How nuclear-inspired liquid cooling could slash data center energy use

The explosive growth of generative AI is pushing data centers to the brink of unsustainable energy consumption. By decade’s end, U.S. data centers could account for up to 17% of the nation’s electricity use, with roughly a third of that power currently diverted to cooling processors that drive AI workloads. A startup emerging from MIT’s nuclear engineering labs is challenging the status quo with a liquid cooling system inspired by reactor physics, promising to cut energy waste while maintaining—or even boosting—AI performance.

A heat transfer breakthrough from reactor science

Ferveret was co-founded in 2021 by Reza Azizian, a former MIT postdoctoral researcher in nuclear engineering, and Matteo Bucci, MIT’s Esther and Harold E. Edgerton Associate Professor in Nuclear Science and Engineering. Their Adaptive Phase Cooling (APC) system reimagines chip cooling by adapting subcooled boiling—a process refined over decades in nuclear reactors—to the needs of modern data centers.

Azizian first encountered the inefficiencies of traditional data center cooling in 2017 while touring a facility packed with energy-hungry air-cooled servers. "I walked in and saw these massive fans spinning everywhere, consuming up to 40% of the data center’s power," he recalls. "It felt like watching a 50-year-old technology propping up an entire industry—inefficient, but tolerated because no one questioned it."

That observation sparked a collaboration with Bucci, whose research focused on optimizing heat transfer in reactor cores, where maximizing thermal efficiency directly impacts energy output. "Heat management in reactors determines how much power you can extract," Bucci explains. "Applying those same principles to chips could unlock dramatic efficiency gains."

Smaller bubbles, faster cooling, zero water waste

Unlike conventional air cooling, which relies on fans, or even basic liquid immersion systems, Ferveret’s APC uses a specialized dielectric fluid that absorbs heat far more effectively than air. The system’s key innovation lies in the size and behavior of the bubbles it generates at the chip surface. These micro-bubbles detach more frequently and recondense rapidly in the surrounding liquid, accelerating the heat transfer cycle.

Most immersion cooling methods boil the liquid to enhance heat removal—a process that adds complexity, requiring systems to capture, reliquify, and manage vapor while controlling pressure and temperature. Ferveret’s approach avoids boiling altogether, instead using a low-boiling-point fluid designed to produce smaller, more manageable bubbles without toxic PFAS compounds common in other systems.

Measurable gains in efficiency and compute output

In a recent study conducted with the Samueli Computer Science Department at UCLA, Ferveret’s APC demonstrated a 15% improvement in computational power efficiency compared to state-of-the-art liquid cooling solutions. When paired with the company’s real-time power optimization software, data centers achieved a 35% increase in AI model token generation—small units of text or data—using the same power input.

"Our mission is to squeeze every possible watt of value from data centers," Azizian says. "By eliminating water use and drastically reducing energy waste, we’re not just cutting costs—we’re enabling data centers to do more with less."

The system’s modular design further simplifies deployment. Instead of large, cumbersome tanks that submerge entire server racks, Ferveret’s APC fits into compact, rack-mounted units, one per server. This scalability allows operators to integrate the technology into existing infrastructure without major overhauls. "We’ve rethought the form factor," Azizian notes. "Our solution is designed to adapt, not disrupt."

Early adopters validate the tech

Ferveret has already begun testing its systems with industry leaders, including CleanSpark, a data center developer and operator; FuriosaAI, an AI accelerator startup; and Switch, one of the U.S.’s largest data center providers. These partnerships aim to validate the system’s performance under real-world conditions, where efficiency and reliability are critical.

The startup’s timing aligns with a growing urgency in the tech industry to address sustainability. As AI models grow more complex, the energy demands of data centers threaten to outpace renewable energy adoption. Solutions like Ferveret’s APC could provide a bridge, reducing the carbon footprint of AI while maintaining the performance that powers today’s most advanced applications.

The road ahead for sustainable AI infrastructure

With AI workloads continuing to expand, the pressure on data centers will only intensify. Ferveret’s approach offers a compelling alternative to the status quo—one that leverages deep technical expertise to deliver measurable efficiency gains without sacrificing performance. As the company scales its deployments and refines its control software, the broader industry will be watching closely to see if nuclear-inspired cooling can become the new standard for AI infrastructure.

The question now is whether data center operators—and the AI models that depend on them—are ready to embrace a future where cooling is no longer a power-hungry afterthought, but a finely tuned, energy-positive process.