AI data center cooling is becoming a physical limit on AI growth as next-generation GPU racks push toward 142kW of power. Why can't conventional air cooling keep up — and can liquid cooling solve the thermal problem?

In this episode of Silicon to Software, Imran Valiani examines the engineering problem hiding underneath the AI infrastructure boom: getting enormous amounts of heat out of increasingly dense GPU systems.

NVIDIA's GPU power trajectory illustrates what's happening.

H100 reaches up to 700W in its SXM configuration. B200 reaches up to 1,000W in the HGX B200 configuration. Blackwell Ultra pushes the maximum envelope to 1,400W per GPU in the mandatory-liquid-cooled GB300 NVL72 rack architecture.

A complete GB300 NVL72 rack can require up to 142kW.

Nearly all of that electrical energy ultimately becomes heat t ...