An engineer walking through a modern AI data hall today might notice something odd: the cooling loops can carry water at temperatures warm enough to feel surprising by traditional data-center standards, yet the system can operate reliably at those higher temperatures. That small detail captures a genuine reversal in how the industry thinks about heat. For years, data center operators chased colder temperatures as a proxy for reliability, pumping chilled water through server racks and treating every stray degree as a threat to uptime. That instinct now faces a challenge from inside the industry, as chipmakers and cooling engineers design the next generation of AI hardware.

Heat Stops Being the Enemy

Nvidia’s Rubin platform, introduced at CES 2026, supports liquid-cooling systems with inlet temperatures of up to 45 degrees Celsius, warmer than a typical hot tub. That number matters less as a standalone specification and more as a signal of the industry’s deliberate move toward warmer cooling loops. Andrew Chien, a computer science professor at the University of Chicago who studies data center efficiency, has highlighted the significance of this shift toward warmer cooling temperatures and the ability to reject heat without relying as heavily on mechanical refrigeration. Rubin’s architecture inverts that assumption by using warmth as a functional feature rather than a byproduct to suppress.

The mechanics explain why this works. Direct liquid cooling can transfer heat far more effectively than air because liquid has vastly greater volumetric heat-transfer capacity, with industry sources commonly citing a difference of roughly three thousand times in heat-carrying capacity by volume. That efficiency gives engineers room to raise the inlet temperature of the cooling loop while keeping the processors within their validated operating limits. Once the facility loop runs warm enough, facilities in suitable climates can reject heat into the outdoor air using large dry coolers, reducing or bypassing the need for energy-intensive mechanical chillers for much of the year.

The Chemistry Behind the Comfort

None of this happens by accident. Rubin’s cooling system can use propylene-glycol-based fluid, with NVIDIA describing PG25 as a common option for closed-loop systems designed around its rack architecture. Motivair, the advanced cooling division of Schneider Electric that has worked alongside Nvidia’s product roadmap for nearly a decade, has noted that liquid cooling became necessary as chip power densities reached levels where air cooling no longer offered a viable option. Richard Whitmore, the company’s president and chief executive, has pointed out that a persistent myth in the industry treats a cold data center as an efficient one. Warm-water cooling directly challenges that myth by showing that efficiency and temperature are not the same variable.

Other cooling companies are testing systems that can operate at even higher facility-water temperatures. Accelsius, which builds two-phase cooling systems, has reported independent third-party testing showing its technology can maintain comparable thermal performance at facility-water temperatures up to 54 degrees Celsius, or up to 59 degrees when facility-water flow increases from 1.5 to 3.0 liters per minute. Two-phase cooling works by letting a coolant boil and condense inside a sealed loop, a process that absorbs far more heat per unit of fluid than a simple temperature rise ever could. Whether this approach becomes standard or remains a specialized solution for the hottest racks, the results show that higher facility-water temperatures are technically achievable as cooling architectures evolve.

Why the Grid Should Pay Attention

The energy implications reach beyond any single data hall. Mechanical chillers consume enormous amounts of electricity, and removing them from the equation, even partially, changes the load profile that utilities must plan around. A facility that relies mainly on outdoor air for heat rejection behaves differently on the grid than one that runs compressors around the clock. This is a design decision with consequences for facility electricity demand, not merely a footnote about pipe temperature. As AI workloads scale, the industry’s willingness to accept warmer operating conditions could reduce the share of facility electricity devoted to cooling and leave more of the available power budget for computation.

There is also a control problem buried inside this shift. Warmer water loops can tighten the operating margin available to the cooling system, since changes in workload, weather and heat-rejection performance become more consequential as supply temperatures rise. Firms like etalytics, which build digital twin systems for data center operations, argue that warm-water cooling does not just change the hardware; it changes what counts as safe operation in the first place. Static, rule-based control systems built for a world of generous cold-water buffers may become less effective when AI workloads ramp rapidly and the cooling system must respond to changing thermal, weather and power constraints.

What Comes Next

The open question is how far this logic extends. If 45 degrees Celsius counts as an efficiency milestone today, the industry will likely test higher thresholds tomorrow, provided materials science and control systems can keep pace. That trajectory suggests future AI infrastructure may earn judgment less by how cold it can run and more by how effectively it can manage and reject heat while maintaining validated performance and reliability. The AI industry has faced growing scrutiny over its water use and energy demand. Warm-water cooling will not resolve every environmental question tied to AI infrastructure, but it does reveal something instructive about where engineering can reduce the resources required to operate increasingly powerful systems.