Liquid Cooling for AI Data Centers: DTC vs Immersion

Direct-to-chip (DTC) liquid cooling handles 40-120 kW per rack at roughly $28,000-42,000 per 42U rack in retrofit costs, while immersion cooling scales past 150-250 kW per rack but requires redesigned chassis and runs $4,500-6,800 per kW of IT load installed. For most operators deploying AI racks in the 40-100 kW range today, DTC is the default; immersion is reserved for the highest-density or most sustainability-constrained deployments.

Key takeaways

• DTC dominates current deployments. NVIDIA’s GB200 NVL72 (120 kW nominal, 130-132 kW measured) ships with direct-to-chip cooling by default, and most 2026 colocation retrofits follow the same path.
• Immersion wins on density ceiling. Single-phase immersion supports 100-150+ kW/rack; two-phase immersion reaches 150-250+ kW with PUE as low as 1.01-1.03, per vendor and integrator data.
• Retrofits beat greenfield on cost. Converting an existing hall to liquid cooling runs about $2 million per MW versus $11 million-plus per MW for a new liquid-ready build — roughly an 80% capital saving, per STL Partners.
• Adoption is still a minority, high-density practice. Uptime Institute’s 2025 survey found 22% of operators use direct liquid cooling, and even among adopters, liquid loops typically serve under 10% of total racks.
• Power, not plumbing, is the retrofit bottleneck. A site without spare electrical headroom cannot support liquid-cooled racks no matter how good its piping design is.
• PUE drops meaningfully. Air-cooled hall PUE of 1.3-1.6 falls to roughly 1.05-1.15 with DTC and 1.02-1.10 with well-run immersion, per Vertiv and integrator field data.
• The liquid cooling market is still small relative to total data center spend — Dell’Oro projects it reaching $7-8 billion in manufacturer revenue by 2029-2030, a fraction of the roughly $1 trillion in 2026 data center capex.

For current per-kW colocation pricing across markets that support high-density deployments, see the colocation price index. For rack-level GPU rental economics that drive density decisions, see GPU cloud pricing.

Why air cooling stops working above 30-40 kW/rack

Air cooling removes heat by moving large volumes of chilled air across server surfaces. The approach scales reasonably to roughly 20-30 kW per rack with hot-aisle containment and in-row cooling, and with engineering effort can reach 30-40 kW. Past that point, the airflow required becomes physically impractical — fan power rises faster than useful cooling, and hot spots form on the densest components (GPUs, HBM stacks) regardless of total room airflow.

AI accelerators pushed past this ceiling within a single hardware generation. NVIDIA’s H100 SXM draws roughly 700 W per GPU; a fully populated 8-GPU air-cooled server can still work within a standard 10-15 kW rack. The Blackwell-generation GB200 NVL72 changes the math: each Grace Blackwell Superchip dissipates up to 1,200 W, and the full 72-GPU rack draws 120 kW nominal — 130-132 kW measured in deployed configurations, per NVIDIA and integrator documentation. That rack ships with direct-to-chip cold plates and an in-rack coolant distribution unit (CDU) as standard equipment, not an option.

Direct-to-chip cooling: how it works and what it costs

Direct-to-chip (DTC) cooling routes liquid coolant through cold plates bolted directly to the CPU, GPU, and other high-TDP components. A manifold distributes coolant to each cold plate; a CDU sits between the facility’s water loop and this sealed secondary loop, controlling flow rate, temperature, and pressure while keeping facility water quality isolated from the electronics.

DTC removes roughly 70-80% of a server’s total heat load directly to liquid — memory, storage, networking, and power delivery components still rely on residual airflow, so a DTC rack still needs some room-level air handling, just far less than a fully air-cooled hall.

Attribute	Direct-to-chip	Notes
Practical rack density	40-120 kW	Matches current AI accelerator platforms (GB200 NVL72: 120-132 kW)
Cost per 42U rack (retrofit)	$28,000-42,000	Cold plates, manifolds, in-rack CDU; 2026 integrator pricing
PUE impact	Falls to ~1.05-1.15 from 1.3-1.6 air-cooled baseline	Vertiv field data
Hardware compatibility	Requires liquid-ready servers with cold-plate mounts	Cannot be added to arbitrary legacy hardware
Facility water temperature	ASHRAE W27-W45 classes typical	Warmer classes (W40/W45) allow dry-cooler-only rejection much of the year

DTC’s main advantage is that it works with server form factors close to today’s standard 19-inch rack, so operators can retrofit existing floor space rather than rebuilding from the slab up — provided the site has spare electrical capacity.

Immersion cooling: single-phase, two-phase, and where it fits

Immersion cooling submerges entire server boards in a dielectric fluid that does not conduct electricity, removing heat from every component simultaneously rather than just the hottest ones.

• Single-phase immersion: fluid stays liquid throughout the cycle, circulated to an external heat exchanger. Supports densities above 100 kW per rack-equivalent tank with PUE approaching 1.03. Installed cost runs roughly $4,500-6,800 per kW of IT load, covering tanks, dielectric fluid, coolant distribution units, and heat rejection equipment.
• Two-phase immersion: fluid boils at the component surface and recondenses at the tank lid, moving heat far more efficiently. Delivers PUE of 1.01-1.03 and supports 150-250+ kW per rack-equivalent, but costs $65,000-95,000 per full rack-equivalent and involves engineered fluids with their own handling and environmental considerations.

Immersion requires servers designed for submersion — no spinning drives, and connectors and thermal paste rated for the fluid — so it is effectively a hardware redesign, not a drop-in retrofit. That makes it best suited to greenfield high-performance computing halls, specialized AI training clusters, or edge/embedded deployments where extreme density in a small footprint matters more than compatibility with commodity servers.

Direct-to-chip vs immersion: side-by-side

Factor	Direct-to-chip	Single-phase immersion	Two-phase immersion
Rack density ceiling	40-120 kW	100-150+ kW	150-250+ kW
Typical PUE	1.05-1.15	~1.02-1.05	1.01-1.03
Retrofit into existing racks	Yes, with liquid-ready servers	Difficult — needs tanks, not racks	Difficult — needs tanks, engineered fluid
Relative cost per rack/kW	Lower	Mid	Highest
Hardware compatibility	Standard chassis with cold-plate mounts	Requires submersion-rated hardware	Requires submersion-rated hardware
Best fit	AI inference/training racks at 40-120 kW, colocation retrofits	Dense HPC clusters, space-constrained sites	Extreme-density research/training clusters

What it costs to retrofit a colocation facility

Facility-level economics favor retrofitting existing halls over building new liquid-ready capacity. STL Partners estimates liquid cooling retrofits at roughly $2 million per MW, against $11 million-plus per MW for a comparable new liquid-cooled build — an approximately 80% capital saving. That gap is why 65% of new CDU installations in North American colocation facilities in 2026 are retrofits rather than greenfield deployments, per Nautilus Data Technologies.

The retrofit bottleneck is rarely the piping itself. It is electrical headroom: a facility needs spare power capacity — and often a busway or switchgear upgrade — before it can support 40+ kW racks, independent of how well the cooling loop is engineered. Facilities with committed power below their nameplate capacity, or with slack in existing substations, retrofit fastest. Fully leased, power-constrained sites (Northern Virginia primary markets, for instance, sit at roughly 0.3% vacancy per CBRE’s H2 2025 data) have the least room to add cooling capacity without a broader power upgrade.

For colocation providers specifically, retrofits carry an added complication that hyperscalers building for themselves do not face: converting a live facility risks disrupting existing tenants, and operators typically want committed customer demand and a migration plan locked in before starting work.

ASHRAE temperature classes: why “liquid cooling” isn’t one standard

ASHRAE’s Technical Committee 9.9 defines liquid cooling facility water temperature classes — W17, W27, W32, W40, W45, and W+ — where the number denotes the maximum supply water temperature in Celsius. All classes share a 2°C lower bound.

The class chosen determines whether a facility needs mechanical chillers at all:

• W17-W27: requires chilled water, similar cooling plant to a traditional air-cooled hall.
• W32: intermediate; reduces chiller runtime hours in most climates.
• W40-W45 and W+: warm enough that dry coolers or cooling towers alone can reject heat for most or all of the year, eliminating mechanical refrigeration and cutting both energy use and water consumption from evaporative cooling.

Facilities designed around W40+ classes get the largest efficiency and water-use benefit, but require hardware validated for warmer inlet temperatures — a design decision made at rack and server procurement time, not something a facility can retrofit after the fact without derating.

Water usage: it depends on design, not the cooling method’s label

Liquid cooling is not automatically water-saving. A direct-to-chip facility still relying on evaporative cooling towers for its remaining air-cooled load sees limited water-use improvement over a comparable air-cooled hall — the savings come from running at warmer ASHRAE classes (W40/W45) that allow dry-cooler-only heat rejection, not from the presence of liquid loops per se. Immersion cooling’s closed dielectric loop can eliminate evaporative water use in the immersion-served portion of the load, which is one reason it gets attention in water-constrained markets. Operators should ask for a facility’s measured water usage effectiveness (WUE) figure rather than assuming any liquid-cooled site is automatically more water-efficient than an air-cooled one.

Market size and adoption trajectory

Uptime Institute’s 2025 Cooling Systems Survey found 22% of data center operators already use direct liquid cooling — unchanged from 2024 — with an additional 61% saying they don’t currently use it but would consider it. Even among adopters, liquid loops typically serve fewer than 10% of racks in a given facility, reflecting how concentrated the demand still is in AI training and inference halls rather than general-purpose compute.

Dell’Oro Group forecasts the worldwide data center liquid cooling market reaching roughly $7 billion in manufacturer revenue by 2029, with direct liquid cooling specifically surpassing $8 billion by 2030 as it shifts from an enabling option to a foundational requirement for AI-dedicated capacity. That is still a small slice of the roughly $1 trillion in global data center capex expected in 2026 — liquid cooling remains a high-density specialization, not the default across the industry’s total footprint.

What this means for colocation buyers

1. Ask for the rack’s actual density support, not a marketing label. “Liquid-cooling-ready” can mean anything from a facility with spare piping stubs to a fully commissioned CDU loop tested at your target kW. Get the tested density and the coolant temperature class in writing.
2. Match cooling type to your hardware roadmap, not the other way around. If you are deploying GB200-class racks now, DTC compatibility is close to mandatory. If you are evaluating extreme-density HPC or custom silicon at 150+ kW, immersion may be the only option — check server compatibility before committing to either.
3. Verify power headroom before cooling headroom. A site can have excellent liquid cooling infrastructure and still be unable to add your rack if it lacks spare electrical capacity — confirm both in the same conversation.
4. Use retrofit economics as a negotiating reference. At roughly $2 million per MW for a retrofit versus $11 million-plus for new builds, an operator quoting premium liquid-cooling surcharges on a retrofitted facility has a materially lower cost base than one at a genuinely new liquid-ready site — factor that into rate discussions.
5. Check the WUE figure, not just the PUE figure, if water availability is a constraint in your target market — the two do not move together automatically.

Compare per-kW pricing across colocation markets that support high-density, liquid-cooled deployments in the colocation price index, browse specific facilities in the data center catalog, or request a quote benchmarked against current index rates.