Direct-to-chip liquid cooling uses cold plates bonded to CPUs/GPUs to circulate liquid (typically water or a water/glycol mix) and extract heat directly from the silicon. In 2026, it is the mandatory standard for high-density AI accelerators exceeding 1 kW TDP, completely displacing traditional forced-air convection in AI clusters.
Air cooling reaches physical limits around 20-30 kW per rack. D2C cooling uses microchannel cold plates where convective heat transfer coefficients are orders of magnitude higher than air. A Coolant Distribution Unit (CDU) manages the secondary liquid loop, interfacing with the facility water system (primary loop) to reject heat.
Cold Plate Design
Skived copper microchannels are engineered into cold plates attached to the chip via a high-performance Thermal Interface Material (TIM).
CDU Operation
Coolant Distribution Units (CDUs) pump the working fluid through the server loops, tightly regulating flow, pressure, and maintaining supply temperatures around 30-45°C.
Manifold Distribution
Rack manifolds distribute coolant to individual servers via drip-less quick disconnect (QD) fittings, ensuring concurrent maintenance capability.
Heat Rejection
The heat is transferred to a facility water loop via a plate heat exchanger inside the CDU, then rejected outdoors via cooling towers or dry coolers.
- Leakage Risk: Despite advanced quick disconnects, introducing pressurized water into server chassis remains a severe catastrophic risk requiring sophisticated leak detection tape.
- Weight and Floor Loading: Liquid-filled racks exceed 1,500 kg, requiring expensive structural reinforcement of raised floors in legacy data centers.
- Supply Chain: The exponential demand for brazed plate heat exchangers and specialized CDUs has created 12-18 month lead times.