Liquid Cooling Is No Longer an Accessory

AI infrastructure is reorganizing around its physical layer. As training and inference clusters scale, the binding constraint is shifting from the chip itself to the substrate around it: coolant, plumbing, facility water, and the energy overhead of moving heat off the chips.

TrendForce's 2026 forecast captures the shift. Liquid-cooling penetration among AI chips rises from 6% in 2023 to 53% in 2026, a 47-percentage-point swing in three years, with 59% forecast for 2027.

High-end AI processors now exceed 1 kW TDP, and advanced rack-scale systems consume hundreds of kilowatts each. At that density, air cooling stops being a serious option, so the choice becomes which liquid loop to use. NVIDIA's Vera Rubin platform goes fanless and all-liquid, AMD's Helios ramps from H2 2026 with full liquid adoption in MI450 and MI500, and rack-scale shipments are forecast to double.

The reframing matters because most coverage reads 53% as adoption. It is closer to category transition: liquid cooling appears to have moved from accessory to co-equal infrastructure tier alongside GPUs, networking, and power. The mechanism, density crossing the threshold where air cooling stops being the default, will repeat in every compute cycle that pushes TDP higher.

Caveats belong in the frame. Coolant supply chains, the energy overhead of liquid loops, and air cooling's edge and lower-density viability are real, and they bound the arc. The winners sit in the physical layer underneath the chip: coolant, plumbing, facility water, the energy grids. The losers are air-cooling incumbents at the high end.