Choosing the Pressure Rating for Liquid Cooling Quick Disconnects

Ask three vendors for the working pressure of their liquid-cooling quick disconnect and you will get three numbers that are not comparable: 2.4 bar, 6.9 bar, 16 bar. All three can be correct, because they describe different interfaces, different fluids and different failure criteria — and the number a datasheet prints is written for a component, while the number a loop has to respect is written for the weakest point in the entire wetted path. Getting from one to the other is the selection problem, and it is arithmetic, not negotiation.

This reference builds the pressure side of a connector selection from the loop outward. It covers the three pressure numbers on every datasheet and which one you design to, the four layers of pressure a running loop actually sees, what each standard and vendor family allows, what a 1.5× proof test proves and what it structurally cannot see, how to size by flow before checking force, and the evidence to demand before you accept a part. For the mechanical and material side of the same decision, see our hot-swap quick disconnect buying guide and the UQD / blind-mate selection piece.

The three pressure numbers on every datasheet — and the one you design to

Every coupling datasheet carries at least three pressures, and they answer three different questions.

Maximum working pressure (MWP), or rated working pressure. The continuous pressure the part may see in service, with its rated service life and its rated temperature window, without losing the seal or the latch. This is your design number. It is not the pressure at which something breaks; it is the pressure at which the part is still doing its job as qualified.

Proof or test pressure. A verification pressure applied at the factory or in the field to demonstrate that the assembly holds. It is typically 1.5× the design pressure, it is not a service condition, and it is not a licence to run the part there.

Burst pressure. The threshold at which the pressure envelope fails — the housing yields, the seal extrudes, or the assembly separates. The Open Compute Project's UQD requirement sets a clean 3:1 relationship: 100 psi (6.89 bar) working against a 300 psi (20.68 bar) minimum burst [1][2][3].

The reason MWP is defined separately from burst is that pressure acts on a coupling in two ways at once. It loads the housing as hoop stress, and it pushes the two halves apart with an axial separating force proportional to the bore area. On small couplings the separating force is modest — a few tens of newtons — but it is the reason a partially latched coupling can pass a leak test at dead head and fail in service, and it is why the latch is part of the pressure rating rather than a separate feature. We come back to the force arithmetic in its own section.

What pressure the loop actually sees: four additive layers

Most badly specified loops were sized on one pressure. A running liquid-cooling loop stacks four, and the sum is what the connector has to survive:

1. Static head and cold fill pressure. Set by geometry and by the fill valve, not by the pump. Present even with the pump off.
2. Pump head at design flow. The sum of every pressure drop between the pump discharge and the point in question.
3. Thermal expansion rise. What the fluid does to the pressure when it warms from fill temperature to operating temperature, minus whatever the expansion tank absorbs.
4. Transients. Water hammer from a valve, a pump start or stop, or a check valve slam. This layer is the largest per unit time and the one most often omitted.

Layers one and two are straightforward and repeatable. Layer three is where an expansion tank either does its job or turns the relief valve into the expansion control. Layer four is where most MWP violations actually happen.

Layer 1 — static head and cold fill pressure

Static head is geometry: coolant density times gravity times the height difference. A 25 % propylene glycol solution at roughly 20 °C sits near 1,017 kg/m³, which gives about 0.0998 bar per metre of elevation — call it 0.10 bar/m in design conversation. For a pump on the slab and manifolds at the top of a two-storey hall, 15 m of rise is 1.50 bar of static head [5][6]. Eight metres is 0.80 bar.

Fill pressure is a choice, and the constraint on it is at the highest point of the loop, not at the pump. The fill pressure minus the static head at the top of the circuit is the residual pressure there, and that residual has to stay positive with margin, or the top of the loop runs at sub-atmospheric pressure, draws air in through the smallest seal imperfection, and starts losing the pump's net positive suction head.

Two consequences follow that people miss. First, the lowest pressure in the loop is at the high point, and the highest pressure with the pump running is at the pump discharge — not at the cold plate, not in the CDU heat exchanger, and not at the connector you were worried about. The QD pair on the topmost tray sees roughly fill plus static-minus-rise plus its own share of pump head, which is a benign number. The QD pair on the pump discharge header sees the whole stack. Second, raising fill pressure to fix a marginal high point raises the discharge pressure everywhere else, which is a trade against MWP rather than a free correction. If the fill has to go up, the pressure class of the pump-discharge hardware has to go up with it.

Layer 2 — pump head: build the budget, do not guess it

Pump head is the sum of every pressure drop on the path at design flow. Build it as a table, component by component, at the flow you intend to run. The following band is representative of a row-level CDU feeding direct-to-chip trays with PG25; the point is the method, not the numbers.

Three corrections decide whether this budget is real. Glycol at 25 % raises the pressure drop of the same circuit substantially against water at the same flow — published correction factors put PG25 near 2.2× the water ΔP for the same loop [7]. Fouling is a permanent upward drift, not a commissioning artefact, and it is why the margin belongs in the budget rather than in the pump's published curve. And the quick disconnect's own share is not negligible once the couplings are smaller than they should be: at its rated flow, one UQD04 pair (supply plus return) contributes about 22 kPa, roughly 9 % of a 250 kPa loop budget [3][8]. That looks acceptable — until the same flow goes through a UQD02 pair, where the contribution becomes about 268 kPa, or 107 % of the budget on its own.

Velocity sets all of this. Push the header velocity up to shrink pipe size, and the ΔP rises with the square of velocity while the surge in layer four rises linearly. Our coolant loop pipe sizing walkthrough works the same trade at the pipe level.

Layer 3 — thermal expansion, and what happens without a tank

A sealed loop filled cold and run hot has to put the fluid's thermal expansion somewhere. For PG25, the volume change from a cold fill at 20 °C to a 65 °C operating point is about 2.5 % of loop volume — roughly 0.056 % per kelvin, consistent with the fluid manufacturers' published volume expansion of about 5.1 % between −10 °C and 90 °C [6][9]. Water in the same span expands about 1.9 %. On a 2,000 L row-level loop, that 2.5 % is 50 litres the system must absorb.

An expansion tank absorbs it by compressing a gas cushion, and its usable share is governed by the acceptance factor:

Both pressures must be absolute. Sizing with gauge pressure is the single most common error in this calculation because it inflates the acceptance factor and makes the tank look smaller than it needs to be [10][11][12]. Note also that the required volume is driven by the ratio between fill and relief pressure, not by the loop size alone: the closer the fill pressure sits to the relief setting, the less room the gas has to compress and the larger the tank has to be. Tall risers need both a higher fill pressure and a larger tank.

Now remove the tank, or close the isolation valve in front of it, and see what the same 50 litres does to a rigid 2,000 L volume:

The gap between 9.9 bar per kelvin and 0.45 bar per kelvin is not a property of the coolant; it is a property of whether a compliant element exists between the coldest and the hottest point of the circuit. A valved-off expansion tank, a closed isolation valve that splits the loop into a dead leg, or a design with no free gas and no compliant hose anywhere will find the 9.9 bar/K path the first time the loop warms up. In the rigid case the relief valve becomes the expansion control, which means it discharges coolant on every thermal cycle and leaves you chasing a fluid-loss problem that looks like a leak.

Layer 4 — transients: the load that actually decides the rating

Water hammer is the pressure rise produced when a moving column of coolant is forced to stop faster than the pressure wave can travel the length of the pipe. For an essentially instantaneous stop, the Joukowsky relation gives the ceiling:

Two things follow from t_c, a relation that appears in every water-hammer reference for a reason [13][17][20]. A powered actuator that closes in 20–200 ms is a rapid closure for the short lengths found inside a rack, so "we have a slow actuator" is not an argument unless you have measured it. And the practical US screening form, with its traditional safety factor of about five already included, is the expression to carry into a review:

Stack that onto the steady state from layer one and two, add the thermal rise, and the result is the number that has to be compared against MWP:

The mitigation list is short and mostly mechanical, and it is worth knowing which item is cheapest. Lengthening the closure to well beyond t_c removes most of the surge and costs nothing but an actuator setting — but check the effective closure time, not the overall stroke time, because a valve with a high initial Cv does not begin controlling flow until late in its travel [16]. Removing check-valve slam matters more than most people expect: a check valve that closes against reverse flow is always a rapid closure by the t_c test, and the fix is a soft-close or spring-assisted design rather than a longer pipe. Reducing header velocity cuts the surge linearly and the steady ΔP quadratically, which is the one change that improves both layer two and layer four at once. A bladder surge arrestor at the pump discharge absorbs the wave locally, and belongs on the discharge header rather than on a branch. And never leave a dead-end branch with trapped air: the gas pocket converts a benign pressure change into a violent one, and it does so at the position where nobody instruments anything [13][16][17].

Map the loop: which pressure class belongs at which interface

Pressure ratings in a direct-to-chip rack are not one ladder; they are a set of interfaces, each with its own governing document and its own number. Selection means pinning a class to every interface and then taking the minimum as the loop's rating.

What a 1.5× proof test proves — and what it structurally cannot see

Pressure testing a liquid-cooling loop borrows its framework from building services and process piping practice. ASME B31.9 sets the hydrostatic test pressure at 1.5× the design pressure at the test temperature, with a pneumatic alternative at 1.25× permitted only under strictly controlled conditions, and it requires the test pressure to be limited by the material's yield strength rather than applied as a blanket multiplier. The minimum hold for leak examination is about 10 minutes, but data centre specifications commonly require two hours or more, and 24 hours is not unusual for a section that will never be accessible again [14][15][19]. Process piping practice under B31.3 adds the temperature correction Pt = 1.5 P (S_t / S) and the requirement to confirm that hoop stress stays below yield at the test pressure [21].

The field mechanics matter as much as the number. Isolate everything not rated for the test pressure — the CDU heat exchanger, pumps, expansion tank and instruments — and use blinds rather than relying on closed valves as a boundary, because a passing valve seat will confuse your hold. Remove or protect relief valves, or they will lift and dump the test. Choose a gauge whose calibration is in date and whose range puts the test pressure near mid-scale, typically 1.5–4× the test pressure, so you can resolve the change you are trying to measure. Vent every high point, because trapped air makes the hold spongy and hides a leak. Repair any leak dry at reduced pressure and rerun the full hold from the start rather than tightening a weeping joint under pressure. And log the water and ambient temperatures with every pressure reading: a pressure drop that tracks a temperature drop is thermal drift, not a leak, and without the temperature log you cannot tell the two apart [14][15].

Two practical rules for testing near connectors. Test a connector subassembly to 1.5× its own MWP — about 10.3 bar for an OCP UQD — as a factory or bench test, never as a whole-loop field test. And never pressure-test through a component whose working pressure is lower than the test pressure, which in an ORv3 rack means the blind-mate chassis interface has to be excluded or separately blanked, because it is below the working pressure of everything else on the branch [3][14].

A related question deserves its own answer: a loop designed to run at sub-atmospheric pressure changes the failure mode, not the rating. On gauge pressure a negative-pressure loop looks stress-free, but the hoses now need an external-pressure or collapse rating, the seals need to be qualified for inward leakage, and the pump's suction margin is the binding constraint. Negative pressure is excellent at containing leaks — a breach pulls air in rather than pushing coolant out — and it does nothing to reduce the transient question, because a valve closure in a sub-atmospheric loop still produces a positive-going pressure wave on the upstream side.

Size by flow, then check the force

Pressure class is a ceiling; flow capacity is the reason the coupling is there at all. The two are connected through Cv, and the connection is quadratic: at the same flow, halving the Cv quadruples the pressure drop, and that drop is a permanent addition to layer two.

The mis-sizing penalty is severe and it compounds. Run the UQD04 flow rate through a UQD02 geometry and the pressure drop rises from 0.112 bar to 1.338 bar — a factor of twelve, exactly the square of the Cv ratio [3][8]. As a pair, supply plus return, that is 2.68 bar of pure coupling loss, more than the entire loop ΔP budget on its own. It also raises the pump discharge pressure, which raises the peak in layer four, which eats the burst margin that the pressure class was protecting. Undersizing a coupling is a pressure decision disguised as a flow decision.

Force is the other half. The separating load on a coupling is the pressure times the effective bore area, and the numbers are not trivial once the size goes up: at MWP a UQD06 is being pushed apart with about 54 N and a UQD08 with about 92 N, and at the burst threshold those become 162 N and 275 N. For scale, one production ORv3 blind-mate coupling publishes a static connect force of about 67 N dry and up to 99 N when wet, with a dynamic requirement of 120 N at 40 mm/s and validation to 400 mm/s [2][22].

Temperature, seals and the three ways heat changes the pressure question

Temperature enters a pressure selection three times, and only one of them is about the metal.

Fluid properties and vapour pressure. A fluid's pressure envelope includes the point at which it starts to boil, and a hot loop running at low absolute pressure at its highest point has the smaller margin. This is also why coolant temperature classes matter: the class an operator designs to sets the water temperatures and flow rates the loop must handle, and the loop's temperature sets both the ΔP and the seal life [8].

The seal, not the housing. At MWP the part that loses first is usually the elastomer. EPDM is the standard choice for water-glycol direct-to-chip loops, and its compression set and extrusion resistance are what define the top of the working temperature window. The published windows reflect that: 5–65 °C for the OCP UQD operating range, 0–70 °C for one vendor's push-button UQD, and −40 to +150 °C for a blind-mate coupling where the envelope is quoted for the metal and elastomer together [26][27][22]. Treat any quoted window as a joint property of fluid, seal and housing, and ask which one sets the limit.

Viscosity, which becomes pump head. Cold glycol is viscous, and a loop commissioned at 15 °C with a cold-start sequence can see a materially higher pump head than the same loop at operating temperature. If the pump is constant-speed and the relief and MWP margins were calculated at design temperature only, the cold-start transient is the case that finds the weak point. This is the same mechanism that pushes fouled loops toward their limits, and it is why the 15 % margin in the pump-head budget is not padding.

Glycol concentration moves every number in this article. A 50 % glycol mix has higher density and viscosity than a 25 % mix, so static head, pump head and the loss across every coupling all rise, while the expansion coefficient and the pressure wave speed change as well. Make the selection against the fluid you will actually run, and recheck it if the concentration changes — a loop specified on water and commissioned on PG30 has different pressures at every point.

The evidence to demand before you accept the part

A pressure class on a web page is a claim. These are the artefacts that turn it into a specification, and each one maps to a specific thing that can go wrong.

• Working pressure with its temperature window, not a single number — and the derating basis where the window ends.
• Burst pressure and its test basis. Ask whether it is a lot-sample or a per-unit result, and to what standard; the OCP UQD requirement is a minimum of 300 psi, so a supplier claiming exactly that is claiming the floor, not an achievement [1][3].
• Proof-test pressure and hold time on the component, with a certificate. Data centre acceptance holds are commonly two hours or longer even where the code minimum is ten minutes [14].
• Leak-test method and sensitivity. Helium leak testing is the practical acceptance test for a coupling; one supplier performs it on 100 % of blind-mate couplings, and one published GB200 cold plate kit quotes a leak rate sensitivity of 1×10⁻⁹ mbar·L/s with every unit tested [22][23][25].
• Cycle count to which the spillage limit still holds, not the cycle count at which the coupling still connects. OCP durability requirements are in the thousands of cycles, and vendors publish more; the two numbers are different properties and they fail independently [1][2][3].
• Separating force versus latch capacity, and for blind-mate parts the misalignment envelope plus the latch-verification method — how the installer confirms full engagement without being able to see it [22][23].
• Wetted materials and the qualified coolant, including whether the qualification covers your glycol concentration rather than water only.

Selection checklist

The outcome of the exercise is usually not a bigger connector. It is a defined loop pressure, a relief setting that protects the weakest part, one mitigated transient, and a coupling whose class matches an interface rather than a catalogue. If you want that run against your own rack, send us the rack power, the coolant and the loop geometry through the quote request form and we will work the pressure stack with you.

References

1. Open Compute Project, "Cooling Environments / Cold Plate" (project page, 2026) — UQD v2, BMQC and PBMC workstreams, manifold specification, coolant guidelines, and the OCP UQD performance requirements including 100 psi operating and 300 psi minimum burst. opencompute.org
2. UQD-USA, "UQD Series Products" — OCP Rev 1.0 performance requirements by size: Cv, fluid loss, pressure, temperature, cycle life, service and shelf life; validation test methods; IEC 62368-1 Clause G.15. uqd-usa.com
3. Amphenol Industrial Operations, "UQD Universal Quick Disconnect" datasheet — 100 psi operating / 300 psi burst, Cv and recommended flow by size, fluid loss per disconnect, 5,000-cycle durability, 10-year service life. farnell.com
4. Stäubli, liquid-cooling coupling technical requirements (industry technical exchange, 2026) — metal-body connector qualification: ≥1,000 connection cycles, 16 bar pressure capability, thermal-shock and humidity cycling, dry-break capability, colour coding and production traceability. 163.com
5. MasterChem, "Thermophysical properties of a 25 % aqueous solution of propylene glycol" — density 1,019 kg/m³ at 20 °C falling to 981 kg/m³ at 80 °C, heat capacity and thermal conductivity tables used for the static-head and viscosity figures. masterchem.fi
6. DOWFROST LC 25 technical data — 25 vol % inhibited propylene glycol for datacom equipment and direct-to-chip loops; specific gravity 1.030–1.036; volume expansion 5.1 % between −10 °C and 90 °C; freeze protection to −10 °C. thermalfluidproducts.com
7. Atlantis Industrial, "Cold plate flow and pressure reference" (2026) — pressure-drop correction factors by coolant: PG25 ×2.22, EG25 ×1.95, EG50 ×5.48; blockage raises ΔP 30–80 % before thermal breach. re-atlantis.tw
8. Sector Signals, "OCP's data center liquid cooling standardization: from cold plates and CDU to UQD" — OCP Cooling Environments scope, UQD specification revision history, and the in-rack technology cooling loop operating band of 140–450 kPa with ASHRAE coolant classes. sector-signals.net
9. EngiCalcsHub, "Expansion Tank Sizing Guide for Hydronic HVAC Systems" — expansion factor from saturated-water specific volumes, acceptance volume = system volume × expansion factor, Boyle's law basis for total tank volume, and the rule that pressures in the acceptance calculation must be absolute. engicalchub.com
10. CalcEngineer, "Expansion Tank Sizing Calculator — Hydronic & Water Heater" — acceptance factor AF = 1 − P_fill_abs / P_max_abs, required tank volume = acceptance ÷ AF, and the statement that using gauge pressure is the most common expansion-tank sizing error because it understates the required volume. calcengineer.com
11. Jaspector, "IRC 2024 Hydronic Expansion Tank Sizing" — acceptance factor worked from pre-charge and maximum operating pressures, bladder versus plain-steel tank sizing, and the requirement that pre-charge match the system fill pressure at the tank elevation. jaspector.com
12. Reuven Engineering Tools, "Expansion Tank Sizing" — ASHRAE method: expansion factor εf = (v_f2/v_f1 − 1) − 3αΔT, diaphragm acceptance factor AF = 1 − P_i/P_f, required volume V_t = V_e/AF. reuven.tools
13. EngiCalcsHub, "Water Hammer Pressure Surge — Piping Transient Guide" — Joukowsky relation ΔP = ρcΔv, wave speed 1,000–1,500 m/s in metal pipe, the critical closure time t_c = 2L/c, and the two regimes of rapid and slow closure. engicalchub.com
14. Anvil, "Chilled water hydro test package field guide for data centers" — 1.5× design pressure per ASME B31.9, code minimum hold of about 10 minutes against a data centre specification acceptance of two hours or more, mid-scale gauge selection, blinds rather than closed valves as boundaries, equipment isolation, and the temperature log needed to separate thermal drift from a leak. anvilfield.com
15. Make Piping Easy, "Hydrotest Procedure for Piping System" — test pressure table by code (B31.9 at 1.5× design, 1.25× pneumatic), gauge range selected between 1.5× and 4× the test pressure, test water temperature limits, and the preparation sequence for a hydrostatic test. makepipingeasy.com
16. Waterhammer.com (Datacor Pipe Flow Modeling Division), "Effective Closure Time" — why overall valve stroke time does not equal the time the valve spends controlling flow, and how initial Cv determines whether a nominally slow closure still produces a full Joukowsky surge. waterhammer.com
17. Wikipedia, "Water hammer" — Joukowsky equation, rigid-column theory with the US practical form P = 0.0135VL/t and the 5× safety-factor form P = 0.07VL/t + P₁, and the definition of the pressure wave-speed terms. en.wikipedia.org
18. Progressive Robot, "Liquid Cooling Failures in AI Data Centers: Causes, Risks, Prevention" — failure-mode taxonomy, joint-count dominance of liquid-cooling failures, weep/drip/spray classification, and commissioning practice that damages couplings before they enter service. progressiverobot.com
19. Fengshecad, "ASME B31.9-2020 Guide: On-Site Installation and Pressure Testing Rules for Building Services Piping" — 1.5× design pressure at test temperature, minimum 10-minute hold, isolation of equipment not rated for the test pressure, staged pressurisation, and the test report and material certification file required for occupancy approval. fengshecad.com
20. Design Calculators, "Water Hammer in Piping: the Joukowsky Equation, Pump Trips" — acoustic wave speed 900–1,250 m/s in typical steel piping, diameter independence of the Joukowsky head rise, and the energy conversion argument behind the surge. designcalculators.co.in
21. UniPiping, "What is ASME B31.3 Pressure Leak Testing?" — hydrostatic test pressure not less than 1.5× design pressure, the temperature-corrected form Pt = 1.5 P (S_t/S), and the requirement to confirm hoop stress stays below yield at test pressure. unipiping.com
22. Hansen Fluid, "DANFOSS Blind Mate Valves (BMQC)" — ORv3 blind-mate coupling data: working pressure 2.4 bar, static connect force 66.71 N dry and 99.3 N maximum, dynamic connect force 120 N at 40 mm/s with validation to 400 mm/s, ±5 mm radial and ±2.7° angular compensation, 303 stainless body, EPDM seal, −40 to +150 °C. hansenfluid.com
23. Danfoss Power Solutions, "New Blind Mate Quick Connector for data center liquid cooling" — ORv3-compliant BMQC with 2.4 bar (35 psi) working pressure, 6 L/min rated and 10 L/min maximum flow, 0.15 bar pressure drop at 6 L/min, 100 % helium leak testing on every unit, and 5 mm size interchangeable with other ORv3 blind-mate couplings. danfoss.com
24. J. Watkins, "Liquid Cooling Interconnects: Connector Types and Standards" (2026) — UQD/UQDB/BMQC/LQC/MQD family map, the ORv3 manifold specification, GB200 NVL72 hydraulic requirements (CDU output ≥750–800 LPM, manifold pressure drop 1.5–2.5 bar, cold plate burst pressure ≥12 bar), and the operational cost of connector-family fragmentation including mis-mating risk under pressure. jaredwatkins.com
25. ToneCooling, "NVIDIA GB200 cold plate kit" product data — 0.6 MPa working pressure with ≥4 MPa burst on lot samples, 35 ± 5 kPa pressure drop at 2.5 L/min, helium leak rate ≤1×10⁻⁹ mbar·L/s on every production unit, and coolant compatibility covering 50/50 EGW and 25 % propylene glycol. tonecooling.com.au
26. DC Market Insights, "Liquid Cooling Manifold Market" (2026) — OCP UQD 2.0.0 maximum operating pressure of 6.89 bar over a 5–65 °C range for water and glycol systems, the ORV3 blind-mate quick connector specification at 3.45 bar maximum working pressure tested with a 25 % propylene glycol mixture, and Parker's push-button UQD at 7 bar (100 psi). dcmarketinsights.com
27. Parker Hannifin, "Push-Button UQD — Universal Quick Disconnect for Liquid Cooling" product bulletin — working pressure 7 bar (100 psi), working temperature 0–70 °C, connection/disconnection durability >5,000 cycles, flow and pressure-drop curves for UQD04 and UQD06, and coolant compatibility covering propylene glycol, ethylene glycol and deionised water. Parker product bulletin (PDF)