Ferveret’s Nuclear Cooling Wins 15% More Compute per Watt in UCLA Test

Ferveret, a San Jose startup founded by two former MIT nuclear researchers, says its waterless server-cooling system lifted compute efficiency by 15% in a benchmark run with UCLA’s computer science department, a result that, if it holds outside the lab, would let AI data centers squeeze more training work out of the same power supply.

The result was published this March by UCLA’s Samueli Computer Science Department using Ferveret’s Adaptive Phase Cooling (APC) hardware and NVIDIA H200 graphics processing units. The benchmark measured server-level output in teraflops per kilowatt and found APC beat direct-to-chip liquid cooling, the workhorse of modern AI infrastructure, by 15%. At the data center level, Ferveret said the system posted a 1.03 power usage effectiveness ratio, meaning about 97% of facility power reached computing.

• 15% server-level TFLOPs/kW gain over direct-to-chip liquid cooling, per the UCLA benchmark
• 1.03 power usage effectiveness ratio at facility level
• Roughly 35% more compute per same power envelope, a modeled outcome
• Zero water consumption at the cooling stage
• Reference platform: NVIDIA H200 GPUs

What UCLA Found in the Benchmark

The headline number came from the UCLA benchmark on adaptive phase cooling, run by the university’s Intelligent Connectivity Laboratory in partnership with Ferveret. The study used NVIDIA H200 GPUs as the reference platform and measured output in teraflops per kilowatt, a metric focused on graphics processor performance rather than facility overhead. A separate facility figure, the 1.03 power usage effectiveness ratio, reflects how much of a facility’s total power reaches compute rather than cooling, lighting, and other overhead.

Reza Azizian, Ferveret’s chief executive, framed the result as a way to “extract more compute from the same power envelope.” Omid Abari, an associate professor in UCLA’s computer science department, went further, tying the cooling to model training itself.

Our recent study shows that Ferveret cooling reduces the time required to train machine learning algorithms by enabling hardware to operate at higher sustained clock speeds. In other words, Ferveret not only provides a more efficient thermal solution but also delivers better performance, resulting in shorter training times.

Abari said in a statement released with the study in March 2026.

How Subcooled Boiling Reaches the Chip

Ferveret’s technology adapts a process used inside nuclear reactors called subcooled boiling. The system submerges a server in a specialized liquid with a low boiling point and no PFAS “forever chemicals.” Where most two-phase immersion systems rely on saturated boiling, which produces larger bubbles that pool in a chamber before being recondensed, Ferveret’s liquid forms much smaller bubbles at the chip surface. Those bubbles detach more quickly and recondense in the surrounding fluid, refreshing the surface and accelerating heat transfer.

“When liquid is boiling, it becomes even better at removing heat because the phase change requires a lot of energy, which is the energy you remove from the chip,” Ferveret co-founder and chief technology officer Matteo Bucci told MIT News. The setup lets the system move large quantities of heat with small temperature differences between the chip and the surrounding liquid, the same physics Azizian and Bucci studied inside reactors in their earlier careers.

The Small Box, Not the Big Tank

Most immersion cooling systems ship as large tanks that hold many servers at once. Ferveret’s hardware ships in smaller modular rack-mounted boxes, each housing one server.

That form factor is the piece Azizian says changes the deployment math. “The physics enable us to get to form factors that weren’t possible in the past,” he told how Ferveret’s founders moved from nuclear reactors to data centers. “Most immersion cooling solutions are large tanks that people submerge the servers in. We have a smaller, modular rack-mounted solution that makes it adaptable to the current infrastructure, so it’s easier for people to deploy our technology.”

Pilot Customers and the Hyperscaler Question

Ferveret is running pilots with CleanSpark, a data center developer and operator, FuriosaAI, an AI accelerator company, and Switch, one of the largest data center operators in the United States, per a June 10, 2026 profile in MIT News. The company is also part of NVIDIA’s Inception program for startups. The same week, Applied Digital signed a $5.2B hyperscaler lease for Delta Forge 2, a reminder of the kind of infrastructure contracts the cooling business is feeding into.

Beyond the named pilots, Ferveret says it is in talks with the large cloud computing firms known as hyperscalers, the same customers accounting for the bulk of new AI capacity. Asked about deployment status in an April 29, 2026 interview, Azizian placed the company squarely in pilot mode in Ferveret’s CEO interview on cooling and power. “We have done a lot of testing at the server level, and now we are doing pilots at the rack level with different customers. We are running pilots in data center environments, but it’s not in full production,” he said. That same article described Ferveret’s 35% more-compute claim as “based on testing and modeled scenarios rather than full production deployments.”

1. CleanSpark, a data center developer and operator
2. FuriosaAI, an AI accelerator company
3. Switch, one of the largest data center operators in the U.S.

The Power and Water Squeeze the Tech Is Built For

The pitch arrives as data center operators hit two walls at once, on power and on water. Ferveret, citing US Department of Energy figures, says US data centers today consume up to 4.5% of total US electricity production and that share is forecast to rise to 12% by 2028. A separate MIT News report put the longer-range projection at 9% to 17% of US electricity by the end of the decade. Around a third of data center electricity is devoted to cooling the chips that run AI models, and Azizian has put the cooling share at “roughly about 20% to 30%” in a typical facility.

On the water side, Ferveret’s own site says US data centers use 230 billion gallons of water a year under current cooling approaches and that consumption could exceed 700 billion gallons a year by 2030. That has put pressure on host communities, with environmental advocates and local residents pushing operators to design around the issue, a thread that a recent report on AI data center water demand tracks in detail.

Ferveret argues its waterless design lets operators build in places where renewable power is abundant but water is scarce. “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 told MIT News. The company has already drawn investor backing for that pitch. Backers include TO VC, Aramco Ventures, Cerberus, Y Combinator, Baruch Future Ventures, Verso Capital, Acclimate Ventures, Cathexis Ventures, Valkyrie, E14, and Climate Capital, according to a March 2026 statement.

For Ferveret, the supply-constrained geography is the opening. The MIT News profile said the company “plans to announce expanded partnerships later this year.”

The Founders’ Path From Reactor to Rack

Reza Azizian was a postdoc at MIT in 2013 when he met Matteo Bucci, then a research scientist. The two worked on heat transfer in nuclear reactors before Azizian left for industry. He cooled chips first on Microsoft’s HoloLens augmented reality headset and then joined NVIDIA. He walked into his first data center in 2017 and was struck by the rows of fans still doing the cooling work. “Holy crap, this is not how you cool facilities,” he told MIT News, noting that air cooling can still take up 40% of the power going into a data center. Bucci stayed at MIT, becoming an assistant professor in 2016, and now holds the Esther and Harold E. Edgerton Associate Professor title in the Department of Nuclear Science and Engineering, per Ferveret’s founding team page.

The two started Ferveret in 2021, were accepted into Y Combinator that same year, and built out an innovation factory in El Paso, Texas, by 2024. The startup’s seed funding closed in 2025, backed by a City of El Paso economic development package. Azizian, who serves as CEO, told Data Center Knowledge the company is now working with original equipment manufacturers and original design manufacturers to integrate the system at hyperscale. “The next test,” Data Center Knowledge wrote, “will be whether those gains hold at scale.”

What’s Still Unproven

The 15% number rests on a single academic benchmark, not on a fleet of operating data centers. The higher “roughly 35% more compute per same power envelope” figure Azizian cites is a modeled outcome, not a measured one. In the Data Center Knowledge interview, Azizian broke the 15% down himself: roughly 4% to 5% comes from running the chip cooler, and the other roughly 10% comes from removing server fans, gains that depend on Ferveret’s hardware being deployed without the fans the rest of the industry still uses.

Capital intensity is also unproven at scale. Ferveret says its base cost is “very comparable to direct-to-chip” liquid cooling and that operators can save on chillers and cooling towers, but the company has not disclosed capex numbers for a full hyperscale deployment. Ferveret’s seed funding closed in 2025, and the company has not announced a Series A.

For now the technology lives in pilots. Whether the smaller-bubble physics travels from a UCLA lab to a hyperscale hall is the question the rest of 2026 will answer.

Frequently Asked Questions

What is Ferveret?

Ferveret is a San Jose-based cooling startup founded in 2021 by former MIT nuclear researchers Reza Azizian and Matteo Bucci. Its Adaptive Phase Cooling (APC) hardware submerges servers in a waterless liquid and uses subcooled boiling, a heat transfer technique borrowed from nuclear reactors, to draw heat away from chips faster than conventional immersion cooling.

What did the UCLA study actually measure?

Researchers at UCLA’s Intelligent Connectivity Laboratory ran Ferveret’s APC system with NVIDIA H200 GPUs and measured 15% more computational output per kilowatt than direct-to-chip liquid cooling, the established approach for AI servers. They also recorded a 1.03 power usage effectiveness ratio at the data center level, meaning about 97% of facility power reached computing.

Is Ferveret’s cooling system in production today?

No. CEO Reza Azizian said in an April 2026 interview that Ferveret is “doing pilots at the rack level with different customers” and that deployments are “not in full production.” Named pilot partners include CleanSpark, FuriosaAI, and Switch.

How is Ferveret different from other immersion cooling companies?

Established immersion systems from GRC, Submer, and LiquidStack typically ship as large tanks that hold many servers at once. Ferveret ships smaller, rack-mounted boxes, one server per unit, and uses subcooled boiling, which produces smaller bubbles that detach and recondense faster than the saturated boiling used in conventional tanks.

How much water and power could a data center save with Ferveret?

Ferveret’s system uses no water at the cooling stage and posts a modeled 1.03 PUE, meaning about 97% of facility power reaches computing. Azizian has said operators can pull roughly 35% more compute out of the same power envelope once server-level gains and facility-level PUE are combined, though that figure is a modeled scenario, not a measured production result.