Leak Detection in Liquid Cooling Ends Where Valves Begin

A leak detection rope fires only when electrically conductive coolant touches it. A solenoid valve that then shuts faster than the pressure wave can cross its pipe and return can add 3,108 kPa to a loop that normally operates at 140 to 450 kPa. The Open Compute Project (OCP) documents on cold plate cooling describe leak control as a chain of joints, sensors and intervention. The chain is usually read from the sensor outward, but the harder limit sits at the valve.

Most leak control is decided at the joints, before any sensor goes in

The OCP white paper on leak detection and intervention starts with pipework, and that order is correct. Its scope is the technology cooling system (TCS), the loop that runs from the coolant distribution unit (CDU) through rack manifolds and cold plates and back. The paper states that every connection point in that loop is a potential leak. It prefers parallel-thread BSPP fittings sealed with an O-ring over tapered BSPT threads sealed with tape or putty, and it recommends brazing or welding any joint that never needs service.

Plastic piping takes the same logic further. The paper describes polypropylene-random (PP-R) pipe joined by fusion, where heated pipe and fitting become a single piece. The earlier OCP cold plate requirements document puts it plainly: the main approach to leak mitigation should be a leak prevention design.

Commissioning is where those joint choices get tested. The white paper describes a gas dwell test that fills the system with dry nitrogen to just below the relief valve setting, 35 psi in its example, and holds it for 8 hours with no pressure loss.

A hydrostatic step follows when the facility water system is filled: 14.5 psi (1 bar) for one hour, again with no drop.

The paper’s two field cases from US National Labs show what these tests do not catch. In the first, an inexpensive O-ring was incompatible with the coolant and leaked slowly, below the trip level of the facility leak detection system, so the loop lost volume until operators traced it. In the second, quick disconnects dripped because hanging hoses pressed sideways on their seals, and supporting the hoses stopped the leaks without replacing a connector. Neither failure was a sensor problem. I read both as evidence that the first layer of leak management is material compatibility and mechanical support, and that a sensor only records the result of decisions made at the joints.

Source: Open Compute Project, “Leak Detection and Intervention” white paper, Revision 1.0 (undated); OCP, “ACS Liquid Cooling Cold Plate Requirements” Revision 1.0 (October 9, 2019)

A sensor detects only the coolant that reaches it

Direct detection has two conditions, and the OCP white paper states both: the liquid must touch the sensor in enough volume, and it must conduct electricity. A rope sensor is two or four electrodes separated by insulation, and conductive coolant bridging them changes the resistance. The four-wire version also reports where along the cable the wetting occurred.

Those two conditions exclude common cases. Dielectric coolants do not complete the circuit. A slow glycol leak often leaves only residue, because the water evaporates before liquid pools, and the paper says a rope would not catch it, while noting that such a leak likely would not damage IT equipment either.

Indirect detection covers part of the gap. The paper lists differential pressure over time, reservoir level, optical spot sensors and turbidity sensors, and it says differential pressure across the TCS pumps should be considered a requirement, since most CDUs already measure it. The catch is baseline drift. Adding or removing racks shifts the differential pressure, and temperature swings shift the reservoir level, so both have to be modeled before a trend means anything.

The paper’s severity matrix makes the ordering explicit. A minor leak with no visible spill trips indirect detection. A major leak, visible inside IT equipment, trips direct detection in one zone, and a critical leak pools outside the rack and trips several. By the time a rope is wet, the leak has already left the pipe, so I read direct sensors as the second line of defense rather than the first.

False alarms are the other constraint. Condensation on uninsulated pipe reads as a leak, and the paper warns that low thresholds teach operators to ignore alarms. Its fix is voting: either a tray sensor or a manifold rope raises an alarm, but automatic intervention runs only when both detect fluid. That rule decides when a valve moves.

Source: Open Compute Project, “Leak Detection and Intervention” white paper, Revision 1.0 (undated)

Closing a valve fast enough creates a second failure

Automatic isolation trades spill volume against surge pressure. The OCP white paper names the limit in one sentence: the plumbing engineer must make sure a valve cannot close fast enough to cause water hammer, the pressure wave produced when moving liquid is stopped suddenly. The paper gives no numbers. The governing equations do.

For a closure faster than the critical time, the Joukowsky equation gives the surge: delta-P = rho × a × delta-v, where rho is coolant density, a is the pressure wave speed in the pipe and delta-v is the change in velocity. The critical time is 2L/a, the round trip of the wave over pipe length L. For slower closure, the Allievi-Michaud approximation, which a 2025 Journal of Fluid Mechanics paper describes as widely used in design standards for a first estimate, gives delta-P = 2 × rho × L × delta-v / T, where T is the closure time.

The assumptions, with units:

• Coolant density rho = 1,000 kg/m3, taken as water; a propylene glycol mixture is denser and would raise every result slightly
• Velocity 2.1 m/s, the upper end of the 1.5 to 2.1 m/s maximum pipe velocities the OCP cold plate requirements cite from ASHRAE
• Wave speed 1,480 m/s for water in very stiff pipe, and 200 m/s as the level some plastic pipe falls below
• Pipe length L = 30 m, an illustrative CDU-to-rack run, not a published figure
• Operating pressure 450 kPa, the top of the typical TCS range

Instant closure in stiff pipe gives 1,000 × 1,480 × 2.1 = 3,108,000 Pa, or 3,108 kPa.

The critical time over 30 m is 2 × 30 / 1,480 = 0.041 s. Any valve faster than that behaves as instantaneous.

A 5 s closure over the same run gives 2 × 1,000 × 30 × 2.1 / 5 = 25,200 Pa, or 25.2 kPa.

Table 1. Water hammer surge in a TCS pipe run, by closure casekPa

* Surge plus 450 kPa, top of typical TCS operating range; 3 times 450 kPa = 1,350 kPa leak test level cited from IEC 62368-1

** Allievi-Michaud estimate, valid only when closure exceeds 2L/a; density 1,000 kg/m3; single straight pipe, friction and branching ignored; calculated estimates, not measurements

Source: calculation from Joukowsky and Allievi-Michaud relations; inputs from OCP, “ACS Liquid Cooling Cold Plate Requirements” Rev. 1.0 (2019) and Fluid Mechanics Ltd, “Water Hammer Analysis” (undated)

These are estimates for a single straight pipe. They ignore friction, branching and hose compliance, and Applied Flow Technology notes cases where the Joukowsky value is not conservative. The comparison still holds. The OCP cold plate requirements note that IEC 62368-1 calls for leak testing at 3 times normal operating pressure, or 1,350 kPa for a 450 kPa loop.

An instantaneous close in stiff pipe peaks at 3,558 kPa, well past that test level, while a 5 s close stays at 475.2 kPa. A valve specified for speed alone can turn a detected drip into a failed fitting.

The CDU is the coarsest valve available. The white paper warns that a CDU shutdown can ramp the pumps down very quickly, that every cold-plated chip must have thermal throttling enabled or stopped flow could cause a boil event, and that a row-level CDU often serves 10 or more racks.

Source: Open Compute Project, “Leak Detection and Intervention” white paper, Revision 1.0 (undated); OCP, “ACS Liquid Cooling Cold Plate Requirements” Revision 1.0 (October 9, 2019); Applied Flow Technology, “Can You Trust the Joukowsky Equation for Waterhammer?” (2021); Fluid Mechanics Ltd, “Water Hammer Analysis” (undated); Journal of Fluid Mechanics, “Weakly nonlinear theory of water hammer induced by slow valve manoeuvres” (2025)

The limit is how much of the loop one valve takes down

Leak management in cold plate cooling is bounded by the isolation zone, not by the sensor. The OCP white paper describes an inverse relationship between operational impact and equipment damage. Stopping flow at the CDU protects hardware and shuts down every liquid-cooled server on it, while only alerting an operator keeps servers running and risks the hardware. Valves closer to the IT equipment shrink the affected zone, but each one adds cost and another potential failure point.

Containment does not remove the tradeoff. The paper says a drip tray should hold enough coolant for an alarm or action to occur, and that a leak from a high-flow manifold or high-pressure facility water is unlikely to be contained for any duration. Negative pressure systems change the premise, because a breach draws air in rather than pushing coolant out. The paper reports National Labs evaluating them but gives no field data, so they remain an option under evaluation.

The design question is therefore not which sensor to buy. It is how small a zone the operator can afford to isolate, and how slowly each valve in that zone may close without exceeding the rating of the weakest fitting. Those two quantities are linked through the loop layout, and sensor coverage follows from them.

The sources set limits on these numbers. The white paper states that it is not a specification for design. The OCP OAI liquid cooling guidelines give an annual failure rate of 0.3% or lower for cold plates and coolant loops as a desired target, not a measured field rate. None of the documents publish valve closure times, wave speeds for specific TCS pipe, or measured surge on an operating installation. Applying any of this to a real system requires review by the design engineer and the installing contractor. Points that review should cover:

• Valve closure time against the critical time of the longest run it isolates
• Estimated peak pressure against the rating of every fitting, including quick disconnects
• The voting logic that decides when automatic isolation runs
• The wetted materials list against the coolant chemistry

Source: Open Compute Project, “Leak Detection and Intervention” white paper, Revision 1.0 (undated); OCP, “OAI System Liquid Cooling Guidelines” (March 2023 update)

References

Open Compute Project, “Leak Detection and Intervention” white paper, Revision 1.0 / Open Compute Project Foundation / undated / https://www.opencompute.org/documents/acs-cold-plate-leak-detection-and-intervention-white-paper-pdf-1

Open Compute Project, “ACS Liquid Cooling Cold Plate Requirements” Revision 1.0 / Open Compute Project Foundation / October 9, 2019 / https://www.opencompute.org/documents/ocp-acs-liquid-cooling-cold-plate-requirements-pdf

Open Compute Project, “OAI System Liquid Cooling Guidelines” / Open Compute Project Foundation / March 3, 2023 update / https://www.opencompute.org/documents/oai-system-liquid-cooling-guidelines-in-ocp-template-mar-3-2023-update-pdf

Applied Flow Technology, “Can You Trust the Joukowsky Equation for Waterhammer?” / Applied Flow Technology / 2021 / https://www.aft.com/about-aft/news/890-can-you-trust-the-joukowsky-equation-for-waterhammer

Fluid Mechanics Ltd, “Water Hammer Analysis” / Fluid Mechanics Ltd / undated / https://www.fluidmechanics.co.uk/hydraulic-calculations/water-hammer-surge-analysis/

“Weakly nonlinear theory of water hammer induced by slow valve manoeuvres” / Journal of Fluid Mechanics, Cambridge University Press / 2025 / https://www.cambridge.org/core/journals/journal-of-fluid-mechanics/article/weakly-nonlinear-theory-of-water-hammer-induced-by-slow-valve-manoeuvres/BDD8119158445500DA7FE19586BF0678