Clivet Magnetic Cooling Splits the Compressor Load

Clivet magnetic cooling introduces a dual-rotor, independently controlled compressor architecture intended to follow the wide load swings of high-density AI data centres. Clivet demonstrated the DUO POWER BOOST system at AI Infra Summit 2026 on 16 September and disclosed the wider launch through a company release on 19 September. The useful idea is a broader operating envelope; the missing evidence is audited performance across real workloads.

How Clivet magnetic cooling is meant to work

A conventional cooling plant has to remain stable as compute demand, water temperatures and outdoor conditions change. Clivet’s announced design gives the two compression stages separate speed control. In principle, that lets the system match different flow and pressure combinations without forcing both stages to follow one fixed response.

The AI Infra Summit agenda independently confirms that Clivet presented “DUO POWER BOOST Independent-Rotor Differential Speed” magnetic cooling for AI data centres. Sina Finance and Viewpoint separately reported the 16 September demonstration and the dual-rotor mechanism. Those reports corroborate the event and architecture, but the detailed performance figures remain company claims.

The portfolio matters more than one compressor

Clivet also showed cooling-source equipment, coolant distribution units, fan walls and terminal equipment, alongside a power-and-cooling coordination display with CLOU. That systems framing is sensible because dense AI racks turn cooling into an end-to-end design problem. A highly efficient compressor cannot rescue undersized distribution, poor controls or a heat-rejection bottleneck.

The company says a large distribution unit can cover up to 3 MW at a 3K approach temperature or 5 MW at 5K, with predictive temperature control and online component replacement. These are specific, testable claims, but Lapaas Voice treats them as attributed specifications until an operator or independent laboratory publishes results.

What operators should ask before deployment

Procurement teams should request part-load efficiency curves, ambient-temperature derating, rotor-failure behaviour, coolant compatibility, service procedures and measured power usage at the system boundary. They should also model whether added control complexity creates a new maintenance dependency.

Infrastructure launches often shift constraints rather than remove them. Lapaas Voice’s report on the Huawei AI Cluster Service rollout separated announced scale from deployment dates. Our coverage of the AI Energy Management Alliance examined how flexible computing load can interact with the grid. Clivet’s proposition is the thermal side of the same operational problem.

The decision is evidence, not novelty

Independent rotor control is an intelligible response to variable AI load, and the event record confirms that Clivet put the design into a public industry programme. That makes the launch more than a concept slide. It still does not prove lower total facility energy, better uptime or cheaper lifecycle operation.

Clivet magnetic cooling deserves a pilot where telemetry can be compared against a conventional baseline under identical load and weather conditions. If the efficiency and reliability curves hold across the full system, the design could widen the useful operating range of AI cooling. Until then, buyers should treat it as a testable engineering claim, not a guaranteed outcome.