Ferveret’s Nuclear Cooling Lifts AI Data Center Efficiency 15 Percent

Ferveret, a startup founded by two MIT researchers, says its nuclear-inspired cooling system lifted AI data center compute efficiency by 15 percent in tests run with the University of California, Los Angeles. The result, reported this month, comes as data centers face mounting pressure to cut both power and water use. Ferveret, founded in 2021, is piloting the technology with CleanSpark, FuriosaAI, and Switch.

The company’s Adaptive Phase Cooling (APC) system submerges servers in a specialized liquid that absorbs heat more efficiently than air. What sets it apart from other liquid cooling approaches is the bubble behavior at the chip surface: Ferveret’s liquid produces smaller bubbles that detach more frequently, accelerating heat transfer. The technique is adapted from subcooled boiling, a process used in nuclear reactors. By combining the cooling gain with its power control software, the company says data centers can get 35 percent more tokens, the small pieces of text or data that AI models produce, from the same amount of power. The system uses no water, a detail that could matter as much as the efficiency numbers for data centers built in arid regions.

The Cooling Bottleneck Behind AI’s Power Bill

Data centers are on track to account for 9 to 17 percent of total U.S. electricity by the end of the decade, according to a June 2026 MIT News report on the technology. The industry has become power-limited, with cooling consuming a growing share of the energy budget. Ferveret co-founder Reza Azizian, who spent years at Nvidia watching GPUs push past the limits of fan-driven designs, argues that air cooling is a 50-year-old technology that no one updated because it wasn’t hurting performance.

The independent comparison, the UCLA study, measured APC against other liquid cooling systems, a higher bar than air cooling. The cooling industry has been moving toward liquid for years, driven by chipmakers packing more transistors onto silicon and pushing power densities past what fans can handle. Two-phase immersion cooling, the most effective liquid approach, is gaining traction, though single-phase cold plate cooling still leads new deployments. Ferveret’s technology page claims even bigger gains relative to air cooling, the baseline most data centers still use.

The startup is entering a market shaped by the same forces that made its founders’ nuclear engineering background relevant again. Demand for compute is climbing faster than the power grid can expand, and every percentage point of cooling efficiency translates directly into more AI tokens per watt. Ferveret’s pitch is that a technique borrowed from one industry can crack a bottleneck in another.

• ~1/3 of data center electricity devoted to cooling AI chips
• 40 percent of data center power consumed by air cooling
• +55% TFLOPs/W improvement vs. air cooling (Ferveret claim)
• 2x chip power enabled vs. air cooling (Ferveret claim)

A Reactor Trick, Repackaged for Server Racks

Azizian was a postdoctoral researcher at MIT in 2013 when he met Matteo Bucci, then a research scientist in the same nuclear engineering group. They worked on heat transfer in nuclear reactors, where the physics of boiling have been studied for decades. Azizian left for industry, first working on Microsoft’s HoloLens augmented reality headset, then joining Nvidia, which makes the GPUs that train and run the latest AI models. Bucci stayed at MIT, becoming an assistant professor in 2016.

Azizian walked into his first data center in 2017 and was struck by the rows of fans. “I thought, ‘Holy crap, this is not how you cool facilities,’” he recalls. “It was not an efficient way of doing things, but since it wasn’t hurting the performance, no one cared that the cooling technology was 50 years old.” He started talking with Bucci about applying what they had learned in nuclear engineering to the data center problem. Heat transfer, Azizian explains, “determines how much energy you can extract from the reactor core, which translates directly to revenue.” The same logic, he argues, applies to AI chips: the better the cooling, the more useful work a chip can do per watt.

Liquid is a better heat transfer medium than air, which is why room-temperature water still feels cold on your skin. When liquid boils, it gets even better at pulling heat away, because the phase change from liquid to vapor absorbs a large amount of energy. The catch is that boiling liquid is messy: operators have to capture and reliquefy the vapor, control pressure and temperature, and manage fluid levels. Most two-phase immersion systems today are large tanks that house many servers, and they rely on a process called saturated boiling, which produces relatively large bubbles that rise away from the chip and collect in a vapor plenum.

Ferveret’s system borrows a different process from nuclear reactors, called subcooled boiling. The company uses a dielectric liquid with a low boiling point and no PFAS, the toxic “forever chemicals” found in some competing fluids. At the chip surface, the liquid produces smaller bubbles than conventional two-phase systems. Those bubbles detach more frequently and recondense quickly in the surrounding liquid, creating a rapid cycle that keeps refreshing the liquid at the chip surface and hastens heat transfer.

Why Conventional Boiling Falls Short

Conventional two-phase immersion cooling works by boiling a dielectric fluid in direct contact with the chips. The phase change pulls large amounts of heat away, which is why two-phase is more effective than single-phase liquid cooling. But the bubbles that form are large and slow to detach, which means the chip surface spends time covered in vapor, a poor conductor. Operators must manage a vapor plenum above the fluid, maintain precise pressure and temperature, and handle fluid loss as vapor escapes. The systems come in large tanks that hold many servers, which makes maintenance disruptive: lifting a server out means draining and disturbing neighboring systems.

Ferveret argues that subcooled boiling sidesteps most of those problems. The liquid stays below its boiling point in the bulk, so bubbles form only at the hot chip surface and recondense before they can rise. That means no vapor plenum, no pressure-tight tank, and far less fluid to manage. The company delivers its system in small boxes, each housing one server, which can be slotted into standard liquid-cooled rack footprints. “The physics enable us to get to form factors that weren’t possible in the past,” Azizian says.

The 15% Efficiency Gain and the Partners Testing It

In a study with UCLA’s Samueli Computer Science Department, Ferveret tested its APC system against state-of-the-art liquid cooling solutions. The result: a 15 percent improvement in computational power efficiency, meaning the same power budget produces more useful work. When the cooling gain is combined with Ferveret’s power control software, which adjusts power delivery to each server in real time, the company says data centers can extract 35 percent more tokens from their AI models. Tokens, the small pieces of text or data that large language models process and generate, are the standard unit of useful output for AI systems. “Our goal is to make data centers as sustainable as possible and help them use every single watt of power to generate tokens, which are the most useful outputs,” Azizian says.

The company is already running pilots with three named partners.

• CleanSpark, a data center developer and operator
• FuriosaAI, a Korean AI accelerator company
• Switch, one of the largest data center operators in the U.S.

Ferveret is also part of Nvidia’s Inception program, which supports startups building on Nvidia’s platform. The company is in talks with hyperscalers, the large cloud computing companies that operate the bulk of the world’s AI infrastructure, and plans to announce expanded partnerships later this year. Bucci frames the offering as more than a cooling box: “We deliver full-stack systems that include the cooling box, the rack, the cooling distribution units, and sensors that measure the temperature and pressure.” Software monitors those sensors and adjusts operating conditions inside each box to minimize energy use.

The company plans to quickly scale the technology as the AI industry continues to grow. Azizian’s framing is that the constraint on AI growth is not just chip supply or grid capacity, but the cooling system’s ability to turn available power into useful work. Ferveret’s pitch is that it can do that more efficiently than any liquid system tested so far.

Zero Water, New Geography

The system uses zero water, and Bucci argues that detail matters as much as the efficiency numbers. Most data center cooling relies on water, either directly through evaporative cooling towers or indirectly through the water footprint of the electricity grid. In water-scarce regions, that constraint has become a binding limit on where new data centers can be built. Ferveret’s dielectric liquid stays in a closed loop and never evaporates into the atmosphere.

The sun shines in places where you don’t have much water, so the advantage of us being water-free is we allow you to build data centers where you have solar energy but nothing to cool the data center down.

Bucci, in an interview with MIT News, pointed to Africa, the Middle East, and parts of the American Southwest as regions where solar power is abundant but water is scarce. Building AI data centers in those locations would pair cheap renewable energy with a cooling system that does not depend on a resource those regions lack. The constraint on AI growth may not be power or chip supply, but the physical resources that current cooling systems require. Growing concern over how AI data center water use is straining local supplies has made water-free cooling more than an efficiency feature.

The water-free angle also changes the calculus of where to build. A system that draws no water and runs quietly removes one of the most common objections to new data center projects. That is a siting and permitting advantage, and it could matter as much as the efficiency gain when operators choose where to build their next facility.

From Pilot to Hyperscaler

Ferveret is one of several startups and established suppliers chasing liquid cooling for AI data centers. Single-phase cold plate cooling still leads in deployed systems because it is simpler and cheaper to install, but two-phase immersion is gaining ground as power densities climb past what air cooling can handle. Ferveret’s bet is that its modular, subcooled approach removes the complexity that has kept two-phase immersion from going mainstream. Some operators are also attacking the power constraint from a different angle through demand-response programs that pay households to cut power use during peak hours. The founders say expanded partnerships will be announced later this year.

The immediate test is commercial. Pilots with CleanSpark, FuriosaAI, and Switch will show whether the efficiency gain holds up in production environments with real AI workloads. Talks with hyperscalers, the companies that buy cooling at the largest scale, will determine whether the startup graduates from niche supplier to industry standard. “The main goal for these data center operators would be to get more tokens from the power they have,” Azizian says. “We’ve shown we can do that.”