Startup's nuclear-inspired cooling system could make data centers more sustainable

Overview

Ferveret, a startup founded by two MIT nuclear engineering researchers, has built a liquid cooling system for AI data center chips that uses zero water and reports a 15% gain in computational power efficiency over current liquid cooling. The company borrows a heat-transfer method called subcooled boiling from inside nuclear reactors, submerging servers in a special liquid to pull heat off chips far faster than air. Paired with its power-control software, the setup lets AI models generate 35% more tokens from the same amount of power. Early testing partners include CleanSpark, FuriosaAI, and Switch.

Key Takeaways

Ferveret adapts subcooled boiling, the heat-removal physics used inside nuclear reactors, to cool AI chips by submerging servers in a low-boiling-point liquid rather than blowing air across them.
The system uses zero water, which opens the door to building data centers in dry regions that have abundant solar power but little water, including parts of Africa, the Middle East, and the Americas.
The company reports a 15% efficiency gain over state-of-the-art liquid cooling, verified in testing with UCLA's Samueli Computer Science Department.
Combined with Ferveret's power-control software, data centers produce 35% more tokens from the same electricity, meaning more useful AI output per watt.
The product ships as modular, rack-mounted boxes that fit existing data center racks, avoiding the large immersion tanks other liquid-cooling approaches require.
Founders Reza Azizian and Matteo Bucci met at MIT in 2013 while studying heat transfer in nuclear reactors and launched the company in 2021.

Stats & Key Facts

#15% improvement in computational power efficiency over state-of-the-art liquid cooling
#35% more tokens generated from the same amount of power when paired with power-control software
#Zero water consumption across the cooling process
#Roughly one-third of a data center's electricity goes toward cooling its chips
#Air cooling alone consumes up to 40% of a data center's total power
#Data centers are projected to use 9 to 17% of all U.S. electricity by the end of the decade
#Founded in 2021 by two MIT researchers; raised about $2.1 million in a 2021 seed round per third-party startup databases

Subcooled Boiling From Nuclear Reactors Applied to AI Chips

Ferveret's core idea takes a cooling method proven inside reactors and points it at computer chips.

The technology, called Adaptive Phase Cooling, adapts subcooled boiling, the same heat-transfer process that pulls heat away from fuel inside nuclear reactors. Servers sit submerged in a specialized liquid with a low boiling point that contains no toxic PFAS chemicals. As the chip heats up, tiny bubbles form at its surface, then detach and recondense quickly.

That fast bubble cycle is what moves heat away rapidly. As co-founder Matteo Bucci explains, liquid is a far better heat-transfer medium than air, which is why room-temperature water feels cold on your hand. When the liquid boils, the phase change absorbs a large amount of energy, and that energy is the heat being removed from the chip.

Zero Water Cooling Unlocks Dry, Solar-Rich Regions

Removing water from the equation changes where data centers can be built.

›Traditional cooling often relies heavily on water, which limits siting in water-scarce areas.
›Ferveret's water-free design suits regions with strong solar power but limited water supply.
›Target areas include parts of Africa, the Middle East, and the Americas where sunshine exceeds water availability.
›The approach ties data center growth to abundant renewable energy rather than scarce freshwater.

Why Data Center Cooling Is Such an Energy Problem

Cooling is one of the largest hidden costs of running AI infrastructure.

›About one-third of the electricity a data center draws is spent cooling its chips.
›Air cooling alone can account for up to 40% of a data center's total power use.
›Data centers are projected to consume 9 to 17% of all U.S. electricity by the end of the decade, driven by AI demand.
›Cutting cooling energy frees more power for the actual computing work.

15% Efficiency and 35% More Tokens Per Watt

The headline numbers describe both raw cooling gains and real AI output.

Ferveret reports a 15% improvement in computational power efficiency compared with current state-of-the-art liquid cooling. That figure was measured in work with UCLA's Samueli Computer Science Department, an outside academic partner rather than the company alone.

When the cooling hardware runs alongside Ferveret's real-time power-optimization software, the company says AI models produce 35% more tokens from the same amount of power. Tokens are the units of text an AI model generates, so more tokens per watt means more useful output for the same electricity bill. Co-founder Reza Azizian frames the goal as helping data centers turn every watt into useful output.

Modular Rack-Mounted Hardware Fits Existing Data Centers

The product is designed to drop into infrastructure operators already run.

›Ships as modular, rack-mounted boxes rather than large immersion tanks.
›Fits into existing data center racks, which simplifies deployment and maintenance.
›Delivered as a full-stack system: the cooling box, the rack, cooling distribution units, and temperature and pressure sensors.
›Bucci says the physics let the team reach compact form factors that were not possible before.

MIT Founders, 2021 Launch, and Early Partners

The company grew out of a decade of shared research at MIT.

Reza Azizian is a former MIT postdoc in nuclear engineering who earlier worked on Microsoft HoloLens and at Nvidia. Matteo Bucci is an associate professor in MIT's Department of Nuclear Science and Engineering. The two met at MIT in 2013 while studying heat transfer in nuclear reactors and founded Ferveret in 2021.

Early testing partners include CleanSpark, FuriosaAI, and Switch, and the company collaborates with UCLA's Samueli Computer Science Department. According to third-party startup databases, Ferveret raised about $2.1 million in a 2021 seed round from investors including E14 Fund and Climate Capital; that funding figure comes from those databases rather than the MIT source article.

Frequently Asked Questions

What is Ferveret's cooling technology?

It is a liquid cooling system called Adaptive Phase Cooling that submerges data center servers in a special low-boiling-point liquid. It adapts subcooled boiling, the heat-removal physics used inside nuclear reactors, to pull heat off AI chips faster than air cooling.

How much more efficient is it than current cooling?

Ferveret reports a 15% improvement in computational power efficiency over state-of-the-art liquid cooling. Paired with its power-control software, it says AI models generate 35% more tokens from the same amount of power.

Why does using zero water matter?

Many cooling methods rely on large amounts of water, which restricts where data centers can be built. A water-free system lets operators build in dry, solar-rich regions such as parts of Africa, the Middle East, and the Americas where sunshine is plentiful but water is scarce.

Who founded Ferveret?

It was founded in 2021 by Reza Azizian, a former MIT nuclear engineering postdoc who previously worked on Microsoft HoloLens and at Nvidia, and Matteo Bucci, an associate professor in MIT's Department of Nuclear Science and Engineering. They met at MIT in 2013.

Which companies are testing the system?

Early testing partners include CleanSpark, FuriosaAI, and Switch. The company also collaborates with UCLA's Samueli Computer Science Department, which verified the 15% efficiency figure.

Ferveret bets that borrowing reactor-grade heat-transfer physics will let AI data centers run hotter chips with far less wasted energy and no water at all. If the efficiency and token gains hold at scale, the approach could reshape where and how the next wave of AI infrastructure gets built.