Ferveret launches nuclear-inspired data-center cooling prototype

Ferveret has unveiled a prototype cooling system that borrows design principles from nuclear reactor thermal management to reduce the energy and water required to cool chips that run large AI workloads. The company, founded by two MIT researchers, disclosed the prototype and its initial test plans on June 10, 2026, targeting hyperscale and edge data centers where chip cooling drives major operating costs.

The approach replaces or supplements conventional air-based and evaporative cooling with a closed-loop thermal fluid architecture that isolates heat at the server rack. Ferveret says the architecture moves heat into a high-heat-capacity working fluid and stores or rejects that heat without relying on continuous evaporative water loss. The system is described as modular, built to retrofit existing rack designs and to integrate with standard data-center power and monitoring systems.

How the system is designed

The core idea adapts steady-state heat management techniques used in nuclear plants, focusing on large thermal capacitance and predictable circulation. Ferveret pairs cold-plate interfaces for processors with a primary coolant loop that captures chip-level heat, then transfers that heat to an intermediate thermal store or a secondary rejection loop. The design emphasizes containment of the working fluid and the ability to smooth peak heat demand, so cooling infrastructure can operate more efficiently under variable AI loads.

Ferveret avoids naming a single coolant in its public materials, describing the medium only as a high-heat-capacity thermal fluid chosen for stability and safety. Control electronics manage pump speed and bypass valves to match circulation to workload, while sensors feed a control plane that regulates fluid temperature and pressure. The company highlights compatibility with facility-level heat rejection systems, including heat exchangers and existing chilled-water plants.

Deployment and testing plans

The startup plans staged testing that begins with lab validation, moves to pilot racks inside a partner data center, and then to larger scale demonstrations. Ferveret declined to name commercial partners at this stage, citing ongoing agreements, but indicated interest from cloud providers and enterprise operators that need to cut water use and operational carbon. The company also said the modular racks are intended to fit existing deployment patterns, reducing the need for wholesale data-center redesign.

Ferveret presented preliminary engineering tradeoffs, noting that higher upfront costs for racks and fluid handling are offset by lower continuous cost for mechanical cooling and reduced potable water consumption. The company argues the design can reduce dependence on evaporative cooling methods that require large volumes of water, and it expects payback periods to vary by climate, electricity price and cooling baseline.

Observers in the data-center industry say the basic idea is not new, pointing to prior liquid cooling efforts and direct-to-chip cold-plate systems, but they also note the novelty in applying reactor-style thermal storage and flow control at rack scale. Analysts expect the economics to hinge on component costs, facility integration complexity, and the degree to which operators accept non-air cooling inside server racks.

Why it matters

If the prototypes scale as Ferveret claims, operators could cut the water footprint and steady-state energy used for cooling AI workloads, which are concentrated and heat-dense. The move signals growing interest in liquid and thermal-storage approaches as conventional air cooling reaches limits in efficiency and sustainability for AI-scale hardware.