MIT uses Ferveret cooling, claiming up to 35% more AI compute for nuclear material research

Ferveret describes its cooling system, a 4U rack‑mount unit called Lambert, as using subcooled boiling to condense vapor bubbles locally, allowing servers to stay within thermal limits while delivering higher performance. In a benchmark with UCLA’s Intelligent Connectivity Laboratory on NVIDIA H200 GPUs, the company reported a 15% increase in server‑level compute measured in TFLOPs per kilowatt and, when combined with a facility‑level power‑usage‑effectiveness improvement from 1.20 to 1.03, an overall gain of roughly 35% usable compute. The announcement notes these figures are company‑issued and lack independent verification or detailed benchmark data.

The platform aims to let researchers run larger models and iterate faster without expanding electrical capacity, but the actual impact on training time, simulation scale, or sustainability depends on many variables, including the specific hardware and workload.