The Next AI Breakthrough Might Be a Heat Exchanger

The most consequential constraint on artificial intelligence in 2026 isn't the cleverness of the next model. It's whether the building it runs in can shed heat fast enough and find enough power to keep going. As chips like Nvidia's Blackwell and Rubin pack more compute into smaller spaces, the watts per rack have climbed past what air can carry away — and every watt spent on cooling is a watt not spent on inference. Into that bottleneck steps an MIT spinout with an unfashionable thesis: the breakthrough that unlocks the next wave of AI might not be a transformer at all. It might be a better way to boil a liquid.

Reactor physics, repurposed

Ferveret, founded in 2021 out of MIT and described in a June 10 profile from MIT News, is built on a process its founders know intimately from a very different industry: subcooled boiling, the heat-transfer regime that keeps nuclear reactor cores from melting. The company's two principals come straight from that world. Matteo Bucci is MIT's Edgerton Associate Professor in the Department of Nuclear Science and Engineering; his co-founder Reza Azizian is a former MIT nuclear-engineering postdoc whose résumé also includes stints on Microsoft's HoloLens and at Nvidia. Azizian recalls his first walk through a data center bluntly: "I thought, 'Holy crap, this is not how you cool facilities.'"

Their system, Adaptive Phase Cooling (APC), submerges servers directly in a specialized liquid with a low boiling point. As the chips heat the fluid, tiny bubbles form at the silicon surface, detach almost immediately, rise, and recondense — and that rapid, controlled churn pulls heat away far faster than air blown by fans or even conventional liquid loops. The reactor analogy is exact: subcooled boiling is prized precisely because the phase change from liquid to vapor absorbs enormous amounts of energy in a small space. Ferveret's contribution is engineering bubbles small enough and frequent enough to make that work on a server board instead of a fuel rod.

No water, no forever chemicals

Two design choices make APC more than an incremental efficiency play. First, it uses no water — a quietly radical feature when the AI build-out has turned data-center water consumption into a live political fight in drought-prone regions from Arizona to Spain. A waterless system that performs best in hot, sunny, water-scarce places aligns almost perfectly with where cheap solar power wants to put the next generation of compute.

Second, the working fluid avoids the toxic PFAS "forever chemicals" that haunt many immersion-cooling approaches. PFAS-based coolants work, but they carry mounting regulatory and liability risk as governments move to restrict the chemicals. By sidestepping them, Ferveret removes a future headache that could otherwise strand a facility's entire cooling architecture.

The metric that matters: tokens per watt

The performance claims are where the physics meets the business case. Ferveret says APC delivers a 15% improvement in computational power efficiency versus state-of-the-art liquid cooling. Pair that with the company's power-control system, which tunes operating conditions to keep chips in their efficient zone, and the combined figure is the one that should make every hyperscaler look twice: 35% more AI tokens generated from the same amount of power.

Read that number through the lens of the real constraint and it reframes the entire problem. The thing in shortest supply for AI right now is not silicon and not capital — both can be acquired on a timeline — it's interconnection to the grid. New large-load power hookups are quoted in years, not months. If a cooling change lets an operator extract a third more useful work from a fixed megawatt allocation, that is functionally equivalent to building a third more data center, without the substation, the transmission upgrade, or the multi-year wait. In a market where compute is rationed by electrons, efficiency is capacity.

From lab to live racks

This is not pure theory. Ferveret is already running pilots with CleanSpark, a data-center developer and operator; FuriosaAI, an AI-accelerator maker that needs to keep its own chips cool; and Switch, one of the largest data-center operators in the United States. Early research was supported through the MIT Energy Initiative. That spread of partners — a developer, a chipmaker, and a colocation giant — matters because each validates a different slice of the pitch: that APC works in new builds, that it pairs with non-Nvidia silicon, and that it can retrofit into the kind of scaled facilities that actually move the industry's power math.

The caveats are the usual ones for hardware at this stage. Pilot results aren't production data; immersion cooling demands new server designs and maintenance practices that conservative operators adopt slowly; and the 35% figure bundles cooling gains with software-side power control, so the cooling-alone contribution is the smaller 15%. None of that is fatal, but it does mean the proof will come from how a Switch or a CleanSpark facility performs at scale, not from a spec sheet.

Still, the strategic point is hard to miss, and it's one a growing chorus inside the industry has started to voice: the frontier is quietly shifting from how big is the model to how cheaply can you run it. Tensordyne is attacking that with arithmetic. Ferveret is attacking it with thermodynamics. Both are betting that the next decisive advantage in AI won't be announced at a model launch — it'll be hiding in the unglamorous machinery that turns a megawatt into an answer. Sometimes the most important thing in the building is the heat exchanger.