Liquid cooling replaces air as the heat-removal path in AI datacentres by moving coolant to the chip (cold plates), around the chip (immersion) or into the die (microfluidics), and analysts now describe it as non-discretionary above roughly 100kW per rack rather than an efficiency option.
Summary
Air cooling removes heat from a server by blowing it across heatsinks. That works while rack power stays modest, but AI accelerator racks have moved from roughly 40kW to 140kW ref, and above about 100kW per rack the air path runs out of practical headroom ref. Liquid cooling substitutes a fluid loop with far higher volumetric heat capacity. There are three main architectures. Direct-to-chip (cold plate) clamps a liquid-carrying plate onto the package lid and pipes heat to a coolant distribution unit; it is the mainstream option and reported to fill the 100 to 175kW density band ref. Single-phase immersion submerges whole boards in dielectric fluid. Two-phase immersion uses a fluid that boils on the hot surface, exploiting latent heat for the highest flux, at the cost of exotic fluids and containment.
The parameters that decide the outcome are rack power density, the thermal resistance budget from junction to coolant, whether the binding resistance sits at the lid TIM or the cold plate (Yole puts the lid TIM as the binding layer above 1.5kW package TDP 2024 Yole Thermal Management 2024), water and facility constraints (Oracle’s February 2026 campuses use closed-loop non-evaporative direct-to-chip ref), and retrofit economics in existing air-cooled halls.
Evidence base
Open questions
- Does the gap between TrendForce’s observed 33% penetration in AI datacentres and Goldman’s forecast 76% of AI servers by end-2026 close, or was the forecast an over-extrapolation?
- How much of the cooling value pool migrates into the package? Does a second top-tier foundry or OSAT beyond TSMC and Intel qualify an in-package microfluidic option, and do design wins follow?
- What is the retrofit economics of converting existing air-cooled halls, and what share of installed capacity is simply stranded for AI workloads? None of the sources address this.
- Does two-phase immersion stay under 15% of new builds through 2030, or does a specific flux or fluid constraint break Yole’s segmentation view?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.