Ferveret, an MIT spin-off, says it has developed a water-free cooling system for AI data centres. The company draws on heat transfer principles from nuclear reactor engineering. Its founders believe the approach can cut both electricity and water consumption at scale.
A Cooling Problem Getting Harder to Ignore
Data centres already consume vast amounts of energy. Industry estimates suggest cooling accounts for roughly a third of total electricity use in these facilities. AI is accelerating the build-out of new capacity. That expansion is sharpening pressure on operators to find more efficient solutions.
Reza Azizian and Matteo Bucci founded Ferveret in 2021. Both researchers came from MIT. Azizian has said his concerns about conventional cooling began when he first entered a data centre in 2017. He saw a sector that had little incentive to change methods that kept machines running, even at significant waste.
Air cooling has dominated data centre design for decades. Fan-driven systems move heat away from servers using large volumes of air. The process works, but it is energy-intensive. Many operators have already shifted towards liquid immersion systems. Ferveret says its approach goes further.
What the Technology Actually Does
The company’s Adaptive Phase Cooling system submerges servers in a dielectric fluid. The system then uses a boiling process called subcooled boiling. This technique originates in nuclear reactor heat transfer research.
In practical terms, the system generates smaller bubbles at the chip surface. Those bubbles detach more frequently and recondense faster. Heat moves away from the chip more quickly as a result. Bucci has said that boiling absorbs large amounts of energy through the phase change itself. That absorption is the source of the efficiency gain.
Ferveret says the system avoids water use entirely. The fluid contains no PFAS chemicals. The system also operates near ambient pressure. That reduces complexity compared with some other two-phase cooling designs.
The company tested the technology with the Samueli Computer Science Department at the University of California, Los Angeles. Ferveret says results showed a 15 per cent gain in computational power efficiency compared with leading liquid-cooling systems. The company says that, combined with its power-control software, the improvement can translate into 35 per cent more AI tokens from the same amount of power. These figures come from Ferveret’s own testing and have not been independently verified.
Modular Design and Geographic Reach
Many immersion cooling systems use large tanks. Ferveret has chosen a different format. Its units are rack-mounted boxes, each housing a single server. The company says this makes installation and maintenance easier in existing data centre facilities.
The system fits standard liquid-cooled rack sizes. Ferveret says it can support very high cooling loads and operate reliably in hot and humid conditions. That environmental tolerance matters for where the technology could be deployed.
Water scarcity has become a genuine constraint on data centre development in many regions. Energy availability alone is insufficient if there is no reliable cooling water supply. Ferveret says a water-free system opens viable sites in areas with strong solar resources, including parts of Africa, the Middle East, and the United States. The claim is plausible given the technology’s design, though no independent assessments of those deployments have been published.
Early Partners and Industry Context
Ferveret says it is already testing the technology with CleanSpark, FuriosaAI, and Switch. The company is also in discussions with major cloud providers. It participates in Nvidia’s Inception programme for startups.
The commercial pathway is still early-stage. No contracts or deployment figures have been disclosed publicly. The discussions with cloud providers remain unconfirmed by those parties.
The broader context matters here. Data centres could account for a substantial share of US electricity demand by the end of this decade, according to industry estimates. That trajectory is drawing serious attention from investors and policymakers alike. Cooling technology is one of the few areas where significant efficiency gains remain available without redesigning the chips themselves.
Ferveret’s argument is straightforward in structural terms. More computing output from each watt requires removing heat more effectively. The company believes nuclear engineering already solved that problem. The question is whether the data centre industry will move fast enough to find out.
