MIT Startup Ferveret Beats Liquid Cooling to Yield 35 Pct

Two MIT researchers looked at how nuclear reactors cool themselves and asked whether the same physics could cool AI chips. The answer, published June 10, 2026, is a company called Ferveret that has demonstrated 35% more tokens from the same power draw compared to conventional liquid cooling. In a sector where efficiency gains typically arrive in single-digit percentages, that number demands a closer look at what they actually built and whether it will scale.

What Actually Happened

MIT News published a detailed profile of Ferveret on June 10, 2026, describing a cooling system the company calls Adaptive Phase Cooling (APC). The technology was developed by Reza Azizian, a postdoctoral researcher in MIT's nuclear science and engineering department, and Matteo Bucci, a professor at MIT who specializes in boiling heat transfer. According to MIT News, Ferveret's system improves computational power efficiency by 15% over state-of-the-art liquid cooling, and when combined with the company's power optimization software layer, achieves the 35% improvement in tokens generated per watt of power consumed.

The system operates without water. Current data center cooling configurations rely on chilled water loops or direct liquid cooling systems that consume millions of gallons of water annually at large facilities. Ferveret's fluid is both waterless and PFAS-free, addressing two distinct regulatory and environmental concerns simultaneously. The company uses a proprietary phase-change fluid with thermal properties matched to GPU heat load characteristics, allowing heat to be removed at the chip level through evaporation rather than sensible heat transfer, the same physical mechanism used in nuclear reactor thermal management systems. The company is an Nvidia Inception program member and has completed early deployments with three named customers: CleanSpark (a publicly listed Bitcoin miner with GPU compute operations), FuriosaAI (a South Korean AI accelerator startup), and Switch (a US-based hyperscale data center operator). A research collaboration with UCLA Samueli School of Computer Science is ongoing. The company has raised $125,000 in pre-seed funding from Y Combinator and Climate Capital.

Detailed technical data supporting the 35% and 15% claims was published as part of the MIT News coverage and is supported by a research collaboration described in additional reporting from TechXplore. The company claims its system can be retrofitted onto existing GPU rack configurations without requiring changes to the underlying server hardware. An independent summary of the efficiency claims appears in Datacenter News, which noted that the 35% gain applies specifically to token throughput per watt when the APC hardware and Ferveret's firmware-level power governor are used together.

Why This Matters More Than People Think

AI inference is rapidly becoming the dominant workload in commercial data centers. Unlike training, which runs in large bursts and can be time-shifted, inference runs continuously in response to live user requests. Token generation throughput per watt is therefore not an academic metric: it is the unit economics of every AI product that charges per token. A 35% improvement in tokens per watt means a proportional reduction in cost per token at fixed hardware cost, or equivalently, 35% more revenue per watt of power capacity. For hyperscale operators who measure power capacity in hundreds of megawatts, that efficiency figure translates directly to hundreds of millions of dollars in additional annual revenue potential from the same licensed power capacity.

The water consumption angle is less obvious but increasingly load-bearing in AI infrastructure planning. Several US cities with major data center concentrations including Phoenix, Las Vegas, and Northern Virginia are facing water scarcity constraints that are now showing up in zoning and permitting decisions. Virginia's Prince William County imposed a moratorium on data center approvals in 2023 partly in response to groundwater concerns. Microsoft's data center water consumption disclosures in 2023 showed that a single facility could consume over 10 million gallons of water annually, drawing broad regulatory scrutiny. A cooling system that is genuinely waterless, not just "water efficient," removes a constraint that is tightening at precisely the same time that AI data center demand is accelerating. For operators in water-stressed regions, Ferveret's waterless cooling may open sites that would otherwise be unavailable.

The phase-change mechanism also has a secondary advantage that hasn't received attention in the initial coverage. Phase-change cooling maintains much more uniform chip temperatures across the GPU die than sensible-heat liquid cooling, because the phase transition occurs at a fixed temperature rather than across a temperature gradient. Thermal uniformity matters for AI performance for the same reason it matters in electronics manufacturing: non-uniform temperatures create thermal stress and cause processor throttling in the hottest die areas while cooler areas are under-utilized. Improved thermal uniformity means more consistent compute utilization across the GPU die area, which contributes to the efficiency gains independently of the raw heat removal improvement.

The Competitive Landscape

The liquid cooling market for AI data centers is large and crowded. Vertiv, CoolIT Systems, Asetek, LiquidStack, and ZutaCore are all active in direct liquid cooling for GPU infrastructure. The category leader is conventional direct liquid cooling (DLC) using chilled water or glycol circuits routed directly to cold plates on the GPU. The market moved aggressively toward DLC from 2023 onward as air cooling became inadequate for high-density GPU clusters, and DLC has become the standard configuration for NVL72 rack deployments. The incumbents have multi-year supply relationships with hyperscalers and have already qualified their hardware with Nvidia.

The direct competitive threat Ferveret poses to DLC incumbents is limited in the near term by the company's current scale: $125,000 in pre-seed funding and three named customers does not constitute a supply chain threat to Vertiv or CoolIT. What Ferveret represents at this stage is a technical proof point that a different physical mechanism can deliver meaningfully better thermal performance than the liquid cooling approach that has dominated the market for the past three years. The commercial pressure on the incumbents is not immediate, but the existence of a validated alternative creates a long-term benchmark they will need to match.

The most relevant competitive parallel is immersion cooling, a different approach to the same problem. Immersion cooling submerges servers entirely in engineered dielectric fluid, achieving excellent thermal performance and eliminating water consumption, but at the cost of full facility redesign to support immersion tanks, high proprietary fluid costs, and limited compatibility with standard rack configurations. Companies like LiquidStack and Submer have been developing immersion cooling for hyperscale deployments for several years with limited adoption due to these integration barriers. Ferveret claims its APC system can be retrofitted to standard racks, which if true would give it a significantly lower adoption barrier than full immersion while delivering comparable or better efficiency gains.

Hidden Insight: Why Nuclear Cooling Is the Right Mental Model for AI

The framing of "nuclear-inspired cooling" is more than a marketing description. Nuclear reactor thermal management is the domain where heat transfer physics under extreme flux conditions has been most systematically studied over the past 70 years. GPU chips running modern AI workloads achieve local heat flux densities that begin to approach those seen in light water reactor fuel assemblies, measured in megawatts per square meter of surface area. The thermal engineering community that has studied this problem the longest is the nuclear engineering community. Azizian's background is not an accident; it's the right physics training for the right problem.

Phase-change boiling heat transfer was intensively studied in nuclear engineering because the transition from nucleate boiling to film boiling (the Leidenfrost point, or critical heat flux) represents a safety-critical failure mode in reactors. Understanding exactly how fluids behave at the onset of phase change, how to maximize heat transfer before the dangerous transition to film boiling, and how to design surfaces that promote stable nucleate boiling without triggering the transition is the core of nuclear thermal hydraulics. Ferveret is applying that hard-won understanding to chip cooling. The "nuclear" framing is unusual for a startup, but it is scientifically accurate and points to a genuine depth of technical provenance that distinguishes this from the typical cooling startup claiming incremental improvement on a commodity process.

The $125,000 seed number is surprisingly small relative to the technical claims. It suggests the company is at a very early stage, likely relying on MIT laboratory infrastructure and academic collaborations to develop and validate the technology rather than on commercially deployed production systems. The three named customers (CleanSpark, FuriosaAI, and Switch) span a useful breadth of AI workload types: Bitcoin mining GPUs, custom AI accelerators, and hyperscale general compute. If efficiency gains hold across that breadth of hardware, the technology is likely genuinely workload-agnostic rather than specifically tuned to one chip architecture. The UCLA collaboration adds academic validation from a computer science perspective, complementing the MIT engineering provenance on the thermal side.

However, critics note that Ferveret's testing partners are relatively small operators, and the company has not yet published results from a hyperscale deployment at the scale of a Microsoft, Google, or Amazon facility. CleanSpark operates GPU clusters for Bitcoin mining rather than training-scale AI workloads. FuriosaAI is a chip startup, not a hyperscale operator. Switch is a legitimate data center operator but operates at mid-market scale compared to the largest AI cloud operators. The skeptic's question is whether the 35% token-per-watt improvement holds when scaled to racks of 10,000 or more GPUs with the thermal management complexity that entails, or whether it was measured on a small cluster that doesn't capture the full range of real-world thermal edge cases. Ferveret's $125,000 in funding is not enough to have answered that question at scale.

What to Watch Next

The 30-day indicator is whether any of the three named customers publishes independent efficiency data from their Ferveret deployment. CleanSpark files public disclosures as a listed company. If CleanSpark references Ferveret in a quarterly report or operational update with specific efficiency figures, it would constitute independent third-party validation of the performance claims beyond what MIT News published. Watch also for whether Ferveret's Y Combinator affiliation translates into a Demo Day appearance and a disclosed follow-on fundraising round; YC Demo Days typically occur in April and November, and a company admitted to YC would likely be in the Fall 2026 batch with a demo day in November.

At 90 days, the signal that matters most is whether a hyperscale operator announces a pilot of Ferveret's system in a production-scale facility. A pilot from a major cloud provider would validate the scalability question that the current testing partners cannot answer. It would also likely trigger a Series A of $10 million or more, as venture capital follows hyperscale operator validation in infrastructure startups. Conversely, if no hyperscale pilot is announced by September 2026, it suggests the company is still in the early-adopter phase and the technology has not yet crossed the threshold required for serious procurement consideration at the largest AI compute operators. The 180-day view is headcount and geographic expansion: if Ferveret grows from its current MIT-affiliated team to a 20 to 30 person engineering organization with data center facilities staff, it signals the company is preparing for commercial deployment at scale rather than remaining a research project with a commercial wrapper.

Ferveret's insight is that the most demanding thermal engineering domain on earth isn't AI data centers: it's nuclear reactors, and the people who solved that problem have now turned their attention to GPUs.

---