Ferveret Turns Nuclear Bubble Physics Into Extra AI Tokens

Ferveret’s Adaptive Phase Cooling posted a 15% gain in TFLOPs per kilowatt over direct-to-chip liquid cooling in a UCLA test on Nvidia H200 GPUs. The San Jose startup, built by two nuclear engineers, is selling that gain as extra AI output from the same power hookup, with no water in the cooling loop.

The UCLA number is a server-level result. Ferveret then stacks a lower facility overhead and says the two layers yield about 35% more compute from a fixed power envelope. That combined figure is the company’s math, not the university’s, and it is the claim operators now take into rooms where megawatts and water permits decide whether a hall gets built.

UCLA Measured 15% More TFLOPs per Kilowatt

The March 26, 2026 study ran with UCLA’s Intelligent Connectivity Laboratory, the ICON Lab inside the Samueli computer science department. Researchers used Nvidia H200 GPUs as the reference platform and scored server output in teraflops per kilowatt against direct-to-chip liquid cooling, the workhorse method for dense AI racks.

Omid Abari, an associate professor of computer science at UCLA, tied the thermal result to training time rather than to a greener scorecard.

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.

Omid Abari, Associate Professor, UCLA Computer Science Department

Ferveret, in the same release, put facility-level power usage effectiveness at 1.03, or about 97% of site power reaching compute. Reza Azizian, the company’s chief executive, said the design “eliminates the need for water while dramatically improving computational efficiency, allowing customers to extract more compute from the same power envelope.”

THE BENCHMARK STACK

Azizian later walked through the 35% arithmetic on X: 15% at the server from heat transfer, then about 17% at the facility from less cooling overhead. Those are different layers. Anyone treating 35% as the UCLA result is mixing them.

How Ferveret Cools Chips Without Water

Adaptive Phase Cooling submerges a server in a dielectric liquid that boils on the hot surface, then collapses those bubbles back into the bath before they collect as vapor. The company runs that cycle near atmospheric pressure, so the box does not need a pressure vessel or a vapor plenum. Heat leaves through a closed loop and, at a 45 C inlet, can go to dry coolers instead of chillers and cooling towers.

The physics come from subcooled boiling in nuclear reactor cores. Typical two-phase immersion uses saturated boiling: larger bubbles grow, detach, and rise into a vapor space that then has to be condensed. Ferveret says its liquid makes much smaller bubbles that detach more often and recondense in the surrounding cooler fluid, which refreshes liquid on the chip and moves more heat for a given temperature difference.

Matteo Bucci, Ferveret’s chief technology officer and the Esther and Harold E. Edgerton Associate Professor in MIT’s Department of Nuclear Science and Engineering, has put the basic point in kitchen terms. Liquid beats air at carrying heat; boiling liquid beats still liquid because the phase change soaks up energy. The hard part, he said, is capturing and reliquefying bubbles while holding pressure, temperature, and fluid inventory. Subcooled boiling is the company’s answer to that mess.

Azizian met Bucci at MIT in 2013, when Azizian was a nuclear-engineering postdoc and Bucci was a research scientist. Azizian later designed thermal hardware for Microsoft’s HoloLens, then worked as a packaging thermal engineer at Nvidia and as a thermal design engineer at Cruise. Bucci spent nearly a decade at CEA Paris before the MIT faculty post. They founded Ferveret in 2021 and went through the Y Combinator Summer 2021 batch.

Azizian has said heat transfer in a reactor core is not a side chore. It sets how much energy you can pull from the core, which is revenue. The same logic now sits on a GPU board: the watt that does not run a fan can run a token, the small piece of text or data a model emits.

The Megawatt Cap Turns Cooling Into Capacity

Lawrence Berkeley National Laboratory’s 2025 update of the United States Data Center Energy Usage Report, published in June 2026, puts a reference case at 11.8 percent of U.S. electricity by 2030. The same study’s range is 9.5 to 15.3%, or 521 to 843 TWh, with a reference of 649 TWh. A 2026 Department of Energy study summarized by the Congressional Research Service put use at 192 terawatt-hours in 2024, about 4.7% of U.S. consumption.

THE POWER ENVELOPE

• 2024 load: About 192 TWh, or 4.7% of U.S. electricity, in the DOE study the Congressional Research Service cites.
• 2030 reference: 649 TWh, or 11.8% of U.S. electricity, in the Berkeley Lab update.
• 2030 bounds: 521 to 843 TWh, equal to 9.5 to 15.3% of U.S. electricity.
• Cooling share: Azizian puts cooling at 30 to 40% of facility energy in a typical hall, and 10 to 20% even in many direct-to-chip designs.

The 2024 Uptime Institute global survey, which Azizian cites, put average power usage effectiveness at 1.56. Direct-to-chip still often wants facility water near 30 C or below to stay inside ASHRAE envelopes, which in hot climates means chillers. Jensen Huang said at CES in January 2026 that Nvidia’s Vera Rubin NVL72 systems are designed around a 45 C supply temperature, with no water chillers required at that point. Ferveret is aiming APC at that same 45 C inlet.

Hyperscalers have been buying nuclear megawatts for the same reason. Microsoft’s deal to restart Three Mile Island Unit 1, now the Crane Clean Energy Center, is aimed at 2027. Google, Amazon, and Meta have their own reactor contracts. Those plants take years. A cooling box that hands 15% more TFLOPs/kW back to the rack is a 2026 lever on an interconnect that is already full.

Operators are already fighting over watts, transformers, and water rights. If that scramble were theater, the interconnection queues and the local permit fights would look a lot quieter.

3M Left Two-Phase Cooling Without Its Workhorse Fluid

Two-phase immersion had a fluid problem before Ferveret had a marketing problem. 3M said it exit of PFAS manufacturing at the end of 2025 was complete. Novec and Fluorinert chemistries, the low-boil dielectric fluids that made a lot of tank-style two-phase systems work, went with that exit. Distributors list the last Novec order window as closed.

WHAT DROPPED OUT OF THE TWO-PHASE AISLE

• The fluids: Novec and Fluorinert lines used in boiling immersion are no longer being made by 3M.

The form: Shared tanks hold many servers in one bath, which makes service a lift-out job and asks for floor load and custom plumbing.

• The pressure: Saturated boiling often wants tighter, costlier vessels than a near-ambient rack box.
• The opening: Ferveret says its fluid has low global-warming potential and is not regulated under PFAS rules.

Direct-to-chip with water-glycol cold plates filled much of that gap because it fits a standard rack and does not depend on a fluorinated bath. Ferveret is trying to keep two-phase heat transfer without the tank, the pressure vessel, or the old fluid. That is a supply-chain story as much as a thermodynamics one.

The Box Fits a Standard Rack

Most immersion products still look like a trough. Ferveret sells a modular chassis that holds one server and slides into the same racks used for direct-to-chip gear. The company’s technology page calls them rack-ready tanks with one server each, near ambient pressure, with less dielectric volume and no shared bath to drain when one machine needs a swap.

Azizian has said the physics is what lets the form factor shrink. Large tanks were never going to retrofit a colo row that already has 19-inch racks and a maintenance ticket system. A sealed box that a tech can service without shutting the neighbor is the product the existing floor can actually take.

Bucci describes a full stack, not a tub of fluid: the cooling box, the rack, coolant distribution units, and sensors for temperature and pressure. Software reads those sensors and trims power to each server so the box stays in the boiling regime that wastes the least energy. The 35% combined claim depends on that control layer as much as on bubble size.

On the company site, APC versus air cooling is a louder scorecard: 55% more TFLOPs/W, double the chip power, 96% less cooling cost, and a 75% smaller physical footprint. Those air-cooling comparisons are Ferveret’s, not UCLA’s. The independent test on the table is still the 15% TFLOPs/kW result against direct-to-chip.

Riyadh Is the Test of the Siting Pitch

Water-free cooling is easy to sell as a climate line. The sharper use is a building permit. Azizian wrote in June 2026 that a 10 to 20 MW hall can burn through up to 110 million gallons of water a year for cooling, about the use of 1,000 households, once cooling towers and evaporative loops are counted. Direct-to-chip does not erase that bill if the heat still dumps into a tower.

Bucci has pointed at the map. Solar is strong in places that do not have spare water. A loop that rejects heat with dry coolers at 45 C is a tool for Texas, Arizona, the Middle East, and parts of Africa, where a river permit can kill a site faster than a GPU allocation. National freshwater share can look small and the local fight can still be ugly.

That pitch had a stage in Riyadh. LEAP 2026 ran from August 31 to September 3. Ferveret took the Rocket Fuel Scaling Star award, a $150,000 prize for a late-stage startup, on September 3, from a field of 100 semi-finalists and 10 finalists. Azizian used the week to meet Tareq Amin, chief executive of HUMAIN, and to talk through halls that have to live with high ambient temperatures and scarce water.

It was an honor to meet @TareqAmin_ , CEO of @HUMAIN, at @LEAPandInnovate in Riyadh this week and to talk through data centers in Saudi Arabia and the cooling challenges that come with them.
What struck me most was how closely our thinking aligned on cooling, and on where… pic.twitter.com/aCDFXowbHW

— Reza Azizian (@Azizian_Reza) August 31, 2026

“Cooling is not a supporting detail in the AI buildout. It is one of the constraints that decides how much compute a region can actually deploy,” Azizian wrote after that meeting. He added that conventional cooling was not designed for Saudi heat, water scarcity, and power that has to stretch. The product claim in that climate is blunt: more compute from the same power, no chillers, and no water.

Aramco Ventures is already on the cap table. PitchBook records a $9.7 million seed round on June 4, 2025. Other named backers include TO VC, E14 Fund, Climate Capital, and Y Combinator. Charles Goodwin, a partner at TO VC, has framed the founders as people who spent years on nuclear thermal problems and are now applying that work where every megawatt is spoken for.

Pilots Remain Short of a Hyperscale Fleet

Ferveret says it is testing with CleanSpark, the data-center developer and operator; FuriosaAI, an accelerator firm; and Switch, a large U.S. operator. It is in Nvidia’s Inception program. In June 2026 the founders said they were in talks with hyperscalers and planned to name broader partnerships later in the year. Those names have not yet shown up as a fleet contract in public filings.

FROM REACTOR LAB TO A PRIZE STAGE

1. 2013: Azizian, then an MIT nuclear-engineering postdoc, meets Bucci, then a research scientist, over reactor heat transfer.
2. 2017: Azizian walks into his first data center and finds halls still cooled by massive fans; he later said air cooling can take 40% of the power going into a site.
3. 2021: They found Ferveret and join Y Combinator’s Summer 2021 batch.
4. June 4, 2025: PitchBook records a $9.7 million seed round.
5. March 26, 2026: UCLA’s ICON Lab result and Ferveret’s 1.03 PUE claim go out in a company statement.
6. September 3, 2026: Ferveret wins LEAP’s $150,000 Scaling Star award in Riyadh.

A seed-stage vendor with a rack box, a UCLA score on H200s, and a prize in Saudi Arabia is still a long way from a 100 MW campus. Direct-to-chip will not vanish because a startup posted 15% TFLOPs/kW. Tank vendors will not vanish because 3M left the fluid business. The test ahead is whether subcooled boiling in a serviceable chassis holds clock speeds on the next, hotter accelerators without a new chemistry scare.

Azizian’s line after the Data Center Knowledge interview still states the bet cleanly. When power is the constraint, cooling becomes capacity. The nuclear part of this story is not a reactor in the parking lot. It is a bubble that forms, sheds, and dies on a chip, and a box that tries to turn that cycle into tokens a grid hookup can actually support.