Direct-to-Chip Liquid Cooling

Overview

Direct-to-chip (DTC) liquid cooling — also called direct liquid cooling (DLC) — removes heat from processor packages by circulating coolant through cold plates mounted directly on CPUs and GPUs. NVIDIA's NVL72 rack at 120 kW/rack in 2024 was the industry inflection point at which air cooling became infeasible for high-density AI compute, making DTC liquid cooling the mandatory rather than optional choice. Rack densities are projected to continue rising from 120 kW/rack (2024) to 160 kW (2025), 300 kW (2026), and approximately 1,000 kW (2027), raising the open question of what the cooling solution for 1 MW/rack will be.

Three approaches exist to meet increasing cooling capacity requirements: new coolant types (nanofluids, two-phase refrigerants) requiring long-term material compatibility validation; improved heat transfer surface area (cold plate structural design, cost, and production feasibility); and higher flow rates (which increase pressure drop and data center footprint). AI training workloads add a control challenge: GPU thermal load varies frequently across training stages, requiring CDUs with rapid, stable response to prevent thermal excursions.