No Water, Less Energy, More Tokens: The Cooling System Inspired by Nuclear Reactors That the AI Industry Was Waiting For

The expansion of artificial intelligence increasingly weighs on global energy consumption. According to projections, by the end of the decade, data centers could absorb between 9 and 17 percent of the entire electricity production in the United States. Of this share, about one-third would be solely related to chip cooling.

It is in this segment that Ferveret, a startup aiming to apply principles borrowed from nuclear engineering to make server cooling significantly more efficient while eliminating water consumption, fits in. The company was founded by Reza Azizian and Matteo Bucci. Azizian conducted postdoctoral research at MIT in the field of nuclear engineering before transitioning to the tech industry: he worked on the development of Microsoft's HoloLens and later at NVIDIA. Bucci, for his part, remained in academia, becoming an Associate Professor at MIT's Department of Nuclear Engineering.

In 2017, Azizian first visited a data center and was struck by the inefficiency of the air cooling system. "A fifty-year-old technology that no one bothered to upgrade because it didn’t affect performance," he commented in the pages of MIT Technology Review. This observation sparked the idea that led to the founding of Ferveret in 2021.

The system developed by the startup, named Adaptive Phase Cooling (APC), falls under the category of immersion cooling: servers are immersed in a dielectric liquid that absorbs heat much more effectively than air. The physical principle behind it—known as subcooled boiling—was already established in the nuclear field, where managing heat extracted from the reactor core is a fundamental engineering parameter.

The difference compared to other immersion cooling solutions lies in the morphology of the bubbles generated at the chip's surface. The liquid used by Ferveret, free of PFAS substances (the so-called "forever chemicals" found in many traditional dielectric fluids), produces significantly smaller bubbles compared to competing systems. These bubbles detach more frequently from the surface and quickly re-condense into the surrounding liquid, accelerating heat transfer.

The result is a system that achieves superior cooling performance without bringing the liquid to a true sustained boil, a condition that in most existing plants introduces complexity in managing pressure, temperature, and fluid inventory.

From a structural integration standpoint, Ferveret opted for a modular approach: each unit is small, houses a single server, and mounts in standard racks. This is a significant departure from the large immersion tanks typical of traditional systems, which require substantial construction work and complicate maintenance. The system is complemented by software that monitors real-time internal temperature and pressure sensors in each unit, dynamically optimizing the power allocation to the servers.

In collaboration with UCLA’s Samueli Computer Science Department, Ferveret conducted a comparative study highlighting a 15 percent improvement in computational efficiency compared to state-of-the-art liquid cooling solutions. By combining this data with the dynamic optimization of its proprietary software, the company believes it can achieve 35 percent more tokens—the unit of measurement for throughput in language models—at parity with energy consumed.

On the commercial front, Ferveret is already conducting tests with CleanSpark (data center operator and developer), FuriosaAI (producer of AI accelerators), and Switch, one of the largest data center operators in the United States. The company is also part of the NVIDIA Inception program dedicated to tech startups and is negotiating with major cloud hyperscalers.

One of the most significant aspects of Ferveret's proposal is the complete lack of water consumption. Modern data centers are among the largest industrial users of water, necessary for evaporative cooling systems. This dependency geographically limits the location of facilities, effectively excluding arid regions or areas under water stress, despite these places offering optimal conditions for harnessing solar or wind energy.

"The sun shines in places where water is scarce," noted Bucci. "Being completely water-free means being able to build data centers where there are renewable energy sources, but not the water resources necessary for traditional cooling systems. Think about Africa, the Middle East, but also many areas of the United States."