Why are we cooling data centers like it's 1975?
As artificial intelligence reshapes the demands placed on global infrastructure, a quiet reckoning has arrived at the data center's door — not in the form of smarter algorithms, but in the humble physics of a boiling bubble. MIT startup Ferveret, drawing on decades of nuclear reactor research, has adapted heat transfer principles from one of humanity's most demanding engineering environments to address a problem the industry had long accepted as inevitable: that cooling powerful machines must consume vast energy and precious water. In doing so, they have opened a question that reaches beyond efficiency — about where, and under what conditions, the computational infrastructure of the future can be built.
- Data centers are on track to consume up to 17% of all U.S. electricity by decade's end, with cooling alone accounting for a third of that burden — a crisis hiding in plain sight behind server room doors.
- The AI boom is accelerating the problem, as ever-larger models demand ever-denser chips that generate heat at a scale that half-century-old fan-based cooling systems were never designed to handle.
- Ferveret's Adaptive Phase Cooling submerges servers in a non-toxic, low-boiling-point liquid and engineers smaller, faster-detaching bubbles that pull heat away from chips with 15% greater efficiency than leading liquid cooling rivals.
- Combined with real-time power optimization software, the system enables data centers to generate 35% more AI output from the same electricity — a gain already being tested with Switch, CleanSpark, and FuriosaAI.
- Because the system uses zero water, it could unlock data center construction in sun-rich but water-scarce regions — the American Southwest, the Middle East, parts of Africa — fundamentally redrawing the map of where AI infrastructure can live.
Reza Azizian walked into a data center in 2017 and noticed something the industry had long stopped questioning: enormous fans running ceaselessly, consuming roughly 40 percent of the facility's power, using technology that hadn't meaningfully changed in fifty years. Nobody was bothered because it worked — just wastefully.
Azizian had spent his career studying heat transfer in nuclear reactors, where moving heat efficiently is directly tied to how much energy a reactor can produce. He reached out to MIT researcher Matteo Bucci, and together they asked a deceptively simple question: why are we still cooling data centers like it's 1975? The answer became Ferveret, founded in 2021.
The company's approach centers on immersion cooling — submerging servers in specialized liquid rather than blasting air across them. The real innovation lies in the bubbles that form as the liquid boils. Phase change, liquid becoming vapor, transfers heat with extraordinary efficiency. Ferveret's system generates smaller bubbles that detach more rapidly from server surfaces, accelerating the heat transfer cycle. The liquid has a low boiling point and contains no toxic PFAS chemicals found in competing systems.
Testing with UCLA's computer science department showed a 15 percent improvement in computational efficiency over state-of-the-art liquid cooling. When paired with Ferveret's real-time power optimization software, the combined system enables data centers to extract 35 percent more AI output from the same electricity. The technology arrives in modular, single-server boxes — far easier to integrate into existing facilities than the large immersion tanks competitors require.
Ferveret is already running pilots with Switch, CleanSpark, and FuriosaAI, and is in conversation with the hyperscalers that underpin much of the internet. But perhaps the most consequential implication is geographic: the system uses zero water. In a world where data centers compete with agriculture and cities for scarce freshwater, a cooling technology that severs that dependency entirely could determine where the next generation of AI infrastructure gets built — opening sun-drenched, water-scarce regions that were previously off the table.
Reza Azizian walked into a data center in 2017 and was struck by something that most people in the industry had stopped noticing: the massive, deafening fans filling the building, working around the clock to keep the machines from overheating. He realized that cooling—the unglamorous infrastructure nobody talked about—was consuming roughly 40 percent of the power flowing into these facilities. The technology doing this work was half a century old. Nobody cared because it worked, even if it worked wastefully.
Azizian had spent years studying heat transfer in nuclear reactors, where efficiency matters enormously. In that world, how well you move heat directly determines how much energy you can extract from the reactor core, which translates to revenue. He began thinking about what would happen if you applied that same rigor to data centers. He reached out to Matteo Bucci, an MIT researcher he had worked with years earlier, and together they started asking a simple question: why are we cooling data centers like it's 1975?
The answer came in the form of Ferveret, founded in 2021. The company adapted a cooling technique from nuclear reactors and applied it to the servers that power artificial intelligence. The timing was crucial. Data centers are projected to consume between 9 and 17 percent of all electricity used in the United States by the end of this decade. Right now, roughly a third of that power goes to cooling—to moving heat away from chips that are generating more and more of it as AI models grow larger and more demanding.
Ferveret's approach uses immersion cooling, submerging servers in a specialized liquid rather than blowing air across them. But the company's real innovation lies in the bubbles. When the liquid heats up, it begins to boil, and those bubbles are where the magic happens. Boiling liquid transfers heat far more efficiently than air because the phase change—liquid turning to vapor—requires enormous amounts of energy, energy that gets pulled directly from the chip. Ferveret's system produces much smaller bubbles than competing immersion cooling approaches, and they detach more frequently from the server surface, accelerating the heat transfer cycle. The liquid has a low boiling point and contains none of the toxic PFAS chemicals that other systems rely on.
In testing with UCLA's computer science department, Ferveret's Adaptive Phase Cooling system delivered a 15 percent improvement in computational power efficiency compared to state-of-the-art liquid cooling solutions. But the company claims something more ambitious: when you combine that efficiency gain with their power control software—which optimizes operating conditions in real time—data centers can extract 35 percent more tokens (the small pieces of text or data that AI models generate) from the same amount of electricity. The system comes in modular boxes, each housing a single server, making it far easier to integrate into existing data center infrastructure than the large tanks that other immersion cooling approaches require.
Ferveret is already testing its technology with major players: CleanSpark, a data center developer and operator; FuriosaAI, an AI accelerator company; and Switch, one of the largest data center operators in the United States. The company is also part of Nvidia's Inception program for startups and is in talks with the hyperscalers—the massive cloud computing companies that power much of the internet.
What makes this technology particularly significant is what it enables beyond just efficiency. Data centers require enormous amounts of water for cooling, and that's a constraint in many parts of the world. Ferveret's system uses zero water. That opens the possibility of building data centers in regions where water is scarce but renewable energy is abundant—places like parts of Africa, the Middle East, and the American Southwest where the sun shines reliably but water is precious. As the AI industry continues to expand and demand for computing power grows, access to both electricity and water will become increasingly limiting factors. A technology that severs the link between cooling and water consumption could fundamentally change where and how data centers get built.
Notable Quotes
Our goal is to make data centers as sustainable as possible and help them use every single watt of power to generate tokens, which are the most useful outputs.— Reza Azizian, Ferveret co-founder
The sun shines in places where you don't have much water, so the advantage of us being water-free is we allow you to build data centers where you have solar energy but nothing to cool the data center down.— Matteo Bucci, Ferveret co-founder
The Hearth Conversation Another angle on the story
Why does cooling matter so much? I thought the real energy cost was in the computation itself.
Cooling is the invisible tax on computation. You're right that the chips themselves consume power, but moving that heat away consumes almost as much. In a typical data center today, a third of all electricity goes to cooling. That's not a small thing.
And the nuclear reactor connection—that seems like an odd place to look for inspiration.
Not really. Nuclear engineers have spent decades optimizing heat transfer because it directly affects how much energy you can extract from the reactor. The physics is the same whether you're cooling a reactor core or a GPU. Azizian and Bucci just asked: why haven't data center people borrowed these techniques?
The bubbles seem like a small detail. Why do smaller, faster-detaching bubbles matter so much?
Because bubbles are where the real heat transfer happens. When liquid boils, the phase change pulls enormous amounts of energy from the chip. Smaller bubbles that detach more frequently mean more of that phase-change energy gets used, and it gets used faster. It's the difference between a slow simmer and a rolling boil.
What about the water-free angle? That seems almost secondary to the efficiency gains.
It's not secondary—it's transformative. You can build a data center anywhere the sun shines now. That changes the entire geography of where computing infrastructure can exist. In a world where water is becoming scarce and renewable energy is becoming cheap, that's not a detail. That's the future.
So this is about more than just making AI cheaper to run?
It's about making AI possible to run in places where it wasn't possible before. The industry is hitting constraints—power constraints, water constraints. This technology addresses both at once.