Liquid Cooling Leak Detection: What Sensors Miss and Inspection Catches

• Liquid cooled AI racks such as GB200 NVL72 put coolant next to powered electronics, while most halls still run mostly on air.
• Liquid cooling leaks start at connections: quick disconnects, couplings, joints, strained hoses, CDU seals and tank lids.
• Rope and point sensors only alarm when coolant reaches them, so slow weeps and dried residue can go unnoticed.
• Inspect every coupler, hose, fitting, manifold and drip tray from several angles, and track heat at pumps, CDUs and heat exchangers.
• FM's data sheet for data centers calls for regular inspection of liquid piping up to the CDUs, plus containment and monitored leak detection.
• Record every leak with a time, a location and photos, starting with the pressure test before first fill.
• Coolant chemistry and the inside of a cold plate need lab sampling and system monitoring; inspection complements them.

Short answer: Because AI racks now bring liquid right to the processors, and a leak there lands on powered equipment. GB200 class racks are liquid cooled by design, and while most halls are still air cooled, the share of liquid cooled racks is growing.

NVIDIA describes its GB200 NVL72 as a rack scale, liquid cooled design, and server makers list these racks at 132 kW and up to 130 kW. At that density, liquid is no longer an option for the GPUs; it is the cooling system.

Adoption is real but early. In Uptime Institute's 2025 cooling survey, 22% of respondents used direct liquid cooling and 75% used perimeter air, and its 2024 survey found nearly half of direct liquid cooling users had it on fewer than 10% of their racks.

So most operators run mixed halls, with liquid at the newest racks and air everywhere else. Air still matters for the optics in liquid cooled racks too, as our guide to AI data center cable management explains.

The stakes are different from a chilled water leak under a raised floor. Coolant at the rack sits inches from powered electronics, and every rack carries many connections that are made, broken and remade during service.

Liquid cooling moves the plumbing into the rack. Leak detection has to follow it there.

Short answer: Mostly at connections, not in pipe walls: quick disconnects, couplings, joints, hoses under strain, manifold connections, CDU pump seals and, for immersion, tank lids and seals. Debris, side load and pressure surges during fill and drain all make those connections weep.

Liquid cooling loops fail where parts meet. Schneider Electric's White Paper 210 puts it plainly: leaks tend to occur at couplings and joints, not at the piping walls (vendor).

Where liquid cooling leaks start

• Quick disconnects: the OCP leak detection white paper notes that a quick disconnect is more likely to leak than a fixed connector when it is disconnected.
• Side load on couplers: at one national lab described in the same paper, hanging cooling lines pulled sideways on quick disconnects and deformed their seals, causing slow leaks.
• Debris: Schneider notes that debris can jam quick connects and cause leaks, which is one reason loops are flushed and filtered.
• Pressure surges: the OCP quick disconnect specification warns that a high pressure drop during fill and drain can wash out a seal.
• Hoses and fittings: kinked, strained or rubbing behind the rack.
• CDUs: pump seals, filters and heat exchanger connections.
• Immersion tanks: lids, seals and seams, with OCP's immersion requirements listing seal failure among the sources of fluid loss.

Notice what these have in common. Almost every one is at a joint you can see from outside the rack, and almost every one starts small.

Leaks start at connections. Connections are visible.

Short answer: Sensors only detect liquid that reaches them. A rope sensor needs coolant to touch the cable, a point sensor sees one spot, and a slow weep can dry to residue before it ever gets there.

Leak detection cable, often called rope, runs under manifolds and along trays and alarms when liquid bridges it. The OCP white paper notes that the coolant must directly touch the cable, and that slow leaks, which leave residue from glycol and additives, would not be caught by a rope.

Point sensors are narrower still. Chemelex's guide notes that a point sensor detects a leak only at its own spot, and fluid can run around it (vendor), while the OCP integration and logistics paper adds that detection only works reliably when leaking fluid is guided to the sensor.

None of this makes sensors less important. They watch every second, they are the right tool for the fast leak, and FM's data sheet for data centers calls for leak detection that alarms to a constantly monitored system.

Operators feel the gap. In The Register's 2024 survey of 812 IT professionals, 41% expressed fears over leaks and spills, and inspection is how you answer that fear with evidence.

Sensors and inspection, side by side

• Sensors: continuous, alarm in seconds and catch fast leaks, but only where liquid reaches them.
• Inspection: covers every fitting in scope and sees weeps, residue, strain, wet or missing drip trays and valves in the wrong position, on a schedule and with a photo.

Sensors catch the leak that reaches them. Inspection catches the one that has not yet.

Short answer: Look at every coupler, hose, fitting, manifold and drip tray from more than one angle, check valves and insulation, and confirm leak detection is in place. At the CDU, watch heat at pumps, motors and heat exchangers, filter pressure drop and local gauges.

Direct to chip cooling runs a technology cooling loop from a coolant distribution unit, or CDU, through manifolds and hoses to cold plates on each processor. Vertiv describes the CDU as the unit that creates an isolated secondary loop, keeps supply temperature above the dew point and runs redundant pumps (vendor).

At the rack and manifold

• Couplers: fully latched and not pulled sideways by hanging hoses.
• Hoses: no kinks, rubbing, strain or damage.
• Manifolds and fittings: dry from several angles, because a weep often shows from only one side.
• Drip trays: present under connections, dry and free of debris.
• Residue: no white or colored deposits around joints, the sign of a slow leak that has dried.
• Leak detection: ropes and sensors present, in place and undamaged.

At the CDU

• Pumps and motors: running heat compared with their own baseline.
• Heat exchanger: connections dry and heat pattern normal.
• Filters: pressure drop tracked over time, as the ASHRAE TC 9.9 paper on water cooled servers recommends.
• Valves and gauges: valves tagged and in position, and local gauges read and recorded, the kind of gauges nobody monitors.

Schneider's paper notes that visual inspection is sometimes the best monitoring, and that long term leaks leave mineral traces in plenums, racks and soffits (vendor). Thermal imaging adds what the eye cannot see, as long as each scan is taken from the same position every round.

A camera at the fitting sees what a sensor on the floor cannot.

Short answer: Check fluid levels through the sight glass, lids and seals, the floor and containment around each tank, and heat at pumps and heat exchangers. Cold boxes in standard racks, rear door heat exchangers and air cooled rows each get the same visual and thermal check.

Immersion tanks hold servers in dielectric fluid, so the leak points move from the rack to the tank. OCP's immersion requirements treat fluid level detection and leak detection as high safety requirements, expect open baths to have a lid, and list seal failure and improper maintenance among the sources of fluid loss.

Inspection points by cooling type

• Immersion tanks: fluid level through the sight glass, lids closed and seals intact, no fluid outside the tank, containment in place and dry.
• Cold boxes in standard racks: read like any other rack, with connections and trays checked at the rack face.
• Rear door heat exchangers: connections and hoses dry, no condensation on the coil or the floor below.
• Air cooled rows: the thermal baseline that still covers most of the hall.

Whatever the cooling, the power gain only holds if the plant stays healthy. The inspection changes shape from one cooling type to the next, but the questions stay the same: is it dry, is it where it should be, and is it running at its normal temperature.

Cold boxes deserve extra attention because they look like ordinary servers from the aisle. Their connections sit at the back of a standard rack, often behind doors, so a closed door should be recorded as not inspected rather than assumed dry.

Every cooling type leaks somewhere. The job is knowing where to look.

Short answer: Regularly, and on every round in the weeks after first fill. FM's data sheet for data centers calls for routine inspection of liquid piping in equipment rooms up to the connections at the CDUs, for early signs of leaks, corrosion and damage.

Insurers have started to write this down. FM's Data Sheet 5-32, revised in 2026, asks for regular inspection of liquid piping inside data processing rooms up to the point where it connects to CDUs or heat exchangers, as well as containment below liquid cooled systems and monitored leak detection.

Equipment makers ask for it too. Dell's preventive maintenance plan for its PowerCool systems includes periodic on site visual inspection for observable leaks and wear, alongside a leak detection check and a coolant sample (vendor).

The first weeks after first fill are when a coupler that was never fully latched shows itself. Check every fitting on every round then, and keep a regular route after that, so the record shows each fitting dry, round after round.

Write the route down. A list of every fitting, valve and tray in scope, in the order they are walked, is what turns regular inspection into a record you can prove.

The best inspection interval is the one that catches the weep before the rope does.

Short answer: The date and time, the exact location, the fitting, photos from several angles, how it was found, what was done and by whom, and its status on the next round. A leak record is a history, not a single event.

The leak record: what to capture for every leak

• When: the date and time it was found.
• Where: the exact location, for example Data Hall 3, CDU 7, and the fitting itself.
• What it looked like: photos from several angles, with any drip, stain or residue visible.
• How it was found: by a sensor alarm or by inspection.
• What was done: the fix, who made it and when.
• What happened next: its status on the following rounds, and whether any stain or residue is growing.

One photo shows a drip. A series of photos from the same position shows whether it is growing, whether the fix held and whether the same coupler keeps coming back, which is the evidence audits and insurers ask for.

Start the record at first fill, not at the first alarm. The leaks found during pressure testing and commissioning are the baseline every later finding is compared with.

A leak with a history can be managed. A leak without one can only be mopped.

Short answer: The whole loop is pressure tested before it is filled, with liquid or with gas at low pressure, and the pressure is watched over time for any drop. The loop is then flushed and filtered so it starts clean, and every leak found is recorded from the first test.

The OCP guidelines for connecting liquid cooled equipment say commissioning should include a pressure test of the entire system, either hydrostatic or low pressure pneumatic, with the method chosen by the design engineer and the commissioning agent. The OCP integration paper adds that the pressure drop over time should be recorded, because a faster than expected decay points to a leak.

After the test, the loop is flushed and filtered. The OCP row manifold guideline walks through that preparation, and the ASHRAE paper sums up the goal in four words: start clean, stay clean.

This is where leaks are cheapest to find. A coupler that weeps during a pressure test is fixed before any server is installed, and the record of that fix becomes the first page of the loop's history, as our data center commissioning checklist describes.

Commissioning also sets the baseline for heat. Thermal images of pumps, CDUs and manifolds under the first real load become the reference that every later scan is compared with.

Test it, flush it and record it, and only then fill it for good.

Short answer: Coolant chemistry, pH, inhibitor levels, biological growth and what happens inside a cold plate. Those need lab sampling and the cooling system's own monitoring.

A camera and a thermal scan see the outside of the loop. They cannot measure what is in the fluid, and they cannot see inside a server or a sealed cold plate.

ASHRAE's 2026 bulletin on coolant integrity, as summarized by StorageReview, names four ways direct to chip loops go wrong: coolant mix drift, air in the loop, poor stainless steel passivation and pressure transients. Most of that is lab and system work, and inspection complements it rather than replacing it.

Short answer: A FacilityOps robot photographs every coupler, hose, fitting, manifold and drip tray in scope from several angles on every round, scans pumps, CDUs and heat exchangers against their own baseline, and records every leak with a time, a location and photos. It works alongside your sensors and lab sampling.

FacilityOps AI (RoboSkill Inc.) builds an autonomous inspection platform for data centers. Whatever the cooling, immersion, cold boxes, cold plate or air, the power gain only holds if the plant stays healthy, and we verify it the same way.

Our inspection robots run fully autonomously, and a technician can take manual control at any time. The platform is hardware agnostic: inspection robots and sensor hubs are the most common way we capture data, but it works with any capture source.

What the robot checks

• Every fitting, every round: a FacilityOps robot photographs every coupler, hose, fitting, manifold and drip tray in scope from several angles.
• Every leak recorded: date and time, exact location on the image (for example Data Hall 3, CDU 7) and photos from several angles per fitting, from first fill and the pressure tests onward.
• Containment and detection: couplers latched, valves tagged and in position, insulation intact, containment present and dry, and leak detection devices present and undamaged.
• Heat at the plant: thermal scans at pumps, CDUs and heat exchangers, and uneven heat visible from outside racks and tanks, against each asset's own baseline.
• Immersion: tank levels through the sight glass, leaks outside tanks and cold boxes, lids and seals.
• Before and after: each round is compared with the last in our 3D digital twin, so a stain that grows or a coupler that moved shows up with dated images.

Built to be trusted

• We check the work. We never do the work. We never connect or disconnect a coupler, turn a valve, open a CDU cabinet or touch the loop. Opening locked CDU cabinets under your approval is on our roadmap.
• Alongside, never instead: inspection works with your leak sensors and your lab sampling, as our guide to why one sensor is never enough explains. Fluid condition stays with the lab.
• People stay in the loop: anything the robot cannot read clearly goes to a person, a closed door is recorded as not inspected, and a person reviews every finding before it reaches you.
• Everything runs on site: detection and the 3D model run on the robot's onboard computer, and you own every image and record. Patent pending.

Where it stands today

Whole facility rounds, the 3D digital twin and change detection run in a production data center today. Since July 28, 2026, a FacilityOps robot has run inspection rounds at the CRSC Data Center, powered by Solutions IS, in Houston, completing 19 autonomous inspection runs and more than 1,600 observations in its first three weeks.

A one row trial on a liquid cooled row gives every fitting its own record, judged on your own data. Most installed halls are still air cooled, so the same rounds keep the thermal baseline on every air cooled row as well.

• Direct to chip: liquid cooling that runs coolant through cold plates mounted on the processors.
• Cold plate: a metal plate with internal channels that carries heat from a chip into the coolant.
• CDU: coolant distribution unit, which pumps, filters and controls the loop that feeds the racks.
• Technology cooling system: the secondary loop between the CDU and the IT equipment.
• Quick disconnect: a coupling that connects and disconnects a hose without draining the loop.
• Manifold: the pipe in a rack or row that distributes coolant to many hoses.
• Drip tray: a tray under connections that catches drops and guides them to a sensor.
• Rope sensor: leak detection cable that alarms when liquid touches it.
• Point sensor: a leak sensor that detects liquid at one spot.
• Immersion cooling: cooling servers by submerging them in dielectric fluid in a tank.
• Cold box: an enclosed liquid cooled unit that fits in a standard rack.
• Rear door heat exchanger: a liquid cooled coil mounted on the back of a rack.
• First fill: the first time a tested loop is filled with coolant.
• Flushing: circulating fluid through a new loop to remove debris before it goes live.

What is the best way to detect coolant leaks in a data center?

Use two layers. Leak detection cable and point sensors under manifolds, in drip trays and around CDUs alarm the moment liquid reaches them. Recurring visual and thermal inspection of every coupler, hose, fitting and drip tray catches slow weeps and residue that never reach a sensor, and records each leak with a time, a location and photos.

Where should leak detection sensors be installed?

Where leaking coolant will actually go: under manifolds and quick disconnects, in drip trays and containment pans, around CDUs and below piping that runs over equipment. Detection only works reliably when fluid is guided to the sensor, so drip trays and pans matter as much as the sensor itself. Follow your cooling vendor's and insurer's guidance on placement.

Why do quick disconnects leak?

Quick disconnects are made and broken during service, and each cycle is a chance for a leak. Seals can deform when hanging hoses pull sideways on them, debris can jam them, and a high pressure drop during fill and drain can wash out a seal. That is why they deserve a check from several angles on every round.

Are rope leak sensors enough for liquid cooling?

No. A rope sensor alarms only when coolant touches the cable, so a slow weep that dries to residue, or a leak that runs somewhere else, can go unnoticed. Rope is the right tool for fast leaks. Pair it with drip trays that guide fluid to it and with regular visual and thermal inspection of every fitting.

What does a CDU do?

A coolant distribution unit runs the liquid loop that cools the servers. It isolates that loop from the facility water, pumps coolant through manifolds and hoses to the cold plates, keeps the supply temperature above the dew point and filters the fluid, usually with redundant pumps. It is the heart of a direct to chip cooling system.

What should be checked on a CDU?

The connections and the area around the unit for any sign of leaks or residue, heat at pumps, motors and the heat exchanger compared with their normal pattern, filter pressure drop over time, local gauges, valve positions and tags, and the leak detection around the unit. Coolant quality needs a lab sample on the vendor's schedule.

How often should liquid cooling piping be inspected?

Regularly, and more often in the weeks after first fill, when a poorly latched coupler shows itself. FM's data sheet for data centers calls for routine inspection of liquid piping in equipment rooms up to the CDUs, and equipment makers include periodic visual inspection in their maintenance plans. Many operators check every fitting on every round.

How is a liquid cooling loop pressure tested before it is filled?

The whole loop is tested with liquid or with gas at low pressure, using a method chosen by the design engineer and the commissioning agent. The pressure is recorded over time, because a faster than expected drop points to a leak. Leaks are fixed and the test repeated, then the loop is flushed and filtered before it is filled.

How do you check for leaks in immersion cooling tanks?

Check the fluid level through the sight glass, look for fluid outside the tank, on the floor and in containment, and confirm lids are closed and seals intact. Watch heat at the pumps and heat exchangers, and inspect any hoses and fittings. Lids, seals and serviced equipment carrying fluid out of the bath are the usual sources of loss.

Do liquid cooled racks still need air cooling?

Usually yes. Many liquid cooled racks cool their processors with liquid but still rely on air for other components, including the switch optics. Most halls are also still mostly air cooled, so air cooled rows keep their own thermal checks alongside the liquid ones.

Can thermal cameras find liquid cooling problems?

They can find several. A thermal scan shows a pump, motor or heat exchanger running hotter than its normal pattern, uneven heat across a rack or tank, and in some cases a cooling path that is not working as it should. It cannot measure coolant chemistry or see inside a sealed cold plate, so it complements sensors and lab sampling.

Can a robot inspect liquid cooling for leaks?

Yes. A FacilityOps robot photographs every coupler, hose, fitting, manifold and drip tray in scope from several angles on every round, scans pumps and CDUs for heat, and records every leak with a time, a location and photos. It never touches the loop, and it works alongside leak sensors and lab sampling, never instead of them.

↑ Back to contents

Liquid cooling moves the risk of a leak into the rack, next to the most expensive equipment in the building. Sensors catch the leaks that reach them, and inspection catches the weeps, residue and strained couplers that have not reached them yet.

Record every leak from the first pressure test, check every fitting on a regular route, and leave the fluid itself to the lab. That is how the power gain of liquid cooling holds up year after year.

The power gain only holds if the plant stays healthy, and the plant stays healthy when every fitting has a record.

• NVIDIA: GB200 NVL72 (vendor)
• Supermicro: GB200 NVL72 rack scale solution (vendor)
• Gigabyte: GIGAPOD with NVIDIA GB200 NVL72 (vendor)
• Uptime Institute: Data center cooling, data report, July 2025
• Uptime Institute: Cooling systems survey 2024: direct liquid cooling, May 30, 2024
• Open Compute Project: Cold plate leak detection and intervention white paper
• Open Compute Project: Liquid cooling integration and logistics white paper
• Open Compute Project: Universal quick disconnect specification, version 2.0.0, 2026
• Open Compute Project: Guidelines for connection of liquid cooled IT equipment, April 2022
• Open Compute Project: Pre commission preparation of technology cooling system row manifolds, March 2025
• Open Compute Project: Immersion requirements, revision 2.0, January 2022
• ASHRAE TC 9.9: Water cooled servers white paper, 2019
• StorageReview: ASHRAE TC 9.9 flags four ways direct to chip cooling loops go wrong, August 30, 2026
• Schneider Electric: White Paper 210, 2025 (vendor)
• Vertiv: Understanding coolant distribution units for liquid cooling, August 2, 2023 (vendor)
• Chemelex: Data center leak detection white paper, 2025 (vendor)
• Dell Technologies: Preventive maintenance for PowerCool DLC systems, 2025 (vendor)
• FM: Property Loss Prevention Data Sheet 5-32, Data Centers and Related Facilities, January 2026
• The Register: Liquid cooling survey, April 22, 2024

• Data Center Commissioning Checklist: What to Inspect Before Handover, Trade by Trade
• Fiber Cable Management in AI Data Centers: Tight Bends, Blocked Airflow and Overheating Optics
• Thermal Imaging in Data Centers: Seeing Heat Before It Becomes Downtime