Direct-to-Chip vs Immersion Cooling for AI Racks

For today’s AI racks, direct-to-chip (DTC) liquid cooling is the mainstream answer and the NVIDIA-specified default, while immersion cooling is the chip-agnostic alternative for operators who want the lowest PUE and are willing to rebuild around tanks. Air cooling is no longer in the running at the top of the rack: its practical ceiling is about 30–41 kW, and GB200/GB300 NVL72 racks land near 120–132 kW, with the Rubin generation expected to push toward 200–250 kW. At those densities airflow physically cannot move the heat. DTC circulates coolant through cold plates on the GPUs and CPUs to a coolant distribution unit (CDU); immersion submerges whole servers in dielectric fluid. Both work; they differ sharply on retrofit effort, PUE, cost structure and how far they depart from a standard rack.

This guide compares DTC and immersion on kW density, PUE, retrofit, cost and lead times. Figures are indicative and as of 2026-08; for components and turnkey loops, see our liquid cooling range.

Why air cooling ran out of headroom

AI rack power climbed from around 20 kW a few years ago to 120 kW and beyond today. Air can realistically remove only ~30–41 kW per rack before airflow, fan power and hot-aisle temperatures become unmanageable. Liquid carries heat far more effectively — DTC is roughly three orders of magnitude better than air at the chip — so once a rack crosses the air ceiling, liquid stops being optional. This is not a future trend: it is already the design point for GB200-class systems shipping now.

Direct-to-chip: cold plates and CDUs

DTC mounts copper microchannel cold plates directly on each GPU and CPU. Coolant flows through a rack manifold, into the plates via blind-mate quick-disconnects, and back to a CDU that transfers the heat to a facility water loop (liquid-to-liquid) or, where there is no facility water, to air (liquid-to-air). In-rack CDUs reach up to ~250 kW; row-based units scale into the megawatts to feed many racks. It is the least disruptive way to hit high density because the rack keeps its familiar shape. As a rough rollup, DTC thermal content runs about $55,000 per GB200 rack — server-level cold plates and quick-disconnects, an in-rack CDU, and manifolds combined. All prices are indicative.

Immersion: single-phase and two-phase

Immersion drops whole servers into a bath of dielectric fluid. Single-phase circulates the liquid to a CDU; two-phase lets the fluid boil and condense to move heat by phase change. Because every component sits in fluid, immersion is chip-agnostic and reaches very low PUE — roughly 1.03–1.06 — and handles up to ~250 kW per rack. The trade-off is infrastructure: tanks, fluid handling, and changed service procedures. Indicative CAPEX runs about $4,300–6,800 per kW, and a 42U tank with fluid and CDU lands around $28,000–36,000. It is a larger departure from a standard rack than DTC.

Retrofit, cost and lead times

For an existing hall, a rear-door heat exchanger (RDHx) is the common bridge — roughly 50–100 kW per rack with little room disruption — buying time before a full DTC or immersion build. New high-density halls go DTC; operators chasing the lowest PUE or chip-agnostic flexibility choose immersion. The sourcing reality matters as much as the technology: cold plates, quick-disconnects and rack CDUs carry roughly 12–18 month lead times, with a few qualified fabricators holding most capacity. OCP UQD/UQDB quick-disconnect standardisation lets qualified alternatives drop in without redesigning the manifold, which opens second-source Taiwan/China supply. We help buyers secure that capacity across our cooling desk before it gates a build.

FAQ

Why can’t air cooling handle modern AI racks? The practical ceiling for air cooling is roughly 30–41 kW per rack. GB200/GB300 NVL72 racks land around 120–132 kW, and the Rubin generation is expected to push toward 200–250 kW. Those densities are physically beyond what airflow can remove, which is why liquid cooling is now the NVIDIA-specified default for rack-scale AI.

What is direct-to-chip (DTC) liquid cooling? DTC circulates coolant through copper cold plates mounted directly on GPUs and CPUs, carrying heat to a coolant distribution unit (CDU) that rejects it to a facility loop or air. It removes heat far more effectively than air and is the NVIDIA-specified default for GB200-class racks. It cools the hot chips directly while other components may still use some airflow.

How does immersion cooling differ from direct-to-chip? Immersion submerges whole servers in a dielectric fluid, single-phase or two-phase, cooling every component at once. It is chip-agnostic and can reach very low PUE (~1.03–1.06), but it needs new tank infrastructure, fluid handling and serviceability changes — a bigger departure from a standard rack than DTC.

Which cooling method is best for retrofitting an existing data center? For a brownfield hall, a rear-door heat exchanger (RDHx) is the usual bridge, handling roughly 50–100 kW per rack with minimal room changes. Full immersion needs the most retrofit work. DTC sits in between and is the path for new high-density AI deployments.

What are the lead times for AI liquid cooling hardware? As of 2026-08, cold plates, quick-disconnects and rack CDUs carry roughly 12–18 month lead times, with a few qualified fabricators holding most capacity. That makes early sourcing and qualified second-source Taiwan/China supply a real advantage — plan cooling procurement alongside the GPUs, not after.

Bottom line

Choose DTC for the standard high-density AI build, immersion for the lowest PUE and chip-agnostic flexibility, and an RDHx to bridge a brownfield hall — and source early, because lead times gate the schedule. Browse our cooling range and request a quote for cold plates, CDUs and manifolds.