Glossary
direct-to-chip cooling
A cooling architecture that pipes liquid coolant directly to a cold plate on each processor, evacuating the 700+ watts per GPU that air cooling cannot handle.
A cooling design where liquid coolant flows through a cold plate bolted to each GPUsiliconA massively parallel processor originally designed for graphics, repurposed since the 2010s as the dominant compute substrate for both training and inference of large neural networks. Open full entry or CPU, removing heat at the source rather than by moving air through the rack. The mechanical complexity is real (each chip is plumbed to a coolant loop) but the thermal capacity is much higher than air: a single rack can dissipate 100+ kW of heat reliably, versus 30 to 40 kW with the best air cooling.
For AI data centers DLC became mainstream in 2024 with NVIDIA Hopper H200 and Blackwell B100/B200 deployments. Blackwell’s GB200 NVL72 rack architecture is liquid-cooled by design and cannot be air-cooled at full power. New AI factoryinfrastructureA purpose-built data center optimized for AI training rather than general cloud workloads, characterized by liquid-cooled high-density GPU racks, gigawatt-scale single-tenant power, and tightly-coupled networking. Open full entry builds are universally DLC-ready or DLC-required; retrofitting older air-cooled facilities is involved and often economically marginal versus new construction.
A related variant, immersion cooling, fully submerges the chips in dielectric fluid. Immersion has higher thermal capacity still but adds operational complexity; direct-to-chip is the production standard through 2026.