The retrofit meeting started badly. The mechanical lead wanted to keep the existing chilled-water plant. The IT director wanted "liquid cooling" for a new 60 kW AI training row.
Procurement asked which one to buy. The facility engineer finally drew two loops on the whiteboard and said the quiet part out loud: these are not competitors. In a modern data center, chilled water is usually the facility heat-rejection path, while liquid cooling describes how heat is captured close to the IT hardware.
The honest answer to "chilled water vs liquid cooling data center" is that most high-density rooms run both, connected through a CDU, and the real design work is deciding where one layer hands heat to the other.
Puntos clave
- Chilled water and rack-level liquid cooling sit on different sides of a heat exchanger, not on opposite sides of a procurement decision.
- The interface point — typically a Coolant Distribution Unit (CDU) — is where the architectural conversation actually happens.
- Warm-water operation on the technology cooling system loop changes what the chilled-water plant has to deliver, sometimes letting it run at higher supply temperatures.
- Retrofits rarely replace chilled water; they extend it with secondary loops for direct-to-chip or rear-door heat exchangers.
- Residual air cooling almost never goes away, even in "liquid-cooled" rooms — power supplies, NICs, and switches still need CRAHs or in-row air.
The Short Answer: These Systems Often Work Together
Treating this as a binary choice is the category mistake that drives bad retrofits. Chilled water is a facility-side utility: a chiller plant, primary and secondary pumps, distribution piping, and terminal units such as CRAHs or fan walls. Liquid cooling, in current data center usage, refers to capturing heat at or inside the IT chassis — direct-to-chip cold plates, rear-door heat exchangers (RDHx), or immersion tanks.
These two layers meet at a heat exchanger. On one side flows facility water; on the other, the technology cooling system (TCS) loop that touches the IT equipment. The CDU is what isolates them, conditions the secondary loop, and keeps IT-side water clean, pressure-controlled, and within a tight temperature band.
Vertiv’s overview of liquid cooling options for data centers frames the same split: facility cooling delivers the bulk thermal capacity, while liquid technologies handle the rack-level heat density that air cannot economically move.
So the right question is not which one to buy. It is: which heat capture method does my IT load require, and how does my chilled-water plant feed it?
Where Chilled Water Ends and Rack-Level Liquid Cooling Begins
The cleanest way to see this is to walk the heat path backward from the chip.
A GPU die rejects heat into a cold plate. The cold plate is plumbed into a manifold inside the rack. The manifold ties into a CDU — sometimes in-row, sometimes in-rack.
Inside the CDU, a brazed-plate or shell-and-tube heat exchanger transfers heat from the TCS loop to the facility water loop. Facility water then carries that heat back to the chilled-water plant, where chillers, cooling towers, or dry coolers reject it to the outdoors.
Up to the CDU primary side, you are in chilled-water territory: ASHRAE W-class facility water, larger pipe diameters, building automation controls, glycol decisions driven by climate. Past the CDU secondary side, you are in technology cooling territory: tighter temperature tolerance, filtration to single-digit microns, biocide chemistry, IT-grade flexible hoses, and quick-disconnects rated for live service. ASHRAE’s HVAC Applications Handbook chapter on data processing and electronic office areas is the standard reference for these water-class definitions.
Rear-door heat exchangers are the hybrid case
RDHx units blur the line. They mount on the back of a rack and use facility water (often directly, sometimes through a CDU) to cool the exhaust air before it leaves the row. The IT chassis itself stays air-cooled.
So an RDHx is "liquid cooling" by piping but "air cooling" from the server’s perspective — useful for retrofits where you cannot open chassis but need to absorb 30–50 kW per rack.
Compare the Two Layers by Function, Temperature, and Retrofit Impact
Attribute | Chilled-Water Plant (Facility Loop) | Rack-Level Liquid Cooling (TCS Loop) |
|---|---|---|
Primary job | Reject building heat to outdoors via chillers/towers | Capture heat at or inside the IT chassis |
Typical fluid | Water or water/glycol | De-ionized water, treated water, or dielectric fluid |
Operating temperature range | Lower supply temperatures when feeding CRAHs; warmer when feeding CDUs | Often warm-water operation, supply temps well above dew point |
Owns this equipment | Chillers, cooling towers, primary pumps, building piping, CRAHs | Cold plates, manifolds, CDU secondary side, RDHx coils, immersion tanks |
Touches IT hardware? | No — stops at CRAH coil or CDU primary | Yes — direct contact with chassis or chip |
Redundancy model | N+1 chillers, looped distribution, dual feeds to CRAHs | Dual CDUs per row, redundant pumps, isolation valves per rack |
Retrofit difficulty | Plant capacity, pipe sizing, valving | Floor penetration, rack manifolds, leak detection, IT downtime windows |
The table also explains why "replace chilled water with liquid cooling" is rarely on the table. You would still need something to reject heat from the CDUs to the outside world — and that something is usually the chilled-water plant you already own.
Three Deployment Patterns for Existing Data Centers
Most retrofits land in one of three patterns. Each one keeps the chilled-water plant and adds a liquid layer on top.
Pattern 1 — RDHx on a dedicated secondary loop. Tap the chilled-water main, add a CDU or a passive heat exchanger skid, and pipe a secondary loop down each row to rear-door coils. Air-cooled servers stay air-cooled. CRAHs handle the rest of the room and residual air cooling for switches and PDUs.
This is the lowest-IT-disruption option.
Pattern 2 — Direct-to-chip with in-row CDUs. Install in-row CDUs per high-density pod. Facility water feeds the CDU primary; cold plates inside servers connect to the secondary manifold. The chilled-water plant may run warmer than before because direct-to-chip tolerates higher supply temperatures.
Vertiv’s guidance on installing and managing CDUs covers the placement, commissioning, and monitoring decisions that drive this pattern.
Pattern 3 — Hybrid row, mixed capture. A single row carries direct-to-chip racks for GPUs, RDHx racks for legacy 2U servers, and air-cooled racks for storage. One in-row CDU can serve both the cold plates and the RDHx coils if the secondary temperature setpoint suits both. CRAHs continue to handle the air-cooled racks and any leakage air from the liquid-cooled ones.
The real mistake teams make in Pattern 2
The most common procurement error is sizing the CDU for nameplate IT power and forgetting the approach temperature of the heat exchanger. If facility water enters the CDU at 18 °C and the heat exchanger has a 3–4 K approach, the secondary loop cannot deliver water cold enough for a cold-plate vendor that specified an inlet near plant supply. The fix is not a bigger CDU — it is either lowering the facility-side setpoint (and giving up some chiller efficiency) or selecting cold plates rated for warm-water operation.
Catching this in design is cheap. Catching it after commissioning means reworking piping or swapping CDUs.
What Changes in the Pumping System
Adding rack-level liquid cooling changes the pumping picture in ways that catch facility teams off guard.
The chilled-water plant gains a new type of load. CRAHs are constant, broadly distributed, and tolerant of pressure variation. CDUs are concentrated, valve-controlled, and can ramp aggressively with IT load.
Secondary pumps that were sized for an even CRAH grid may struggle with the new flow profile when several CDUs modulate at once.
On flow: each CDU pulls a defined GPM at a defined delta-T. Aggregate that across the deployment plan, not just the first pod. On head: long secondary runs to in-row CDUs, plus the heat exchanger pressure drop inside the unit, can exceed what the existing secondary pumps were curve-matched for.
On redundancy: the TCS side has its own pump pair inside each CDU, but the facility-side feed is still a single point of failure unless you provide dual valved connections per CDU.
Retrofit constraints usually show up here first. A plant with adequate thermal capacity can still be flow- or head-limited the moment you add CDUs at the far end of a long pipe run.
Choose the Architecture From the Heat Load Backward
The decision sequence that holds up in practice runs the other direction from how most RFPs are written.
Start with the IT roadmap: what is the per-rack power in 12, 24, and 36 months? Below roughly 20 kW per rack, well-designed air with chilled-water CRAHs is still defensible. Between 20 and 50 kW, RDHx on a chilled-water secondary loop is often the lowest-disruption path.
Above 50 kW per rack — and certainly for dense GPU training clusters — direct-to-chip with CDUs becomes the default, with immersion as a niche option where chassis-level liquid is acceptable.
Next, audit the chilled-water plant. Can it supply water warm enough to make direct-to-chip efficient, and cold enough to still feed CRAHs for residual air cooling? Many plants can, but only with control changes.
Some need a dedicated medium-temperature loop split off the main header.
Then size the interface. CDU count, location, dual-path facility feeds, and leak detection zoning all flow from the rack plan, not from a generic "liquid cooling" line item.
Finally, lock down residual air cooling. Even fully liquid-cooled racks reject 10–20% of their heat to air through power supplies, drives, and unsoftened components. The CRAHs do not retire — they shrink in role.
Preguntas frecuentes
Can I run direct-to-chip cooling without a chilled-water plant at all?
Only in narrow cases. A dry cooler or adiabatic cooler can feed CDUs directly when ambient conditions and warm-water tolerances allow. In most climates and most facilities with mixed loads, the chilled-water plant stays, because storage, networking, and residual air loads still need it.
Does adding CDUs let me raise my chilled-water supply temperature?
Often, yes — but not unilaterally. Cold plates and RDHx coils tolerate warmer facility water than CRAH coils do. If you still feed CRAHs from the same loop, the loop temperature is bounded by the air-side dehumidification and sensible capacity needs of those CRAHs. A split loop or a dedicated medium-temperature header is the usual answer.
How do I handle leak risk in a retrofit above a live white space?
Zone the secondary loop with isolation valves per rack and per row. Use leak-detection cable under manifolds and at every quick-disconnect. Plan the routing so that any failure drops into a contained pan or trench, not onto an energized rack. Commission with water before energizing IT.
What changes for the operations team after a liquid retrofit?
Water chemistry becomes a daily item. The TCS loop needs scheduled sampling, filter changes, and biocide management — the chilled-water plant does too, but at a different cadence and chemistry. Spare parts shift from belts and bearings toward CDU pump cartridges, quick-disconnects, and hose assemblies.
Where do rear-door heat exchangers fit if I am already planning direct-to-chip?
They are useful for handling the residual air heat from direct-to-chip racks, especially in dense rows where CRAH airflow is constrained. Some operators feed both the cold plates and the RDHx from the same in-row CDU, which simplifies piping and balances loop return temperatures.
Conclusión
Anyone framing the decision as chilled water vs liquid cooling in a data center is usually looking at two different floors of the same building and asking which floor to keep. Keep both. The chilled-water plant remains the facility’s heat-rejection backbone; rack-level liquid cooling — direct-to-chip, RDHx, or immersion — is how the densest IT loads get their heat into that backbone without melting through air.
Design the interface, size the CDUs and pumps for the real flow and head profile, plan for residual air cooling, and let the chip-level heat load, not the procurement form, choose the architecture.