Nuclear-Derived Cooling Claims 35% More AI Output Per Watt: the real signal is the immediate adjustment required in cash, risk, and execution
The Number That Leads
Cooling consumes roughly one-third of total data center electricity today. That allocation is the quiet drag on every capacity decision your team makes: power purchased but never converted to compute. Against that baseline, MIT-founded startup Ferveret is reporting a 15 percent improvement in computational power efficiency compared to state-of-the-art liquid cooling, based on a study conducted with the Samueli Computer Science Department at UCLA. The company further claims that when its Adaptive Phase Cooling system is combined with its power control software, data centers can extract 35 percent more AI tokens from the same power envelope.
Both figures originate from Ferveret and its academic collaborators. The study has not been independently replicated in production environments at scale, and the company is early-stage. Read the numbers as directional evidence of a credible mechanism — not benchmarks ready to enter a procurement model.
What Sits Behind the Number
The mechanism draws on subcooled boiling, a heat transfer technique refined for nuclear reactor cooling. In a reactor, the efficiency of heat removal determines how much energy can be extracted from the core — a constraint that drove decades of materials and fluid dynamics research. Ferveret’s founders, one a former MIT nuclear engineering postdoc and one a current MIT faculty member in nuclear science, are applying that lineage to chip-level thermal management.
The system submerges servers in a proprietary liquid with a low boiling point, free of PFAS compounds. At the chip surface, the liquid produces smaller bubbles than conventional immersion cooling approaches. Those bubbles detach faster and recondense in the surrounding liquid, tightening the rewetting cycle and accelerating heat transfer without requiring large open-tank infrastructure. Each server sits in a modular rack-mounted enclosure rather than the large immersion tanks that dominate current deployments.
A software layer adjusts power delivery to each server in real time, monitoring temperature and pressure sensors to minimize energy consumption at the box level. The compound efficiency claim combines the physical cooling gain with that software optimization — the two effects have not been separately audited in available public data.
What This Is Worth in Your Operation
If the efficiency gain holds at production scale, the arithmetic across a large portfolio is material. A facility running 100 MW of IT load currently devotes roughly 30 to 35 MW to cooling under air-dominated or hybrid configurations. As chip power requirements climb with next-generation AI accelerators, the gap between efficient and inefficient cooling widens in absolute megawatt terms — meaning a genuine cooling efficiency improvement compounds in value rather than staying flat.
The zero-water claim carries a separate strategic dimension independent of the efficiency numbers. Water availability is an increasingly explicit constraint in site selection, particularly in the U.S. Southwest, the Middle East, and parts of sub-Saharan Africa — regions where solar irradiance is high but water access is restricted or politically contested. A cooling architecture that removes water dependency entirely reconfigures the geographic optionality available when evaluating greenfield sites near large solar generation assets. For any team negotiating PPAs in locations where water permitting moves as slowly as grid interconnection, that intersection is worth mapping against your current pipeline.
The modular, per-server form factor also lowers the barrier for a controlled pilot. Large immersion tanks require significant civil and facility modifications. Rack-integrated enclosures make it feasible to test the technology in a live environment without committing to a full facility redesign — relevant when assessing vendor risk against a pre-commercial company.
What the Data Does Not Say
The UCLA study result has not been verified in a hyperscale or large colo production environment. Ferveret’s current deployment partners — CleanSpark, FuriosaAI, and Switch — are in testing, but no published operational data from those deployments has been made available as of June 2026. The proprietary liquid’s long-term stability, compatibility across chip generations, and total cost of ownership relative to direct liquid cooling or single-phase immersion are not addressed in available public data.
Ferveret is part of Nvidia’s Inception startup program and in discussions with hyperscalers, but no commercial agreements with hyperscale operators have been announced. For operators managing multi-GW portfolios, the risk profile of a pre-commercial vendor is as operationally significant as the efficiency claim itself. The wider projection that data centers will account for 9 to 17 percent of U.S. electricity by end of decade reflects genuine uncertainty in load forecasting — and the upper bound is the scenario in which cooling efficiency improvements carry the most systemic weight.
The Implementation Question
Before the claimed efficiency improvement can inform a capital or procurement decision, route one question to your infrastructure team: under your highest-density planned workloads — next-generation GPU clusters or inference farms — what would a meaningful reduction in cooling power draw unlock in additional IT capacity within your existing power envelope, and at what deployment scale does a structured pilot with Ferveret become the most cost-effective way to answer that question with real operational data?
Sources
- Mit — Startup’s nuclear-inspired cooling system could make data centers more sustainable (Link)
