What if cooling a data center came down to the same physics used in nuclear reactors? That is exactly the premise behind Ferveret, a startup that is bringing sub-cooled boiling technology into the AI data center market. The performance gains for data centers potentially huge: 35% more compute from the same power envelope. We sat down with Reza Azizian, co-founder and CEO of Ferveret to talk about the company’s technology, its origins, and why the data center industry is now ready for a cooling approach that was once considered too complex to adopt at scale.
The story of Ferveret starts at MIT, where Azizian was researching ways to optimize heat transfer in nuclear reactors. The principle is straightforward: better heat transfer means more energy can be extracted from the nuclear core, which is directly tied to the revenue of those facilities. That same logic, he realized, applies perfectly to data centers.
“Cooling has become a big bottleneck, and improving cooling means you can get more compute out of your data center,” Azizian explains. “That is directly tied to the revenue of data centers.”
Before founding Ferveret, Azizian worked on the Microsoft HoloLens and then joined NVIDIA, where he designed cooling for various GPUs and worked directly with TSMC to improve heat transfer at the package level. It was during that time that he first walked into a large-scale air-cooled data center. He was struck by how inefficient it was. “The noise hits you first. It is like a plane taking off,” he recalls. “Then you realize: air is compressible. You have to spend a lot of energy to squeeze it between all the components. And it needs a huge surface area, which is why you need massive heat sinks on top of every chip.”
Why air cooling is running out of road
As chip power keeps increasing in line with Moore’s Law, air cooling becomes increasingly untenable. The fans that dominate today’s data centers consume enormous amounts of energy, require cold aisle and hot aisle infrastructure, and demand air handlers throughout the facility. Even in direct-to-chip liquid-cooled deployments, fans remain necessary because direct-to-chip only cools the processors. Everything else still relies on airflow. As long as anything in the rack is air-cooled, the entire facility-level overhead remains, Azizian says. The real energy waste lives in that overhead, he says.
The case for immersion
From a pure physics standpoint, immersion cooling is far more efficient than any air-based approach. That has been clear for a long time already. The form factor and the operational complexity held it back.
Traditional immersion cooling relies on large tank-style systems. Multiple servers are submerged in a single bath of dielectric fluid. If one server needs maintenance, all servers in that tank must come offline. The infrastructure required is entirely different from a standard data center deployment, demanding significant capital investment and carrying real operational risk. These barriers, Azizian argues, are the primary reason immersion never achieved widespread adoption. “We truly believe that is the main reason that thing never took off,” he says. “So we asked: how can we solve this problem?”
Adaptive Phase Cooling: a new category
Ferveret’s way to solve the problem alluded to above is what it calls Adaptive Phase Cooling. This is an approach that sits between single-phase and full two-phase immersion. The company uses sub-cooled boiling. A heat exchanger design developed by Ferveret keeps the top layer of fluid cooled, so bubbles that form at the chip surface collapse within the liquid rather than rising to the surface and building pressure.
The practical consequences are significant, Azizian says. The system operates at near-atmospheric pressure, which eliminates the venting and fluid loss problems that plague conventional two-phase systems. The rewetting cycle, where cold liquid quenches the chip surface after each bubble departs, is faster and more consistent, improving thermal performance. Crucially, the fluid can have a low boiling point without the system becoming unstable.
The core intellectual property sits in what Ferveret calls the “cooling brain”. This is the heat exchanger design that makes sub-cooled boiling viable at the chassis level. It is where the company holds most of its patents and trade secrets.
Ferveret fits the existing infrastructure
Perhaps the most commercially significant aspect of the Ferveret design is its form factor. Each server slides into a Ferveret chassis that sits on exactly the same rack infrastructure used for direct-to-chip cooling. Nothing needs to change on the data center floor, Azizian claims. No new CDUs of a fundamentally different type, no tank cranes, no new power distribution architecture.
Maintenance follows the same workflow as any standard server deployment: pull the chassis from the rack, take it to the maintenance room, drain the fluid, perform the service, refill, and reinstall. The operator never comes into direct contact with the liquid during normal operation.
Current production units are 4U, optimized for NVIDIA DGX systems, though Ferveret is actively developing 1U form factors and can adjust chassis height based on server density and customer requirements.
Significant efficiency gains
Ferveret has conducted benchmarking in collaboration with UCLA on servers including the NVIDIA H200 and Blackwell. The results show approximately 15% improvement in teraflops per watt at the server level when comparing Ferveret to direct-to-chip cooling. That gain comes from two sources: eliminating all fan energy at the server level, and running the chip at a lower steady-state temperature.
The facility-level gains of the benchmarks are even more significant. Ferveret systems support a 45°C liquid inlet temperature. That is a threshold that NVIDIA’s Vera Rubin platform is also targeting. At that inlet temperature, chillers can often be eliminated entirely in favor of free cooling via dry coolers, cutting the largest single component of data center overhead.
The PUE achievable with full-rack Ferveret deployment drops to 1.03–1.04, compared to approximately 1.15–1.20 for typical liquid-cooled facilities. When the full-rack approach is applied, covering compute trays, networking trays, and power shelves, and the math is run end-to-end from facility power input to compute output, Azizian claims up to 35% more compute from the same power envelope.
Conclusion
The physics of immersion cooling has always been superior. Ferveret claims to have solved the problem around adoption. It offers an intermediate immersion cooling solution that has the right form factor and doesn’t bring the operational complexity that other solution had and still have. In addition, it offers chassis-level design, near-atmospheric operation, PFAS-free fluid, and full-rack coverage.
The company claims very significant gains in efficiency, with a near-perfect possible PUE for data centers that deploy its solution. In a market where getting as much performance out of every watt consumed, this is potentially ground-breaking. Not only for data centers being built or outfitted at the moment, but also moving forward. According to Azizian, Ferveret is ready for any of the upcoming challenges too, whether it’s the rise of (silicon) photonics or much higher voltage DC power delivery systems.
Also read: From water to watts: TES Group brings unique expertise to data centers