Flexible Film Detects Sub-5 µL Liquid-Cooling Leaks in Under a Second

A printed, flexible sensor film that claims detection of under 5 µL of coolant in under one second — independent of fluid conductivity — is Serinity's answer to a leak problem the liquid-cooling market has largely measured only by guesswork. The startup, led by CEO Joshua Windmiller, is targeting the protected zones beneath racks, along CDU rows, and in piping galleries of AI data centers where incumbent sensors struggle.

The timing matters. Liquid cooling hit a $3 billion market in 2025 and Dell'Oro projects $7 billion by 2029. "A coolant leak can lead to anything from a single failed component to an unplanned data-center outage," Windmiller told Electronic Design. Beyond the data center, the same platform could serve EV-battery thermal management, oil-filled power transformers, and semiconductor fabs.

Why leaks are getting harder to catch

Connection counts are climbing while coolant chemistry is shifting away from conductivity. An NVL72-class rack carries roughly 125 to 270 quick-disconnect pairs depending on generation, and connection counts roughly double from NVIDIA's GB200 to GB300 as shared cold plates give way to per-device plates.

No field leak-rate data exists for data-center liquid cooling, so Serinity estimates it from process-safety databases (IOGP and PLOFAM), which put leak frequencies for flanged joints at 5×10⁻⁶ to 1.3×10⁻⁵ per joint-year. Applied to a GB300 rack's connection count, that implies roughly one leak event every two months across a 1,000-rack deployment at the central estimate — with a wide uncertainty band.

The first-order cost is downtime, not hardware. A leak caught early means a scheduled isolation of one rack; caught late, it becomes an unplanned cluster shutdown where lost compute revenue on GB200-class infrastructure usually exceeds the cost of the hardware. In tests of direct-liquid-cooled servers cited by Windmiller, the CPU throttled within 23 seconds of a total pump failure at full load.

What's trending up and down in cooling

The industry is consolidating around single-phase direct-to-chip (DTC) cold plates. Lenovo's GB300 NVL72 guide puts liquid at roughly 90% of rack heat capture at a 135 kW TDP, and NVIDIA describes Vera Rubin NVL72 as liquid-cooled down to every chip, with no fans. Supply temperatures are rising too — Vera Rubin is designed for roughly 45°C inlet and 55°C return water, largely eliminating chillers. Google's Project Deschutes CDU specification is rated to 2 MW.

Trending up next: two-phase DTC, which needs roughly 1/5 the coolant flow per kilowatt and makes sense as GPU TDPs push past 1.5 kW. Microsoft's lab tests of microfluidic on-chip cooling report up to three times the heat removal of cold plates and a 65% cut in peak temperature rise. Trending down: two-phase immersion, stalled by 3M's exit from PFAS manufacturing and the loss of Novec-class fluids, and rear-door heat exchangers, which top out around 70 kW per rack.

Where today's sensors fall short

Conductive sensing cable — two conductors under a porous jacket — requires liquid to pool at the cable and the fluid to be conductive. That fails with deionized water, now the recommended coolant in some OEM guides, and with hydrocarbon, silicone-oil, and dielectric fluids. Spot "puck" sensors monitor single points; optical prism sensors are fast and fluid-agnostic but need leak locations predicted in advance. NVIDIA's GB200/GB300 architecture stacks three tiers — tray-level sensors read through the BMC, rack-level ropes and spot sensors, and CDU sensors — yet microliter-per-minute leaks stay buried in flow and pressure noise. What's missing, Windmiller argues, is continuous area coverage at microliter sensitivity that discriminates coolant from condensation.

An impedance spectrum, not a resistance reading

Serinity's film performs time-domain electrochemical impedance spectroscopy (EIS), resolving each fluid's frequency-dependent permittivity rather than a single resistance value. The approach descends from transcutaneous biosensing of pico-ampere signals. Printed roll-to-roll on a flexible substrate, the film conforms to tray floors, manifold runs, and underfloor zones without drip pans.

The signal chain is the hard part. A 5 µL droplet disturbs only a tiny fraction of a large-area surface, so the front end's input bias current must sit in the low picoampere to femtoampere range. Printed traces carry sheet resistance orders of magnitude above copper, so a long film behaves like a distributed RC line. Data-center EMI from busbars and pumps demands guarded, shield-driven measurements. Battery-operated wireless mesh nodes — running three to five years on a battery by sleeping almost constantly — pair the film with a precision analog front end and an MCU running pattern recognition. Cloud analytics add an API for anomaly detection, leak localization, and DCIM integration.

Classification, not thresholding, separates coolant from condensation or a cleaning spill: an algorithm weighs waveform features against an internal library, with a reference channel and AI/ML models compensating for temperature, humidity, and aging.

All performance figures so far are vendor claims without independent verification. The open question is whether hyperscalers will embed such films directly into rack-level cooling hardware — Serinity's stated roadmap with OEMs — before connection counts and DI-water chemistry outpace the incumbent sensor base.