Water leak detection for data centers, from construction through operations

Data centers have a strange relationship with water. The cooling plant pushes water around the clock through branches, loops and headers that cross the white space. The load underneath tolerates none of it. As rack densities climb and liquid cooling moves closer to the chip, the volume of water inside the hall goes up, not down.

Two more systems have nothing to do with cooling and cross the building anyway. The domestic main feeds washrooms and make-up water. The fire sprinkler line covers the same ceiling as everything else. Both can fail quietly, and both usually fail after hours: across Eddy's monitored portfolio, 59% of water events land outside business hours, with the peak at 2 AM.

Where the water actually sits

Map the water before you spec the coverage. In a typical facility:

• Cooling branches and expansion loops. Chilled water or glycol runs in drip trays above or beside the racks. A single hall can carry dozens of branch runs plus a header measured in hundreds of feet.
• CRAH and CRAC units. Condensate pans, pumps and humidification lines that run around the clock, whether or not anyone is in the building.
• The domestic main. Washrooms, janitor closets, make-up water. The one system that can and should shut itself off automatically.
• The fire sprinkler line. A life-safety system that must never be auto-closed. It can still be metered, so movement on the line raises an alarm in seconds.
• During construction: the temporary supply. Pressurized early, extended floor by floor and unattended at exactly the hours when leaks do the most damage.

Construction is the highest-risk phase

Across Eddy's portfolio, 69.6% of active construction sites confirmed at least one leak, roughly 3 times the rate of operating buildings, and 52% of construction leaks trace to pipes: fitting failures, pressurization errors, connections that let go on a Friday night. Zurich attributes 25% of builders risk claims to internal water system failures, and Nationwide's 2024 construction report has median water claim costs up 21% year over year.

Most of those losses never reach a claim. They land below the deductible, so the general contractor absorbs the rework, the cleanup and the schedule slip out of pocket. On a data center the stakes are steeper: the same event that costs a rework week on an office job can wet switchgear with lead times measured in months.

Remediating water damage during construction typically costs 5 to 10 times less than the same failure after occupancy. And a data center's handover is not one date: phased turnovers put operating halls next to active construction, and every seam between phases is a seam in responsibility for water.

What a data center leak detection spec should require

The MEP specs we price against converge on the same requirements. Writing or reviewing one, this is the checklist:

• Continuous rope sensing in every drip tray, covering the full length of the tray rather than spot probes at intervals. Leaks travel along trays and show up between points.
• Leak location, not just leak presence. Long runs should be sectioned or addressable so the alert names the branch. On headers that run hundreds of feet, 10 ft resolution is achievable and worth specifying.
• Nuisance-alarm immunity. The sensor must alarm on standing water, not on humidity, condensation or dust. A hall that condenses around the clock will otherwise bury the operations team in false alarms.
• Coolant compatibility. Sensors verified against the actual loop fluid, including propylene glycol blends, not just city water.
• Serviceability. Sensors field-replaceable without dismantling the drip tray.
• BMS integration. Relay outputs and an API so monitored points surface in the building management system alongside everything else the operations team watches.

Everything above is a detection spec, and detection is the easy half. Most water damage is not only a detection problem. It is a response problem. An alarm that lands in an unstaffed inbox at 2 AM protects nothing.

The response half of the spec matters more: 24/7 human monitoring with phone escalation, per-line response rules, commissioning with baseline learning and a documented report after every confirmed event. The metric to hold any vendor to is time to verification, the gap between an alert and a confirmed action. Per line, that means automatic shutoff on the domestic main, after-hours lockdown policies and alert-only metering on the sprinkler line. Life safety is monitored, never isolated.

From turnover to steady state

The strongest version of this program starts with the build and never leaves. A course-of-construction deployment meters the temporary supply from the first day water is live, covers the domestic main and puts wireless sensors in mechanical and electrical rooms. As phased turnovers complete, the same hardware, network and monitoring center convert to the operating program, with the flow baseline already learned and the incident history already documented.

That documentation carries weight beyond operations. Aon's 2023 whitepaper on the Eddy system found leak detection would positively affect 58% of water-loss claims, and monitored buildings support renewals with premium savings of up to 20% and deductible reductions of up to $150K. For a facility whose insurance file already carries heavy business-interruption exposure, a documented water-response record is underwriting evidence, not a brochure claim.

The short version

Treat water in a data center like the operational risk it is. Put continuous, located, glycol-verified rope sensing in the trays. Meter every line, including the ones you will never auto-close. Configure response per system, and put a human escalation chain behind the alarms. Start during construction, and make the program convert at turnover instead of starting over.

Eddy runs this as one managed program: detect, verify, isolate, resolve. See how it maps to your facility on the data centers page, or talk to a leak-detection advisor about a specific build.