Last updated: October 2026
Most data center racks don't draw much power. Uptime Institute's 2026 survey puts the typical rack at about 11 kW, and the average at just 7.8 kW once facilities built for more than 30 kW are set aside. AI racks live in a different world. NVIDIA's GB300 NVL72 reference architecture calls for up to roughly 142 kW per rack, more than ten times what a typical rack draws. That gap is why data center liquid cooling has become a facility-level planning question.
At AI densities, the problem usually isn't total cooling capacity. It's that the facility has no path to carry heat from the surface of a chip all the way out of the building. Air can't handle heat flux that high. Liquid can.
This guide looks at liquid cooling at the scale of the whole data center: the loops heat travels through on its way outdoors, the components that make up the system, how the heat load splits between liquid and air, the rack densities where liquid becomes necessary, and how to think about cost. If you're new to the topic, start with our introduction to liquid cooling.
What Is Data Center Liquid Cooling?
Data center liquid cooling uses liquid to absorb heat directly at the server or rack. Coolant in the technology cooling system (TCS) carries heat to a coolant distribution unit (CDU), which passes it to the facility water system (FWS) for rejection outdoors. In most deployments, air handles the remaining heat.
A note on definitions: rear-door heat exchangers are often grouped with liquid cooling because they bring water to the rack. Under a stricter definition, though, a system only counts as liquid cooling when liquid makes the first heat transfer from the IT equipment. By that test, a rear door is close-coupled air cooling, because server heat goes into the air before it reaches the water.
The data center liquid cooling heat path at a glance
| Stage | Where it sits | Medium | What it does |
|---|---|---|---|
| ① Chip and server | Cold plate or immersion tank | Coolant | Absorbs heat at the source |
| ② Technology cooling system (TCS) | Rack manifolds, quick disconnects, secondary piping | Clean coolant (e.g., PG25, deionized water) | Carries heat from servers to the CDU |
| ③ CDU | In the rack or in the row | Heat exchanger (liquid-to-liquid or liquid-to-air) | Isolates the IT loop; controls temperature, pressure and flow |
| ④ Facility water system (FWS) and heat rejection | Facility piping and outdoor plant | Facility water or water-glycol | Rejects heat outdoors |
How Liquid Cooling Works in a Data Center: The Heat Path from Chip to Outdoors
Stage 1: Capturing heat at the source. With cold plates, coolant runs through metal plates mounted on the CPUs and GPUs and never touches the chip itself. With immersion, the whole server sits in a dielectric fluid. Either way, the goal is to pick up heat as close to where it's generated as possible.
Stage 2: The technology cooling system (TCS). The TCS is the closed loop that leaves the CDU, runs through the rack manifolds and IT equipment, and returns to the CDU. Under the Open Compute Project's definitions, the CDU gives the IT equipment its own isolated loop, so the TCS and FWS exchange heat but never mix fluid. Because this loop has to pass through the fine channels inside each cold plate, it has to stay very clean. Manifolds split the flow inside the rack, and dripless quick disconnects let technicians pull servers without draining the loop.
Stage 3: The CDU. In a liquid-to-liquid design, the CDU is where the two loops meet. Its heat exchanger moves heat from the TCS to the FWS without letting the two fluids mix. The CDU also controls supply temperature, pressure and flow, and keeps the supply temperature above the dew point to prevent condensation.
Stage 4: The facility water system and heat rejection. The FWS carries heat from the CDU to equipment outside the white space: chillers, cooling towers, dry coolers or some combination. The two loops have very different needs. The TCS needs tightly controlled, clean coolant so it doesn't clog small channels, while the FWS mostly runs ordinary facility water, with glycol added in cold climates to prevent freezing.
Liquid Cooling Technologies Used in Data Centers
"Liquid cooling" in a data center isn't one technology. The table below maps the main options by what absorbs the heat first, where each fits, and how hard it is to add to an existing site.
| Technology | First heat transfer | Best fit | Retrofit difficulty |
|---|---|---|---|
| Direct-to-chip (D2C) cold plates | Liquid | High-power CPU/GPU racks; today's mainstream choice for AI | Moderate: needs CDUs and secondary piping |
| Single-phase or two-phase immersion | Liquid (dielectric fluid) | Very high density, new builds or specialized sites | High: tanks, floor loading and service practices all change |
| Rear-door heat exchanger (RDHx) | Air (then carried off by water) | Moderate density increases, as a supplement to air | Low: mostly a door swap plus a water connection |
| Liquid-assisted air cooling (LAAC) | Liquid inside the server, then air via a radiator | Sites with no facility water that want to pilot | Low |
- D2C cold plates: We cover system-level loops, single-phase vs. two-phase and the pros and cons in our direct-to-chip liquid cooling guide. For the cold plate itself, see what is a direct-to-chip cold plate.
- Immersion: For the trade-offs against D2C, see our direct-to-chip vs. immersion cooling comparison. APALTEK also builds single-phase immersion systems, covered below.
- LAAC: A practical bridge technology. Learn more about liquid-assisted air cooling.
Real projects often mix these. A common pairing uses cold plates for the chips and a rear door for the rest of the rack's air-cooled heat.
Data Center Liquid Cooling System Components
A typical data center liquid cooling system has four parts: the IT-side loop (TCS), the CDU, the facility side (FWS and heat rejection), and controls. Grouping the components by loop makes it easier to see what each one does.
IT Side: The Technology Cooling System
- Cold plates: Mount on CPUs, GPUs and other high-power devices and transfer their heat into the coolant.
- Rack manifolds: Distribute supply coolant from the CDU to each server and collect the return flow.
- Quick disconnects: Let technicians remove a server while the loop stays pressurized, without draining coolant.
- Secondary piping and leak detection: The hoses or hard pipe running from the CDU to the racks, plus leak sensors along the route.
For coolant selection (PG25 vs. deionized water), water quality and corrosion control, see our direct-to-chip coolant guide.
The Middle Layer: CDUs
CDUs are usually classified two ways. By location, in-rack CDUs serve a single rack, while in-row CDUs serve an entire row. By heat exchange, liquid-to-liquid (L2L) units need facility water, while liquid-to-air (L2A) units reject heat back into the room.
Two more decisions belong in the design phase:
- Redundancy: Critical loads often get N+1 CDUs, meaning one extra unit so the system still runs at full load if any single unit fails or goes down for service. That costs floor space and budget, so weigh it early.
- Commissioning and water quality: ASHRAE recommends building a list of every wetted material, planning the clean-and-flush process during design, and following clean construction practices during installation. Whether a closed loop runs reliably for years is largely decided before it ever goes live.
Facility Side: Facility Water and Heat Rejection
The FWS includes primary piping, heat exchange equipment and whatever finally rejects the heat outdoors. Uptime Institute's 2025 cooling systems survey found that air-cooled chillers are still the most common heat rejection method, at about 35% of respondents.
One advantage of liquid cooling is that it can run on warmer supply water, which cuts chiller use, sometimes eliminates chillers entirely, and adds hours of free cooling. Water use varies widely by heat rejection method, and we'll cover that in a separate article.
Controls and Monitoring
A liquid cooling system needs continuous monitoring of supply and return temperatures, differential pressure, flow, coolant quality and leak alarms, with the data feeding a building management system (BMS) or data center infrastructure management (DCIM) platform. When loads swing sharply, control tuning determines whether supply temperature holds steady.
Liquid Cooling Doesn't Remove All the Heat: Planning the Liquid-Air Split
This is one of the most commonly overlooked points in planning. Cold plates cover the main heat sources, mainly CPUs and GPUs. Memory, power supplies, NICs, drives and the rack's network switches often still rely on fans.
- As a rule of thumb, cold plates typically capture 70–85% of a server's heat, leaving 15–30% for air.
- A 2026 study of a hybrid cold-plate-plus-air design found that about 26% of the heat in that case still left through air cooling.
- The share can go higher. In one lab test, a single 53 kW rack reached a liquid heat capture ratio of about 94%.
That means cooling capacity in a liquid-cooled data center has to be planned in two parts: what liquid handles, and what's left for air. There are three common hybrid approaches:
- Cold plates + existing room cooling: The leftover heat goes to the existing air handlers. This works when the density jump is modest.
- Cold plates + L2A CDUs: With no facility water, the CDU rejects the cold plate heat into the room as well. It's the fastest to deploy, but it adds load to room cooling.
- Cold plates + rear-door heat exchangers: The rear door captures the remaining air-side heat and moves it into water, so room cooling load can drop very low.
Coverage also affects efficiency. ASHRAE notes that cooling only the CPU with liquid while leaving memory on air keeps system fan power high and erodes the TCO benefit of liquid cooling. Plan cold plate coverage and the room's remaining air cooling headroom together.
When Do You Need Liquid Cooling? A Rack-Density Decision Guide
Most data center professionals surveyed by Uptime Institute see liquid cooling as necessary above about 20 kW per rack, and air cooling becomes impractical above roughly 50 kW. Rack power is the most direct way to decide whether you need it.
| Rack power | Air cooling status | Typical approach |
|---|---|---|
| Below about 20 kW | Air with good hot/cold aisle containment is usually enough | Air; LAAC for the occasional high-power server |
| 20–50 kW | Air is struggling; required airflow and fan power climb fast | Rear doors, L2A CDUs, or cold plates on the hottest racks |
| 50–100 kW | Air is no longer practical | D2C cold plates with L2L CDUs become the norm |
| Above 100 kW | The hall has to be designed for liquid | Full cold plate coverage, rear doors for residual heat; immersion as an alternative |
Here's what the thresholds are based on:
- 20 kW: In Uptime Institute's 2025 survey, most respondents said air cooling becomes too costly or inefficient above about 20 kW per rack, making direct liquid cooling necessary.
- 20–50 kW: In this range, airflow becomes the bottleneck. ASHRAE notes that a 40–50 kW rack can need up to about 5,000 CFM, while even a best-in-class raised-floor tile delivers about 1,900 CFM.
- 50 kW: According to Uptime, air cooling becomes impractical above roughly 50 kW per rack.
For GPU platform power and cold plate selection, see our guide to choosing a server GPU cold plate and how NVIDIA GPU cold plates have evolved. A new build can be designed around liquid from day one. An existing site is usually better served by a phased retrofit, which we'll cover in a separate article.
How Much Does Data Center Liquid Cooling Cost?
Plenty of people want a single cost-per-MW number for liquid cooling. That number doesn't exist, because different reports measure very different scopes.
Start by getting the scope right. JLL projects that average global data center construction cost will reach about $11.3 million per MW in 2026. That figure covers shell and core only, and it is benchmarked on a single-tenant, air-cooled facility, so it excludes tenant fit-out for AI equipment. Add AI tenant technology fit-out, and costs can climb by up to about $25 million more per MW. That fit-out budget covers high-density power distribution, rack infrastructure, cooling loops and more, so liquid cooling is only one slice of it.
Then look at what liquid cooling changes. Liquid cooling does add piping, CDUs, heat exchange equipment and installation labor. But it usually comes with densification: the same IT load fits in fewer racks, which cuts power distribution, network cabling and building footprint. ASHRAE notes that first cost per MW falls as rack density rises, and past a certain density, building for liquid can cost less up front than air-cooling the same load. For a line-by-line breakdown across the server side, rack distribution, facility side, and commissioning and operations, see the cost section of our direct-to-chip guide.
Operating savings come from three places:
- Fan power: ASHRAE reports that in some dense air-cooled servers, fans can draw 10–20% of server power, or at least 5 kW in a 50 kW rack. Server fans run on the same UPS as the servers, so a rise in fan power from 2% to 10% effectively costs 8% of UPS capacity. Liquid cooling gives much of that power back to compute.
- Cooling plant energy: Liquid cooling can run on warmer supply water, which shortens chiller runtime and extends free cooling hours.
- Heat reuse: Warmer return water is worth more for heat recovery.
Building Liquid-Cooled Infrastructure with APALTEK
APALTEK designs and manufactures liquid cooling solutions for AI data centers, from the cold plate on the chip to the CDU that serves a row. We come at this from a different angle than most industry articles. Our focus is whether each part can be customized to your platform, built consistently at volume, and connected on site without surprises.
- Cold plates: Our server cold plates use vacuum-brazed copper construction, can be customized to your CPU or GPU platform, and can integrate a vapor chamber. Browse our cold plate range.
- CDUs: Designed and built in-house, including in-rack units rated at 5 kW and 150 kW and in-row units rated at 1.5 MW and 2.5 MW.
- Ready to install: Cold plates and manifolds can ship with quick disconnects preinstalled and leak-tested as a complete assembly, which cuts on-site work and risk.
- Quality: Every product we ship passes helium leak testing. Other validation tests are defined around your requirements.
- Immersion: Beyond cold plates, we offer single-phase immersion cooling, including an IDC immersion cooling system for data centers and an all-in-one edge immersion cooling system.
- Engineering support: Customers who order paid prototypes usually receive thermal simulation at no extra charge, and we reply to quote requests within one business day.
Data Center Liquid Cooling FAQ
What are the main components of a data center liquid cooling system?
A typical data center liquid cooling system has four main parts. The IT-side loop, or technology cooling system (TCS), includes cold plates, rack manifolds, quick disconnects and secondary piping, and carries heat away from the chips. The coolant distribution unit (CDU) transfers that heat out of the IT loop. The facility water system (FWS) and heat rejection equipment move it outdoors. Controls and leak detection tie the system together.
What is the difference between TCS and FWS?
The technology cooling system (TCS) is the closed loop that runs from the CDU through the rack manifolds and IT equipment and back to the CDU, using tightly controlled, clean coolant. The facility water system (FWS) carries heat from the CDU to chillers, cooling towers or dry coolers, and its water quality requirements are less strict. The two loops exchange heat in the CDU's heat exchanger, but their fluids never mix.
What percentage of server heat does liquid cooling remove?
Cold plates typically remove 70–85% of a server's heat, depending on which components they cover. Lab tests have measured liquid capture ratios above 90%. The rest is handled by room cooling, L2A CDUs or rear-door heat exchangers.
Is a rear-door heat exchanger considered liquid cooling?
It depends on the definition. Rear doors are often grouped with liquid cooling because they bring water to the rack. Under the stricter definition, where liquid has to make the first heat transfer, they don't qualify: server heat goes into the air first, and the door's coil then transfers it to water. In practice, rear doors are often paired with cold plates to handle the remaining air-side heat.
Can a liquid-cooled data center reject heat without chillers?
In many climates, yes, for much or all of the year. Cold plate systems can accept relatively warm supply water, so heat can often be rejected through dry coolers or cooling towers alone, with mechanical cooling reserved for the hottest days. Whether you can drop chillers entirely depends on local climate, the supply temperature your chips require, and how you balance energy use against water use. Plan with some margin: ASHRAE expects rising chip power to push required facility water temperatures down over time.
Next Step: Map Where the Heat Goes, Then Choose the Hardware
Data center liquid cooling isn't finished once the cold plates are installed. It's a complete system that runs from the chip to the outdoors. A sound order of operations is to set rack power and decide how much heat liquid and air will each carry, then confirm whether you have facility water and at what supply temperature, and only then choose cold plates, CDUs and heat rejection equipment.
If you're planning liquid-cooled racks or a liquid-cooled hall, send your rack power, chip platform and facility water conditions to APALTEK's engineering team. We'll evaluate the right combination of cold plates and CDUs and get back to you with a quote within one business day.