Rack-Mount vs Cabinet CDU: Matching the Form Factor to the Deployment

Two photographs of two coolant distribution units. In the first, a cabinet-style CDU stands at the end of a row of liquid-cooled racks in a finished white space, expansion vessels at the top, stainless headers and pump skid below, facility piping coming up through a glass access floor. In the second, a rack-mount CDU sits on a test bench with its cover removed — pumps, motorized valves, sensors and wiring harness exposed — while a laptop logs its functional test. Same job. Completely different buildings around them.

That gap — not the hardware itself but everything the hardware forces you to build around it — is what this article is about. Choosing between a rack-mounted CDU and a cabinet (floor-mounted) CDU is not a capacity comparison; it is a decision about who owns the piping, where the redundancy lives, and how your room grows. We will define both form factors precisely, put their parameters side by side, walk through the deployments where each one is the right answer, and cross-check the flow arithmetic that decides the pipe between them.

What the CDU does — regardless of shape

Every liquid-to-liquid CDU, from a 5U in-rack chassis to a 2.5 MW floor-standing skid, performs the same set of functions. ASHRAE's liquid cooling guidance defines it as the assembly that separates the technology cooling system (TCS) from the facility water system (FWS) while transferring heat between them, and lists its working set: move heat from TCS to FWS, circulate TCS coolant, permit a coolant other than water, prevent condensation, hold a coolant chemistry different from the FWS, and provide flexible coolant temperature control [1][2]. The Open Compute Project adds the operational framing: pumps, filtration, sensors, valves and controls make the CDU the system node of the secondary loop, not just a heat exchanger [3][4].

Because the two loops are hydraulically separated, everything on the IT side — coolant chemistry, pressure class, dew-point margin — is owned and protected by whatever box you choose. The form factor does not change that job description. What it changes is the scale at which the job is done and the infrastructure you must build to feed it.

The two form factors, precisely defined

OCP describes an in-rack CDU as a standard rack-mount chassis — typically 4U or 5U — connected directly to the rack manifold and serving the IT equipment in that one rack [5]. Cabinet-style (floor-mounted) CDUs occupy roughly the footprint of a single rack, stand beside or between rows or in a mechanical room, and serve a group of racks through distribution piping. Schneider Electric draws the practical capacity line between them: rack-mounted units typically handle loads of 100 to 300 kW, while centralized floor-mounted units are the better choice above roughly 300 kW [6]. Rittal's current portfolio brackets the same reality from the other side: in-rack units around 150 kW per unit in a compact form, in-row/cabinet units that scale past 1 MW in 250 kW increments [7].

Side by side: the parameters that actually differ

One number in that table deserves its own sentence: the 5U of rack space a rack-mount CDU consumes is not just 5U — in a 42U cabinet it is roughly 12% of the enclosure you paid for, occupied by equipment that will never compute. At 70 kW per cabinet that may be an excellent trade. At 130 kW per cabinet it starts to hurt.

Where the rack-mount CDU wins

1. Pilots, labs and proof-of-concept rooms. A liquid-cooling pilot is a question, not a building. A 4U–5U in-rack unit turns one high-density rack into a complete, isolated liquid loop without a single meter of facility piping, which is exactly why OCP's integration guidance treats in-rack units as a primary distribution method for individual racks [5]. When the experiment ends, the loop leaves with the rack. The test-bench photo above is the prelude to this story: rack-mount units are validated as complete assemblies precisely because they are bought one at a time, by people who cannot afford a dud.

2. The single high-density rack in an air-cooled room. One AI rack at 80–120 kW inside an otherwise air-cooled space is the classic retrofit problem. An in-rack CDU brings cooling directly to the heat source with rack-level thermal control, and Schneider explicitly positions it for standalone high-density racks and incremental scaling as demand grows [6]. You do not build a TCS main for one rack.

3. Edge and constrained sites. Edge rooms rarely have a mechanical room, a raised floor, or staff on site. A rack-mount CDU ships inside the IT enclosure, connects to one facility-water pair, and its short secondary runs — CDU to manifold to cold plate, all within one cabinet — minimize the piping that can leak, foul, or be installed wrong [6]. Rack-level isolation also means a coolant problem in the liquid-cooled cabinet cannot propagate into anything else, because there is nothing else on the loop.

4. Phased buildouts where capital follows demand. If liquid cooling arrives rack by rack over quarters, rack-mount units let capacity track load exactly. Each new high-density rack brings its own CDU, its own redundancy domain, its own billing unit. Schneider's framing of in-rack units as a way to "scale efficiently while aligning investment to demand" is precisely this pattern [6]. The economics invert at scale — there are only so many in-rack units you can deploy before the cabinet option becomes cheaper [6] — but below roughly four liquid-cooled racks, per-rack infrastructure usually wins.

5. Wherever independent control per rack is a requirement. HPC labs running mixed job types, university clusters with one GPU rack and one CPU rack, test houses validating customer hardware: rack-level CDUs give each enclosure its own flow setpoint, its own coolant chemistry, its own maintenance window. One rack can be drained while its neighbor runs.

Where the cabinet-style CDU wins

1. Rows of liquid-cooled racks — the AI production default. NVIDIA's NVL-class platforms typically deploy four to eight liquid-cooled racks per pod, and Schneider's guidance is unambiguous: a group of racks means a floor-mounted CDU [6]. The unit in Fig. 2 is this answer made physical — one enclosure, standing at the end of the row, feeding the whole group through a row manifold while facility water rises through the access floor behind it.

2. Loads above 300 kW. Centralized units are the better choice above roughly 300 kW, and the modern capacity ladder runs 600 kW, 1–1.3 MW, and 2.5 MW per unit at the top — Schneider's Motivair line alone spans 105 kW to 2.5 MW per unit [6][8]. OCP's Deschutes CDU specification pushes the ceiling further, targeting roughly 2 MW heat rejection at 500 GPM for hyperscale AI [10].

3. Concurrent maintainability. OCP's CDU guidance is explicit that service and component replacement must be possible without interrupting IT cooling — hot-swappable pumps, bypass provisions, unit-level N+1 with seamless changeover [3]. That architecture is realistic in a cabinet unit standing in a service aisle with room to work, and much less realistic in a 5U chassis sandwiched between servers inside an occupied rack. ASHRAE’s 2024 resiliency guidance for cold-plate deployments makes the same point from the operations side: CDU redundancy, leak response and coolant-quality management through hardware refresh cycles are now part of the availability design, not afterthoughts [11]. If your availability target assumes live maintenance, the cabinet form factor is doing part of the engineering for you.

4. When the CDU belongs in a mechanical room. Some operators want the white space entirely free of mechanical plant: CDUs behind a wall, in a separate plant area, or in a dedicated zone with its own containment. Schneider describes exactly this spectrum — end-of-row for efficiency, walled-off mechanical rooms to save data hall space, or a combination for N+1 across zones [6]. A cabinet CDU can live anywhere the FWS reaches; a rack CDU by definition lives with the servers.

5. Fewer things, better watched. One 1 MW cabinet unit replaces six 150 kW rack units — six pumps to service become one pump skid, six controllers become one BMS-integrated panel, six coolant inventories become one managed loop. OCP's chemistry-control guidance (integrated filtration, air and gas removal, continuous monitoring) is far easier to implement once, centrally, than six times, locally [3].

The flow arithmetic that seals the decision

Form-factor comparisons go soft in the middle until you attach numbers, and the number that matters most is flow — because flow sets the pipe, and the pipe sets the build. Using the OCP rating basis of 1.5 L/min per kW [9] (which, as we showed in the pipe-sizing article, corresponds to a ~10 °C secondary ΔT on water or a PG25 coolant):

Read that ladder as an architecture statement. At 70–150 kW the secondary loop is hose-sized — something that fits through a rack's cable egress and needs no pipefitting crew. At 1 MW you are routing a five-inch pressurized main across a room, with slope requirements, supports, pressure tests and expansion compensation. That is not a component choice anymore; it is a construction project. The rack-mount CDU exists precisely so that small loops never have to become construction projects, and the cabinet CDU exists so that big loops become one well-managed project instead of forty small ones.

There is a second, quieter reason the large unit wins at scale: the FWS side. A rack-mount CDU still needs facility water delivered to the rack — the loops are separated, but the heat has to leave somehow. Feeding twelve racks with twelve in-rack units means twelve independent FWS terminations, twelve points where building water pressure, chemistry and leak risk touch the IT space. One cabinet unit concentrates that interface at a single, monitorable, containable location — which is exactly the containment-boundary role ASHRAE’s guidance assigns to the CDU, and the reason automated leak detection and coolant-quality monitoring converge on it as the anchor point [12].

The decision, in one table

Most production rooms end up running both

The mature answer is rarely either/or. A typical AI buildout grows like this: a rack-mount CDU proves the loop in a lab; the first production pod gets one cabinet unit at the end of the row; a second pod adds another cabinet unit with unit-level N+1 between them; meanwhile a test rack with its own in-rack CDU validates every new hardware generation before it touches the production loop. OCP's modular-TCS guidance is written for exactly this life cycle — CDUs right-sized for peak and partial load, capacity modules added as the row fills, serviceability designed in from the start [3][7].

If you take one rule from this article, take the crossover rule: price the infrastructure, not the boxes. The rack CDU looks cheaper until you count twelve facility terminations, twelve controllers and 60U of lost IT space. The cabinet CDU looks expensive until you notice it is one commissioning project, one coolant inventory and one maintenance contract. The crossover typically lands between three and six liquid-cooled racks — compute it for your own room with your own numbers, and the architecture decision makes itself.

Procurement checklist — eight questions before you order either one

1. What is the aggregated kW of the racks the unit must serve, at the worst-case hardware generation you will deploy — not today's?
2. What flow does that require at your design ΔT, and does the unit's rated flow at its test conditions meet it? (1.5 L/min/kW is the OCP rating basis [9].)
3. Does the available pump head cover your entire TCS loop drop — manifold, quick disconnects, cold plates — at design flow? (OCP rates rack-class units at ≥ 25 psi and row-class at ≥ 40 psi [9].)
4. Where does redundancy live — internal N+1 pumps, dual power feeds, or a second unit? Does it meet your concurrent-maintainability expectation [3][11]?
5. What is the dew-point strategy? The CDU must hold TCS supply at least ~2 °C above room dew point to prevent condensation [2] — ask how the control loop does it and how it fails.
6. What chemistry does the unit enforce — filtration rating, air/gas removal, monitoring points — and is the wetted-material list compatible with your planned coolant [3][4][12]?
7. What does service look like at 3 a.m. — hot-swap from the front aisle, or unplug half a rack?
8. What is the growth module? If the room doubles next year, does the same product family scale with it [7]?

Both form factors do the same job; the difference is the shape of the building you are willing to build around them. If you are weighing a specific deployment — rack counts, kW targets, facility water constraints — send us the configuration and we will quote the loop that fits it. For the numbers behind the pipes referenced above, read how to size coolant pipe diameter from CDU operating conditions, and for the condensation constraint every CDU controller enforces, see how to calculate dew point in a data center.

References

1. ASHRAE Technical Committee 9.9, Liquid Cooling Guidelines for Datacom Equipment Centers (Datacom Series, Book 4) — definition of the CDU as the demarcation between the Facility Water System and the Technology Cooling System, its functional set, and TCS/FWS fluid-quality requirements. ashrae.org
2. Harris Thermal, “AI data center liquid cooling: where heat exchangers fit in the cooling infrastructure” — summary of the ASHRAE white paper’s six CDU functions, the requirement that the CDU monitor ambient dew point and hold TCS supply at least 2 °C (3.6 °F) above it, and the W-class/S-class temperature frameworks governing the two loops. harristhermal.com
3. Open Compute Project, OCP CDU Performance & Specification Guidance for Modular TCS — CDU role, chemistry control (integrated filtration, air/gas removal, real-time monitoring), flow control and redundancy strategies, right-sizing for peak and partial load, concurrent-maintainability requirements, and SLA frameworks. ocp-all.groups.io (PDF)
4. Open Compute Project, Liquid Cooling Integration and Logistics White Paper, Rev. 1.0 — CDU definition, TCS/FWS separation with no coolant flow between loops, in-rack and end-of-row CDU distribution methods, and integration/commissioning practices from nVent, Vertiv, CoolIT, Meta and other contributors. vertiv.com (PDF)
5. Cooltechx, “Rack-Mounted vs In-Row vs Room-Level CDU: How to Choose the Right CDU Architecture for AI Data Centers” — OCP’s description of the in-rack CDU as a standard 4U–5U rack chassis serving its own rack, the in-row CDU as a larger unit serving an array of racks, and the scenarios each fits. cooltechx.com
6. Schneider Electric Blog, “In-rack CDU v. floor-mounted CDU: Which option is better for your data center?” — in-rack units typically handling 100–300 kW and floor-mounted units the better choice above 300 kW; NVIDIA NVL-platform deployments of four to eight racks pointing to floor-mounted; placement options including end-of-row, mechanical room, and combined N+1; Motivair CDU capacity spanning 105 kW to 2.5 MW. blog.se.com
7. Rittal, “How to Plan AI-Ready Cooling for High-Density Racks” — in-rack CDUs in a compact 4OU form up to ~150 kW per unit, in-row CDUs growing beyond 1 MW via 250 kW incremental N+1 additions, and CDU placement as the first decision in a DLC deployment. rittal.com
8. Data Centre Magazine, “Schneider Electric Reveals New Liquid Cooling Portfolio” — Motivair CDU family scalable from 105 kW to 2.5 MW per unit, certified for NVIDIA’s latest hardware; rack densities surpassing 140 kW with projections toward 1 MW and beyond. datacentremagazine.com
9. OCP Modular TCS rating basis for liquid-to-liquid CDUs — base flow of 1.5 L/min·kW, rack-class (in-rack) pump head ≥ 25 psi, row-class (in-row) pump head ≥ 40 psi, as compiled in industry summaries of OCP qualification requirements. ocp-all.groups.io (PDF)
10. HPCwire / AIwire, “OCP Launches New ‘Open Data Center for AI’ Strategic Initiative” — the Deschutes CDU specification targeting ~2 MW heat loads at 500 GPM and 80–90 psi, assembled from multi-vendor components for hyperscale AI. hpcwire.com
11. ASHRAE Technical Committee 9.9, Liquid Cooling: Resiliency Guidance for Cold Plate Deployments (Technical Bulletin, 2024) — high-availability operation of liquid-cooled IT, CDU redundancy, leak response and coolant-quality management during hardware refresh cycles. tpc.ashrae.org
12. Dober, “Data Center Leak Detection and Glycol Monitoring” — ASHRAE TC 9.9’s structural recommendation to deploy a CDU as both thermal interface and containment boundary, OCP wetted-material and monitoring guidance, and the convergence on coolant-quality monitoring and automated leak detection as baseline requirements. dober.com