Why a Mixed Approach Is the Norm, Not the Exception
It's tempting to think of liquid cooling and air cooling as a binary choice for a data center, but the reality in most modern GPU-dense facilities is a hybrid split: the highest-heat-density components (GPUs, and increasingly CPUs) are cooled via direct-to-chip liquid cold plates connected to a facility or dedicated cooling loop, while lower-heat-density components (memory modules, VRMs, storage drives, network interface cards) remain air-cooled by traditional rack-level fans and CRAH/CRAC air handlers. This isn't a transitional compromise — it reflects a genuine engineering tradeoff between liquid cooling's superior heat-transfer efficiency and its added plumbing complexity and cost.
Why GPUs Specifically Drive the Liquid-Cooling Decision
Modern AI accelerator GPUs can draw 700W-1000W+ per chip, concentrated in a die area of only a few square centimeters — a heat flux density that air cooling struggles to remove efficiently without extremely high, noisy, energy-intensive airflow. Liquid's much higher volumetric heat capacity (roughly 3,500 times that of air) lets a cold plate remove this concentrated heat with a comparatively small, quiet, low-velocity coolant flow. This is the specific engineering reason GPU-dense racks have driven the recent adoption of liquid cooling in data centers, more than any other single equipment category.
What Typically Stays Air-Cooled
Components that generate less concentrated heat, or that are impractical/undesirable to plumb (removable storage drives, for example, where a liquid connection would complicate hot-swapping), commonly remain air-cooled even in a heavily liquid-cooled rack. This includes memory modules (individually lower power than a GPU die, though aggregate memory heat is non-trivial), voltage regulation modules near the board edge, network interface cards, and storage. The rack-level air handling — usually still delivered by a raised-floor or in-row CRAH system — has to be sized for this residual air-cooled fraction, not for the pre-liquid-cooling total load.
What Actually Determines the Split Percentage
The liquid-cooled fraction of a rack's total heat load depends on server architecture (how many components the manufacturer's cold-plate design covers), GPU density (racks with a higher proportion of GPU-vs-general-compute hardware tend toward a higher liquid fraction), and deliberate design choice (some operators push toward near-100% liquid via immersion cooling specifically to minimize air-handling infrastructure and unlock warmer, more energy-efficient cooling operation). A commonly cited practical range for 2026-era GPU-focused deployments is roughly 60-80% liquid-cooled by heat load, with the remainder air-cooled — though this varies meaningfully by specific hardware generation and facility design philosophy.
Why the Split Matters for Chilled-Water and Air-Handling Sizing
Because liquid and air cooling loops are typically served by separate (though sometimes thermally linked) infrastructure — a chilled-water loop feeding direct-to-chip cold plates or CDUs, and a separate air-handling system feeding CRAH/CRAC units — the split percentage directly determines how much capacity each infrastructure type needs. Underestimating the liquid fraction leaves the chilled-water/CDU system undersized for actual GPU heat rejection; underestimating the air fraction leaves residual air-cooled components under-cooled even if the liquid loop is performing exactly as designed. Both halves of the split need independent sizing, not just a single combined tonnage figure.