What a CDU actually is
A coolant distribution unit (CDU) is the interface skid between an IT-side liquid-cooling loop and a facility-side water loop. It contains pumps, a plate heat exchanger or brazed heat exchanger, filtration, makeup/expansion tank, pressure and temperature sensors, flow meters, leak detection, and controls. The CDU does not reject heat to atmosphere; it transfers heat from the warm IT coolant to the facility water loop, which then rejects heat through a cooling tower, dry cooler, or chiller. Company product documentation describes CDUs as the “heart” of liquid-cooled data-center racks, but that is vendor self-reported language and should be treated as Tier 2 marketing context.
Source: Vertiv coolant distribution unit product page — Vertiv, self-reported, accessed 2026-05-07.
Common CDU topologies:
- Facility-water-to-technology-cooling-system (FW-to-TCS) CDU: primary facility water enters a plate heat exchanger; secondary treated coolant circulates to cold plates.
- Direct-to-chip CDU loop: coolant is pumped to server cold plates at inlet temperatures often around 30–45°C, depending on facility water temperature and control strategy.
- Rear-door heat exchanger loops: some CDU-like devices serve rear-door heat exchangers instead of direct cold plates.
- Immersion CDUs: some vendors offer CDUs for single-phase immersion tanks, where dielectric oil or synthetic fluid is circulated to an external heat exchanger.
Source: Open Compute Project Cooling Environments / Liquid Cooling wiki — industry community specification page, updated continuously, accessed 2026-05-07. OCP documentation is standards-body-adjacent but not a formal IEEE/ISO/NIST standard.
The CDU is often called a “liquid-to-liquid” device because the primary technical function is hydraulic separation and heat exchange between two liquid loops. It is not equivalent to a chiller. A chiller provides mechanical cooling and changes the facility water temperature; a CDU only moves heat between two loops and maintains secondary-loop pressure, flow, and water quality.
Source: Schneider Electric, White Paper 265, “Liquid Cooling for Data Centers” — vendor/white paper, needs primary source validation for exact document landing page; accessed 2026-05-07.
Technical deep-dive: loops, components, and failure modes
2.1 Hydraulic separation and heat exchange
The CDU isolates the secondary IT loop from the facility water loop. This is operationally important because facility water may be chemically untreated or variable; secondary-loop coolant is typically filtered, inhibited deionized water, propylene-glycol/water mix, or a dielectric fluid. The plate heat exchanger allows high thermal effectiveness at a relatively small volume, but it also adds pressure drop on both loops. CDU pumps must be selected to overcome:
- secondary-loop piping resistance,
- cold-plate pressure drop,
- quick-disconnect fittings,
- heat exchanger secondary side pressure drop,
- valves and strainers,
- elevation head if the CDU is above or below the rack manifold.
This is ordinary hydraulic design; the governing relation is standard pump-system energy balance. The specific pressure drops are vendor- and installation-specific; generic published curves are not a substitute for measured system curves.
2.2 Key components
- Pumps: often N+1 redundant variable-speed centrifugal pumps in a skid.
- Heat exchanger: plate-and-frame or brazed plate, usually stainless steel or copper-brazed depending on coolant compatibility.
- Filtration and air separation: micro-bubble separators and particulate filters prevent cold plate fouling.
- Makeup/expansion: maintains pressure and accommodates thermal expansion.
- Sensors and controls: flow, temperature, pressure, conductivity, leak rope/tape, and optional BACnet/Modbus interfaces.
- Power and controls enclosure: VFD drives, control logic, and alarm outputs.
Source: Vertiv coolant distribution unit product page — self-reported component overview, accessed 2026-05-07.
2.3 Failure modes
Liquid cooling changes the risk profile of a data hall. The most commonly cited risks in vendor docs and industry guides are:
- Leakage at quick disconnects and cold plates, causing IT hardware damage or corrosion.
- Filtration failure leading to cold plate microchannel clogging and rising GPU/CPU junction temperatures.
- Pump failure causing immediate loss of coolant flow; redundancy mitigates but does not eliminate systemic failure.
- Heat exchanger fouling on the facility-water side, degrading thermal performance over time.
- Control sensor drift causing supplied coolant temperature to fall below the facility room dew point, creating condensation on fittings and manifolds.
- Water-quality incompatibility between facility and secondary fluids causing corrosion or biological growth.
Source: ASHRAE Technical Committee 9.9 Thermal Guidelines for Data Processing Environments — standards-body document, accessed 2026-05-07. The exact edition and page numbers require primary-source validation; ASHRAE publishes liquid-cooling guidance under TC 9.9.
Worked calculation: CDU flow and pump power at 100 kW rack load
This calculation estimates the secondary-loop flow and CDU pump electrical overhead for a 100 kW rack. Inputs are a mix of assumed design points and standard fluid properties.
Formula:
\[
Q_v = \frac{P}{\rho c_p \Delta T}
\]
where:
- \(P\) = rack heat load absorbed by the coolant (W)
- \(Q_v\) = volumetric coolant flow (m³/s)
- \(\rho\) = coolant density (kg/m³)
- \(c_p\) = coolant specific heat (J/(kg·K))
- \(\Delta T\) = supply-to-return coolant temperature rise (K)
Calculation at 100 kW, ΔT = 10 K, water properties:
\[
Q_v = \frac{100{,}000}{998 \times 4180 \times 10} = 0.00239\ \text{m}^3/\text{s}
\]
\[
Q_v = 2.39\ \text{L/s} \approx 143\ \text{L/min}
\]
Pump electrical power at 345 kPa pressure drop, 70% efficiency:
\[
P_{\text{pump}} = \frac{\Delta p \times Q_v}{\eta}
\]
\[
P_{\text{pump}} = \frac{345{,}000 \times 0.00239}{0.70} \approx 1{,}178\ \text{W}
\]
Pump overhead relative to IT load:
\[
\frac{1.18\ \text{kW}}{100\ \text{kW}} \approx 1.18\%
\]
Sensitivity:
- If \( \Delta T = 5\ \text{K} \), flow doubles to 4.78 L/s; at the same pressure drop and efficiency, pump power doubles to approximately 2.36 kW, or 2.36% of IT load.
- If \( \Delta T = 15\ \text{K} \), flow falls to 1.59 L/s; pump power falls to approximately 0.79 kW, or 0.79% of IT load.
- If pressure drop is 100 kPa instead of 345 kPa, pump power at ΔT = 10 K falls to approximately 0.34 kW, or 0.34% of IT load.
Interpretation: CDU pump overhead is typically low single-digit percentage of rack IT load, but the result is highly sensitive to the design supply/return temperature spread and the installed pressure drop. This calculation is generic engineering and does not replace a vendor-specific system curve.
Parameter table
| Parameter | Value | Source (URL + ID + date) | Tier | Measured/derived/assumed/inferred | Cross-verified |
|---|---|---|---|---|---|
| Rack heat load | 100 kW | Assumed design point; NVIDIA GB200 NVL72 page cites “up to 120 kW per rack” as vendor self-reported marketing, accessed 2026-05-07 | Tier 2 / assumed | Assumed | No |
| Water density | 998 kg/m³ at 25°C | NIST Chemistry WebBook fluid properties — NIST, accessed 2026-05-07 | Tier 1 | Derived from standard tables | Yes |
| Water specific heat | 4.18 kJ/(kg·K) | NIST Chemistry WebBook fluid properties — NIST, accessed 2026-05-07 | Tier 1 | Derived | Yes |
| Supply/return ΔT | 10 K | Common direct-to-chip design point; vendor-specific confirmation needed for a given CDU | Tier 2 / assumed | Assumed | No |
| Secondary flow at ΔT=10 K | 2.39 L/s = 143 L/min | Calculation above | Derived | Derived | Yes |
| Pump head | 345 kPa | Assumed representative secondary-loop pressure drop; needs primary source validation for installed system | Assumed | Assumed | No |
| Pump efficiency | 0.70 | Assumed typical centrifugal pump efficiency | Assumed | Assumed | No |
| CDU pump electrical power | ≈1.18 kW | Calculation above | Derived | Derived | No |
| Pump overhead vs 100 kW IT load | ≈1.18% | Calculation above | Derived | Derived | No |
Intellectual property, patents, and research evidence
4.1 Tier-1 patent gap
No granted patent is cited in this note because no grant number was verified against USPTO/EPO/WIPO during this pass. A patent is evidence of recorded claims, not proof of performance or commercial viability. Therefore this note does not assert a moat based on a patent. The correct USPTO CPC search class for liquid-cooling enclosures/racks is H05K7/20272; a second pass should retrieve a granted patent from that class and cite the grant number, title, assignee, and grant date.
Source: USPTO Patent Public Search — search portal, accessed 2026-05-07.
4.2 Peer-reviewed literature
A candidate review paper is:
Zhang H., Shao S., Xu H., Zou H., Tian C., “A review of data center cooling technology, operating conditions and the corresponding low-grade waste heat recovery opportunities,” Renewable and Sustainable Energy Reviews, vol. 31, 2014, pp. 584–593.
Candidate DOI: 10.1016/j.rser.2013.12.007 — needs primary source validation. I could not resolve this DOI in the current environment, so I am treating it as a literature pointer rather than a load-bearing Tier-1 source. It would support general context on liquid cooling and waste-heat recovery, but it is not a CDU-specific experimental study.
Because the required granted patent was not verified and the paper DOI is unresolved, no Tier-1 technical performance claim in this note should be treated as patent- or paper-backed.
Adoption and deployment evidence; marketing vs SEC filings
5.1 Vendor technical documentation
NVIDIA states that direct-to-chip liquid cooling is used for high-density GPU-accelerated systems and that it reduces data-center PUE and allows greater compute density. This is Tier 2 vendor self-reported marketing context.
Source: NVIDIA Data Center Liquid Cooling page — NVIDIA, self-reported, accessed 2026-05-07.
NVIDIA’s GB200 NVL72 product page markets “up to 120 kW per rack” for the liquid-cooled GB200 NVL72. That is vendor marketing, not a standards-body measurement; on a 100 kW rack assumption, the flow calculation above remains illustrative.
Source: NVIDIA GB200 NVL72 product page — NVIDIA, self-reported, accessed 2026-05-07.
Vertiv, Modine, nVent, Supermicro, Dell, and HPE all publish liquid-cooling and CDU-related product pages. These are useful for component existence and architecture but not for independently verified performance.
- Vertiv coolant distribution units — self-reported.
- Modine liquid cooling — self-reported; includes CDUs and heat exchangers.
- nVent liquid cooling — self-reported.
- Supermicro liquid cooling solutions — self-reported.
- Dell liquid cooling — self-reported.
- HPE liquid cooling — self-reported.
5.2 SEC filing cross-check
SEC filings are legal documents and should be used to constrain marketing claims.
- NVIDIA 10-K: NVIDIA’s risk factors discuss product complexity, manufacturing capacity, and supply-chain dependencies; they do not disclose a CDU revenue line or liquid-cooling failure-rate data. Marketing pages describe capabilities, but the filing does not quantify liquid-cooling margins or attach rates. This is an information gap rather than a contradiction.
Source: SEC EDGAR, NVIDIA 10-K search, CIK 0001045810 — SEC, accessed 2026-05-07.
- Vertiv 10-K: Vertiv discloses thermal-management products and risk factors related to supply-chain constraints, customer demand, and component availability; it does not separately break out CDU gross margin on the primary financial statements. CDU gross margin therefore requires segment-level or earnings-call validation.
Source: SEC EDGAR, Vertiv 10-K search, CIK 0001674101 — SEC, accessed 2026-05-07.
- Super Micro Computer 10-K: Supermicro discusses liquid-cooled rack solutions in its business description but does not quantify liquid-cooling revenue separately in the 10-K headline financials. Its risk factors include component availability and rapid technology transitions.
Source: SEC EDGAR, Supermicro 10-K search, CIK 0001375365 — SEC, accessed 2026-05-07.
Discrepancy statement: Marketing material describes CDUs as enabling high-density AI deployments; SEC filings describe liquid cooling as one of many product categories and emphasize supply and demand risk rather than guaranteed adoption. No contradictory dollar figure was identified in this pass. The principal discrepancy is scope: marketing is forward-looking and capability-focused; 10-K disclosures are backward-looking and risk-focused.
Contradictions and open questions
- kW-per-rack claims are not standardized. NVIDIA markets “up to 120 kW per rack” for GB200 NVL72, but facility water temperature, altitude, coolant chemistry, and CDU pressure drop change the actual sustained cooling capacity. Marketing does not provide the full operating envelope.
Source: NVIDIA GB200 NVL72 product page — self-reported, accessed 2026-05-07.
- “Liquid cooling reduces PUE” is system-dependent. Liquid cooling can reduce fan power and improve heat-rejection temperature, but actual PUE depends on the heat-rejection plant, not the CDU alone. A low-flow CDU installed in a facility with an inefficient chiller plant will not automatically achieve the marketed PUE.
Source: ASHRAE TC 9.9 Thermal Guidelines — standards-body guidance, accessed 2026-05-07.
- Reliability claims are often qualitative. Vendors mention N+1 pumps and leak detection, but published MTBF/leak-rate statistics for CDUs are not consistently disclosed in the product pages reviewed. This is a data gap, not a demonstrable contradiction.
Source: Vertiv CDU product page — self-reported, accessed 2026-05-07.
- Standards remain fragmented. OCP, ASHRAE TC 9.9, and individual hyperscale buyers have different manifold sizes, coolant types, and connector standards. A CDU built for one liquid-cooled rack ecosystem may not be compatible with another.
Source: Open Compute Project Liquid Cooling wiki — community specification page, accessed 2026-05-07.
Unit economics / business model
Who monetizes: CDU makers sell hardware skids as capital equipment, often with commissioning, startup, and service/warranty contracts. Integrators include CDUs in full liquid-cooled rack or row deployments. Facility operators monetize the resulting power-density increase by charging higher per-kW colocation or cloud compute rates.
How value is captured:
- CDU vendor: one-time skid sale plus service, replacement filters, and controls upgrades.
- Server/rack integrator: CDU value is embedded in the price of a liquid-cooled server/rack.
- Data-center operator: captures value through higher power density per square foot and potentially lower annualized fan energy.
- Chip vendor: captures value indirectly by enabling higher-power accelerators that would be difficult to air-cool.
Margins: CDU-specific gross margin is not standardized in public filings. Vertiv’s audited consolidated gross margin is available in its 10-K, but the exact figure should be re-read from the filing; I am not quoting an unverified number here. Vertiv’s 10-K search URL is provided above; the exact gross margin for FY2024/FY2025 is needs primary source validation.
Source: SEC EDGAR, Vertiv 10-K search — SEC, accessed 2026-05-07.
Illustrative CDU pump energy cost at 100 kW rack:
Using the calculation above, pump power is approximately 1.18 kW at ΔT = 10 K. At a hypothetical electricity cost of $0.10/kWh and 8,760 hours/year:
\[
1.18\ \text{kW} \times 8{,}760\ \text{h} \times \$0.10/\text{kWh} \approx \$1{,}034/\text{year}
\]
At ΔT = 5 K, pump power approximately doubles to 2.36 kW, and annual pump energy cost approximately doubles to $2,068/year. The result is small relative to the cost of 100 kW of GPU hardware but material in large fleets.
Hardware pricing: CDU list prices are not standardized and are often project-quoted. Public prices are rarely disclosed by major vendors. Any specific $30k–$100k per CDU figure would be needs primary source validation.
Disconfirming Evidence
The CDU thesis would be wrong if:
- High-density liquid cooling stalls. If GPU/ASIC power per rack stops rising or plateaus below levels that justify direct-to-chip liquid cooling, air cooling remains sufficient and CDU attach rates fall. Evidence would be a sustained plateau in announced rack power below current NVIDIA GB200 marketing levels.
Source: NVIDIA GB200 NVL72 page — self-reported design point, accessed 2026-05-07.
- Leak and reliability problems trigger costly warranty claims. If CDU or cold-plate leaks cause field failures at scale, hyperscale buyers could delay deployment or impose expensive containment/redundancy requirements. Evidence would be vendor recall or 10-K liability disclosure.
Source: SEC EDGAR, Vertiv 10-K search — risk factors, accessed 2026-05-07.
- Cooling architecture shifts away from CDU-mediated direct-to-chip loops. If immersion cooling or chip-level two-phase cooling becomes the dominant high-density architecture, the stand-alone liquid-to-liquid CDU market could be commoditized or bypassed. Evidence would be multiple hyperscale announcements moving away from CDU/water loops.
Source: Open Compute Project Liquid Cooling wiki — architecture alternatives, accessed 2026-05-07.
- Buyers vertically integrate CDU functionality. If NVIDIA, Supermicro, or hyperscalers integrate pump/heat-exchanger modules directly into racks and eliminate the discrete CDU as a saleable skid, standalone CDU vendors lose value capture. Evidence would be major buyer procurement shifts to internally designed manifolds and skids.
Source: Supermicro liquid cooling — rack integration, accessed 2026-05-07.
Recent developments
_Checked 2026-10-26 by the technology material-change monitor._
- 2026-09-23 — Schneider Electric launches the Motivair WCDU (liquid + air in one CDU). Schneider claims it is the only CDU combining liquid and air cooling in a single unit, aimed at operators deploying CDUs outside the white space; with it, the Motivair line spans 105 kW to 3.5 MW. Unit begins shipping in select regions October 2026, U.S. pre-order early 2027. This raises the top of the vendor capacity curve (Vertiv's CoolChip CDU 2300, 2.3 MW, went on sale in EMEA in May 2026) and is a genuine product launch, not a roadmap slide. Source: se.com
- 2026-07-31 — nVent adds 160,000 sq ft of liquid-cooling manufacturing capacity. Third data-center liquid-cooling capacity expansion in three years (>400,000 sq ft added cumulatively), production expected H1 2027, >200 employees; nVent cites >2 GW of liquid cooling deployed to date. Capacity — not design — has been the binding constraint this cycle. Source: investors.nvent.com
Read-through: the competitive axis in CDUs is consolidating around megawatt-class capacity per unit and outdoor/technical-corridor placement, with the two largest Western thermal vendors both expanding within the same quarter. That is a share-shift risk for sub-1 MW row CDU specialists and a tailwind for the high-capacity incumbents already in this note's ticker set (VRT, NVT, JCI).