On most direct-to-chip (D2C) liquid cooling projects, the cold plates and server platform get locked in early. Then the team reaches the coolant distribution unit (CDU) and stalls on the in-rack vs. in-row question: does every rack get its own CDU, or does a whole row share one?
It sounds like a question about where to put a box. It isn't. That choice decides how you build redundancy, how much compute a single fault can take down, whether you give up floor space, how the secondary piping runs, and how you'll scale later. Once the racks are populated and the pipework is in, changing course gets expensive.
If you need a refresher on what a CDU does, start with what a CDU is and how it works, or see where it sits in a full data center liquid cooling system. This guide goes straight to the trade-offs, then closes with a decision matrix organized by deployment type.
What Is the Difference Between an In-Rack and an In-Row CDU?
An in-rack CDU mounts inside the cabinet and serves only that rack, typically from a few kW to about 250 kW. An in-row CDU is a floor-standing unit that feeds several racks through row manifolds, with capacities from several hundred kW to multiple MW. Use in-rack for rack-by-rack rollouts and in-row for dense rows at scale.
Key takeaways:
- Failure domain: An in-rack CDU failure affects one rack; an in-row CDU failure can affect a whole row.
- Space: In-rack CDUs use rack units but no floor space; in-row CDUs use floor space but no rack units.
- Redundancy: In-rack relies on redundancy inside each unit; in-row usually adds N+1 units on top of that.
- Scaling: In-rack grows one rack at a time; in-row is sized for the full row up front.
In-Rack vs. In-Row CDU at a Glance
| Factor | In-Rack CDU | In-Row CDU |
|---|---|---|
| Location | Inside the rack, uses rack units | Floor-standing cabinet in or at the end of a row |
| Capacity per unit | A few kW to about 250 kW | Several hundred kW to multiple MW |
| Racks served | One | Multiple racks in a row |
| Failure domain | Limited to one rack | Can affect an entire row |
| How redundancy is built | Mostly inside the unit: redundant pumps, sensors and power | Inside the unit, plus N+1 units across the row |
| Secondary coolant volume | Short loop, small volume | Long loop, large volume |
| Piping complexity | Low, mostly contained in the rack | High, needs row headers and manifolds |
| How you scale | Add one unit per rack | Deploy per row with built-in headroom |
Above the in-row tier sits a third category: gallery or facility-level CDUs installed along the perimeter or in mechanical corridors, serving several rows or an entire hall. They behave more like facility plant and suit large new campuses, so we won't cover them here. This guide also assumes liquid-to-liquid CDUs connected to facility water; if your site has none, see the liquid-to-air path in our air-cooled data center retrofit guide.
Capacity and Rack Density: Where Each CDU Type Fits
Two numbers drive the decision first: the liquid heat load per rack and how many racks you need to cool.
An in-rack CDU is capped by the space inside the cabinet, since the pumps, heat exchanger and filters all have to fit in a few rack units. Take APALTEK's 150 kW in-rack CDU: it's 4U tall (175 × 445 × 900 mm), runs PG25 on the secondary side at a rated 115 L/min, and removes 150 kW from a single rack at rated conditions. If your GPU racks stay at or below 150 kW of liquid heat load, one unit per rack covers it.
An in-row CDU isn't limited by rack space, so it can carry a much larger heat exchanger and pump set. APALTEK's 1.5 MW in-row CDU is rated at 1,500 kW at a 4.5°C approach temperature difference (ATD) with 1,800 L/min of secondary flow, and the 2.5 MW in-row CDU is rated at 2,500 kW at 5°C ATD with 2,200 L/min. Across the industry, 2 MW in-row CDUs are now built to an open hardware specification, a sign that this capacity tier is becoming standard for large AI clusters.
A quick sanity check: at 130 kW of liquid heat load per rack, a 1.5 MW in-row CDU can support roughly 11 racks at rated conditions, before you add redundancy. Once you're past ten or so racks clustered in one row, in-row starts winning on unit count.
When you compare capacity numbers, always check the rating conditions: approach temperature, facility water supply temperature, secondary flow and whether the filter was installed during the test. Two units that both say "150 kW" can perform very differently if they were rated under different conditions. That's why the OCP test methodology and performance rating for liquid-to-liquid CDUs exists: to put vendor data on the same baseline.
Failure Domain and Redundancy: Plan N, N+1 or 2N at Four Levels
Capacity tells you whether you can cool the load. Redundancy tells you whether a fault will take that load offline. This is where in-rack and in-row CDUs differ the most.
A failure domain is how far a single fault can spread. An in-rack CDU's failure domain is one rack: if it goes down, the worst case is that one rack throttles or shuts down while its neighbors keep running. An in-row CDU's failure domain is the row: if it feeds 10 racks with no backup, one fault can hit all of them at once.
That's why the two types build redundancy differently:
- In-rack: redundancy lives inside the unit. APALTEK's 150 kW in-rack CDU uses 1+1 redundant pumps, and either pump can hold rated flow on its own. The pumps are hot-swappable with blind-mate liquid and electrical connections, and the critical pressure and temperature sensors are also 1+1 redundant. A failed pump or sensor can be replaced without shutting the rack down.
- In-row: redundancy inside the unit, plus extra units. APALTEK's in-row CDUs also use 1+1 redundant pumps and add an automatic transfer switch (ATS) on the power input. But because one unit covers a whole row, critical loads usually add N+1 CDUs as well: one extra unit, so the rest can still carry full load when any single unit fails or is down for service.
Redundancy doesn't stop at the CDU. Availability has to be specified separately at four levels: the pumps inside the CDU, the number of CDUs, the secondary distribution loop and the facility heat rejection plant. The system is only as available as its weakest level. Buy N+1 CDUs but run them on a single secondary header or a single facility water path, and that path is still a single point of failure.
It's the same logic behind Tier III concurrent maintainability: any component can be taken out of service for maintenance or replacement without shutting down IT operations, which takes redundant components plus a second distribution path. ASHRAE likewise recommends active redundancy so cooling continues through the switchover from primary to backup systems, with a CDU providing a clear boundary between the facility water system (FWS) and the technology cooling system (TCS).
Stacking Two In-Rack CDUs vs. Moving to One In-Row Unit
Two in-rack CDUs can share one cabinet, either as 1+1 redundancy at the rack level or for more combined capacity. The cost is double the rack space (8U for two 150 kW units) plus extra in-rack piping and control coordination. If most racks in a row end up needing two units, that's usually the point where an in-row CDU with N+1 becomes the better deal.
Floor Space, Rack Units and Piping
An in-rack CDU uses no floor space, but it does use rack units. Those are slots that could otherwise hold servers, and in an AI rack where every U is expensive, that's a real opportunity cost. The upside is that the secondary loop stays almost entirely inside the cabinet.
An in-row CDU uses no rack units, but it takes floor space. APALTEK's 1.5 MW unit measures 1,340 × 1,400 × 2,200 mm, and the 2.5 MW unit measures 1,400 × 1,950 × 2,100 mm and weighs about 3,000 kg, so floor loading, service clearance and connection orientation all need planning (both models connect from the top). An in-row CDU also needs a row-level secondary header that feeds each rack's liquid cooling manifold, which connects to the server cold plates through quick disconnects. Longer pipe runs and more fittings mean more pressure drop and more potential leak points.
Coolant Volume, Filtration and Water Quality
An in-rack loop holds little coolant, so filling or replacing fluid on any one unit is quick, but more units mean more filters and water quality checkpoints. An in-row loop holds much more coolant and centralizes filtration, which is more efficient. The flip side: if the secondary loop gets contaminated, say by debris from an incomplete flush during installation, every cold plate in the row is exposed.
APALTEK's in-row CDUs filter the secondary side to 25 μm, while the 150 kW in-rack unit filters to 100 μm. Match the filtration to what your cold plate channels can tolerate, especially if you use microchannel cold plates with very narrow passages. ASHRAE TC 9.9's white paper on water-cooled servers covers both CDU design considerations and water quality, and our direct-to-chip coolant guide covers fluid selection and corrosion control.
Serviceability and Day-2 Operations
In-rack operations are distributed but low-risk per task: more units to inspect and more filters to change, but each task touches one rack, and hot-swappable pumps let much of the work happen while the rack stays live. In-row operations are centralized but high-stakes per task: fewer units, but every service window requires confirming that redundant units have picked up the load.
Confirm monitoring integration early, too. APALTEK's 150 kW in-rack CDU offers RS485 (Modbus RTU) and RJ45 interfaces with a 3.5-inch color touchscreen, while the in-row CDUs support Modbus, TCP/IP and SNMP, with Redfish and a web interface available as options. If the CDU has to report into an existing DCIM or building management system, make sure the protocols, point lists and alarm logic line up.
Scaling and Phased Rollouts
In-rack CDUs fit phased rollouts. Add a liquid-cooled rack, add a CDU. Capital spending grows with compute, and you don't have to build out piping for a full row on day one. That makes in-rack a good option for piloting a few liquid-cooled racks in an air-cooled hall or for retrofitting an air-cooled data center for liquid cooling in stages.
In-row CDUs work best when you plan the full build up front. Piping, floor space and facility water connections have to be sized for the row's final capacity in phase one, so early utilization may be low. After that, adding racks is mostly a matter of connecting manifolds. For new high-density AI halls, that's usually the simpler path.
Cost Structure: CapEx, OpEx and Stranded Capacity
Comparing unit prices alone will mislead you. Weigh these instead:
- Equipment cost per kW: In-row units carry more capacity each, so their cost per kW is usually lower.
- Installation and piping: In-row needs row headers, floor space and more installation labor; in-rack needs far less.
- Stranded capacity: N+1 leaves one unit idle in normal operation, and 2N leaves half of installed capacity on standby. With in-row N+1, the idle unit is an entire large CDU.
- Rack space: Every U an in-rack CDU takes is a U that can't hold a server.
The usual pattern: with fewer racks and phased deployment, in-rack has the lower total cost. With many racks, high density and a single build-out, in-row comes out ahead. Where the crossover falls depends on your rack count, power per rack and redundancy target.
Decision Matrix: Which CDU Fits Your Deployment?
| Deployment scenario | Recommended approach | Why |
|---|---|---|
| Edge site or a single GPU rack | In-rack | Small failure domain, no row piping |
| Piloting a few liquid-cooled racks in an air-cooled hall | In-rack | Deploys rack by rack without changing the room layout |
| Phased liquid cooling retrofit | In-rack, moving to in-row as scale grows | Capital spending tracks compute |
| Multi-tenant colocation racks | In-rack | Keeps each tenant's cooling loop isolated |
| New high-density AI training cluster | In-row with N+1 | Lower cost per kW, centralized management |
| Mostly standard loads with a few ultra-dense racks | Hybrid | In-row covers the bulk; in-rack handles the outliers |
In a hybrid deployment, plan the control logic and secondary loops for each CDU type separately to avoid load imbalance or interference between them.
What to Put in Your CDU Spec
Whichever type you choose, spelling out the following in your RFQ cuts down on back-and-forth and puts every vendor's quote on the same baseline:
- Heat load: Liquid heat load per rack, rack count and headroom for the next hardware generation
- Rating conditions: Approach temperature, facility water supply temperature, load point and whether the filter is installed
- Flow and pump head: Secondary design flow and total pressure drop on the worst-case circuit (farthest rack, hoses, manifolds, quick disconnects and a dirty filter)
- Redundancy levels: Internal redundancy for pumps, power and sensors; number of redundant units; loop and facility-side redundancy
- Coolant: PG25, deionized water or another fluid, plus compatibility with all wetted materials
- Installation: For in-rack, U height, depth and rack standard; for in-row, footprint, floor loading and connection location
- Monitoring and communications: Protocols, point lists, alarm logic and any DCIM integration
- Leak detection and condensation control: Sensor locations and the actions they trigger
For the flow, temperature rise and pressure drop calculations, see the CDU sizing steps in our data center liquid cooling design guide. If you're still weighing the overall cooling architecture, see how direct-to-chip liquid cooling systems are put together.
How APALTEK Builds Both CDU Types
APALTEK is a liquid cooling manufacturer that designs and builds both in-rack and in-row CDUs in-house, and both are in production. We customize capacity and configuration to each project's rack power, rack count and interface requirements.
Because we also build liquid cold plates, we can deliver the CDU, cold plates and manifolds with pre-installed quick disconnects as one complete loop, leak tested as an assembly to reduce on-site integration risk. Every APALTEK product is helium leak tested before it ships. See our testing and validation capabilities for more, or browse our full range of AI data center thermal management products.
FAQ
How many racks can one in-row CDU support?
Divide the CDU's rated capacity by the liquid heat load per rack. At 130 kW per rack, a 1.5 MW in-row CDU supports about 11 racks at rated conditions. For N+1 redundancy, add one more unit as a standby.
Can in-rack and in-row CDUs be used in the same data hall?
Yes, and it's a common approach: in-row CDUs cover most of the racks, while in-rack units handle a few ultra-dense racks or racks added later. Plan the two sets of secondary loops and controls separately.
Which is better for an AI data center, an in-row CDU or an in-rack CDU?
For a new AI training cluster with many high-density racks in a row, an in-row CDU with N+1 redundancy is usually the better fit because it costs less per kW and centralizes management. For a few AI racks added to an existing hall, or a rollout spread over several phases, in-rack CDUs are usually the better fit because each rack is deployed and protected on its own.
When should you switch from in-rack to in-row CDUs?
Consider switching when most racks in a row would need two in-rack units for capacity or redundancy, when a single row grows past roughly ten liquid-cooled racks, or when you're building a new hall where piping and floor space can be planned from the start.
Should a CDU use a plate or a shell-and-tube heat exchanger?
Most data center CDUs use plate heat exchangers because they're compact, transfer heat efficiently and can achieve a tight approach temperature. Shell-and-tube exchangers tolerate fouling and higher pressures better, so they're more common on the facility side where water quality is poor or pressures are higher.
Start with Your Rack List, Then Pick the CDU Type
Neither CDU type is better across the board. In-rack keeps the failure domain to a single rack and suits rack-by-rack deployment and phased growth. In-row serves more racks with fewer units and suits dense, large-scale new builds. The right answer comes from your power per rack, rack count and layout, facility water conditions and redundancy target.
Send your rack list to APALTEK's engineering team, including power per rack, rack count and layout, facility water temperature and redundancy target. We'll recommend an in-rack, in-row or hybrid approach and quote the CDUs together with the matching cold plate and manifold loop, with a reply within one business day. Request a quote