MIT Startup Unveils Water-Free Cooling for AI Data Centers
Ferveret's nuclear-inspired system targets 35% efficiency gain and zero water use
MIT-born startup Ferveret has launched a modular cooling system that adapts nuclear reactor heat transfer technology to cool AI chips without using a single drop of water. By significantly reducing the energy required for thermal management, the company claims its solution allows data centers to generate 35% more AI tokens using the same amount of power.
Key details
Ferveret’s Adaptive Phase Cooling (APC) solution utilizes a process known as subcooled boiling, a technique originally developed for high-heat-flux nuclear reactor cores. The system submerges servers in a specialized, non-toxic liquid that produces micro-bubbles at the chip surface. These bubbles detach rapidly and recondense, transferring heat far more effectively than traditional air-cooled or standard immersion systems.
In a collaborative study with UCLA, the company demonstrated a 15% improvement in computational power efficiency compared to existing liquid cooling benchmarks. Combined with real-time software that optimizes power delivery to individual servers, data center operators can achieve a 35% increase in AI model output (tokens) within their existing power constraints.
Why this matters
As AI data centers are projected to consume up to 17% of U.S. electricity by 2030, cooling remains one of the largest non-computational energy drains. Furthermore, the industry is facing increasing pushback over water consumption. Ferveret’s system addresses both bottlenecks simultaneously: it eliminates the need for cooling towers and evaporated water while freeing up a massive portion of the power budget for actual AI processing.
Context
The transition from air cooling to liquid immersion is already underway as chip power densities exceed the physical limits of fans. However, most current immersion systems require complex infrastructure to manage pressure and fluid boiling. Ferveret’s modular, rack-mounted "boxes" aim to simplify this transition by making advanced liquid cooling compatible with existing data center layouts.
What happens next
Ferveret is currently testing its APC solution with major industry players, including data center operator Switch and AI accelerator firm FuriosaAI. The company is also part of Nvidia’s Inception program. Following successful pilot programs, Ferveret plans to scale production later this year, targeting hyperscale cloud providers and data centers in water-scarce regions where traditional cooling is no longer viable.
Source: MIT News Published on AI Usage Global, author: AUG Bot
