Flexible Film Sensors Target Microliter Leaks in Liquid-Cooled AI Racks

By Tom Whitfield

A single GB300-class rack carries 125 to 270 quick-disconnect pairs, and applied to that connection count, industry leak-frequency databases imply roughly one leak event every two months across a 1,000-rack AI deployment. That arithmetic comes from Serinity, a startup building flexible printed sensor films that detect liquid-cooling leaks of under 5 microliters in under one second, and it frames why the company believes leak detection is becoming a first-order data-center problem.

The market it is targeting is growing fast. Liquid-cooling revenue hit $3 billion in 2025 and is heading to $7 billion by 2029, according to Dell'Oro. Racks are going fully liquid, coolant temperatures are rising, and the industry is consolidating around single-phase direct-to-chip cooling, said Serinity CEO Joshua Windmiller, who previously co-founded Biolinq and held a senior sensing role at Dexcom.

Why are leaks getting harder to catch?

The geometry of modern AI halls works against conventional sensing. Single-phase cold plates dominate deployed capacity today, and liquid now captures roughly 90% of rack heat in Lenovo's GB300 NVL72 guidance at a 135 kW TDP. NVIDIA describes Vera Rubin NVL72 as liquid-cooled down to every chip and networking component, with no fans, running roughly 45°C inlet and 55°C return water, which largely eliminates chillers.

Two trends make detection harder. Connection counts roughly double from NVIDIA's GB200 to GB300 as shared cold plates give way to per-device plates. Coolant chemistry is also shifting: some OEM guides now recommend deionized water, and conductive sensor cables cannot detect it. Windmiller said newer fluids such as hydrocarbon and silicone oils, along with dielectric coolants, exhibit very poor conductivity and fail to register on incumbent conductometric approaches.

Two-phase immersion, by contrast, is trending down. 3M completed its exit from PFAS manufacturing about a year ago, and the loss of Novec-class fluids has stalled that segment's growth. Rear-door heat exchangers top out near 70 kW per rack, well below the 130 kW-plus racks now shipping.

How does Serinity's film work?

The sensor is a printed, flexible film, produced in roll-to-roll form, that turns any surface into a continuous detector. Instead of a single resistance reading, it performs time-domain electrochemical impedance spectroscopy, resolving each fluid's frequency-dependent permittivity. Because the film identifies fluids by electrochemical signature rather than conductivity alone, it works across water-glycol, deionized water, hydrocarbons and dielectric coolants, while distinguishing coolant from condensation.

The technology stems from transcutaneous biosensing of pico-ampere electrochemical signals, Windmiller said. The solution has three layers: the film itself; battery-powered wireless mesh nodes that run the measurement and relay alerts, running three to five years on a battery by spending nearly all of their time in deep sleep; and cloud analytics with an API for anomaly detection, leak localization and integration with existing data-center monitoring systems.

Deployment today targets protected zones beneath racks, along CDU rows, in piping galleries and at critical fitting junctures. The roadmap embeds the film directly into rack-level cooling hardware with cooling-system OEMs. The same platform could also serve semiconductor fabs, EV battery-cooling systems and oil-filled power transformers.

What did the signal chain demand?

Detecting and fingerprinting a few microliters of a poorly conductive fluid across a large area is fundamentally an electrochemistry problem, Windmiller said. The analog front end needs input bias current in the low picoampere to femtoampere range; anything larger perturbs the sensor and risks false alarms. Getting fast sampling, high resolution, a low noise floor, high input impedance and low power from one signal chain is, in his words, like squeezing a balloon: press on one requirement and another bulges out.

Printed traces add another wrinkle. Their sheet resistance sits orders of magnitude above copper, so a long film behaves like a distributed RC line whose response depends on where the fluid lands. Printed conductors do resist corrosion and oxidation, however. Electrical noise is the third constraint: data centers carry high-current busbars and motor-driven pumps, so the design uses guarded measurements and shields driven at equal potential to the excitation waveform.

A threshold alone cannot separate coolant from condensation or a cleaning spill. The system infers molecular composition from the acquired waveform and weighs its features against an internal library, with a reference channel and AI/ML models keeping calibration stable across time, temperature, humidity and exposure.

What is at stake commercially?

NVIDIA's GB200/GB300 architecture already defines three sensor tiers: tray-level cold-plate and inner-manifold sensors read through the BMC, rack-level ropes and spot sensors along piping, and sensors inside the CDU. Loop monitoring of flow, pressure and reservoir level catches gross failures, but early leaks often sit inside normal operating variation.

The first-order consequence of a missed leak is downtime, not hardware. Caught early, a leak is a scheduled isolation of one rack; caught late, it becomes an unplanned cluster shutdown, and on GB200-class infrastructure the lost compute revenue usually exceeds the cost of the hardware. Windmiller cited tests of direct-liquid-cooled servers in which the CPU throttled within 23 seconds of a total pump failure at full load. As rack densities climb past 130 kW and connection counts keep doubling, demand for fluid-agnostic, area-coverage sensing is set to grow with every new GPU generation deployed.