a boiling pot of water by blowing a fan across the kitchen, rather than putting it directly in the sink with cold water.
The counter-intuitive breakthrough: 45°C liquid cooling
Here’s the weird part. NVIDIA, a company at the heart of AI hardware, found a way to make AI servers run more efficiently by using cooling liquid that’s actually warmer than a hot tub. Yes, you read that right. While hot tubs usually sit around 38 to 40 degrees Celsius (about 100-104°F), NVIDIA’s newest AI servers can run their cooling liquid at up to 45 degrees Celsius (113°F).
This isn’t some minor tweak. It’s one of the biggest leaps in data center efficiency we’ve seen. The NVIDIA Rubin generation, released in 2026, is the first AI infrastructure to achieve 100% liquid cooling. Every single chip and networking component is cooled entirely by liquid, in a closed loop, with no noisy fans anywhere in the system.
Think of it like this: your car engine gets hot, right? Instead of blasting it with cold air from a giant fan, your car uses a radiator (a special heat exchanger) and coolant (a liquid mixture, often water and antifreeze) to draw heat directly from the engine. That warm coolant then flows to the radiator, which uses outdoor air to cool it down, cycling it back to the engine.
NVIDIA’s system works similarly. Cold plates (metal blocks with channels for liquid) sit directly on the hot chips, absorbing their heat. This liquid, a mix of 75% water and 25% propylene glycol, then flows through the system. Running this coolant at a relatively warm 45°C means that in many climates, the data center can simply use large outdoor dry coolers (just like big car radiators) to reject the heat into the outside air.
*AI servers send hot coolant to a coolant distribution unit, which then sends the hot coolant to an outdoor dry cooler.*This means they don’t need to turn on huge, energy-guzzling mechanical chillers (refrigeration units that super-cool water) or those loud fans for much of the year. The liquid coolant entering a chip at 45°C exits at roughly 55°C, having absorbed all that heat, but the chip still runs at full speed because the liquid cooling keeps it within its safe operating temperature.
Why warmer coolant saves millions
For business leaders, this isn’t just a clever engineering trick; it’s a significant financial and environmental win.
First, the energy savings are huge. Historically, cooling could account for up to 40% of a data center’s electricity bill. Industry estimates suggest that raising the temperature of the water in a chiller system by just one degree Celsius can cut cooling energy costs by about 4%. By operating at 45°C, these systems can often bypass chillers entirely.
Then there’s water. Traditional cooling towers (which often work with chillers) rely on evaporation to cool water, consuming staggering amounts. A conventional data center might use around 2.6 million gallons of water per megawatt of power per year. With NVIDIA’s 45°C liquid cooling and dry coolers, that water consumption drops to near zero, as the system is a closed loop that’s filled once and then simply recirculates its coolant.
The financial impact is substantial. For a large hyperscale facility (a massive data center, like those run by major cloud providers) consuming 50 megawatts of power, the shift to liquid-cooled infrastructure can save over $4 million annually in cooling-related energy and water costs. That’s real money that hits the bottom line.
*The new 45°C liquid cooling dramatically reduces water consumption compared to conventional cooling, and drastically cuts the overall power consumed by cooling in data centers.*Beyond the numbers, there are other benefits. Imagine walking into an AI factory that’s almost silent. Traditional data centers, with all those fans, can be loud enough (85 decibels or more) to require ear protection. Liquid cooling eliminates that noise. It also takes up significantly less physical space, freeing up valuable real estate within the data center. And, in some cases, the residual heat from the AI factory can even be recovered and repurposed to heat nearby buildings.
The catch: geography matters
Now, there’s always a catch. While 45°C liquid cooling offers massive benefits everywhere, its full “chiller-less” potential depends on where the data center is located. A facility in the Scottish Highlands will have a much easier time rejecting heat through dry coolers than one in Phoenix, Arizona.
Even in warmer climates, though, the shift to 45°C coolant is a huge step forward. It means mechanical chillers might only need to run a few days a year when outside temperatures are truly extreme, rather than constantly. That still translates to significant energy and cost savings.
A new industry standard
The fact that NVIDIA’s Rubin platform integrates 100% liquid cooling means that every major cloud provider and data center operator building for it is making this transition. This isn’t just a niche solution; it’s becoming the new standard. Companies like Motivair, a cooling division of Schneider Electric, have worked with NVIDIA for years, confirming that once chip power densities crossed a certain point, air cooling simply wasn’t a viable option anymore. Liquid cooling became mandatory.
This isn’t just about making AI run faster. It’s about making the entire infrastructure that powers AI sustainable, cost-effective, and efficient. And that matters for every business considering how to build or use AI.
Sources: blogs.nvidia.com