An MIT spin-out borrowed a nuclear reactor trick to cool AI chips, and won $150,000 for it

Ferveret won the Scaling Star Award at Rocket Fuel 2026, the $150,000 prize given to the competition's strongest growth company. It builds immersion cooling for the chips that run AI workloads. The company was founded in 2021 by Matteo Bucci, an associate professor at MIT, and Reza Azizian, a former MIT postdoc. Its approach is adapted from subcooled boiling, a heat transfer process used in nuclear reactors, and applied to a problem that has become one of the binding constraints on AI infrastructure. The physics is worth explaining because it is where the claim sits. Immersion cooling works by submerging chips directly in a liquid rather than blowing air across them. When the liquid boils at the chip surface, bubbles form, carry heat away and detach. The rate at which that cycle repeats sets how fast heat leaves the chip. Ferveret's liquid produces smaller bubbles that detach more often and recondense quickly in the surrounding fluid, which speeds up the rewetting cycle at the surface and moves heat faster. The second claim is chemical rather than thermal. Many immersion cooling systems rely on fluorinated liquids from the PFAS family, the compounds known as forever chemicals, which do not break down in the environment and are under tightening regulatory pressure in several jurisdictions. Ferveret's liquid has a low boiling point and, according to the company, none of them. For an operator building capacity meant to run for twenty years, a coolant with a regulatory question attached to it is a liability on the balance sheet as much as an environmental one. The packaging is the third piece, and commercially it may be the most important. The system ships as small boxes each housing a single server rather than as a tank holding many. That makes it something an operator can deploy incrementally and service one unit at a time, instead of a facility-level commitment that has to be designed in from the start. Retrofit is where most cooling technology fails to sell, because the buyer already owns a building. Why this wins a growth award at a Saudi event is not hard to see. Cooling is the constraint the Kingdom's data-centre programme runs into hardest. TechScoop has reported this year on district-scale developments, on capacity commitments starting at 250 megawatts, on a company committing $1.2bn to reach 192 megawatts, and on immersion cooling agreements signed at this same event. Every one of those megawatts becomes heat, in a climate where rejecting heat to the outside air is expensive for much of the year and where water for evaporative cooling is a strategic resource rather than a utility line item. That combination — high ambient temperature, constrained water, and a buildout measured in gigawatts — makes the Gulf one of the strongest markets in the world for a cooling technology that reduces both energy and water use. A company with the right product here does not need to explain the problem to the buyer. What the award does not establish is commercial traction, and the Scaling Star category name should not be read as evidence of it. No customer count, deployment size or revenue figure has been published, and a laboratory result and a shipping product are different things in this category — data-centre operators are conservative buyers because a cooling failure takes the compute with it. The number that would settle it is a named operator running Ferveret hardware in production at scale, and it has not been announced.