Feature Story

The Hidden Risk in the AI Buildout: Why Leak Detection Is the Next Frontier in Uptime

For twenty years, the first rule of data center design was simple: keep liquid away from the equipment. That rule is now obsolete.

Inside the buildings training the world’s AI models, coolant runs within inches of the most expensive silicon ever manufactured. A single NVIDIA GB200 NVL72 rack carries roughly $3 to $4 million in hardware, draws about as much power as a hundred American homes, and is plumbed with well over a hundred fluid connections. Multiply that across a hall full of racks and the picture comes into focus: the industry has deliberately introduced water into the one place it spent two decades keeping it out.

It had little choice. Modern AI accelerators generate more heat per square centimeter than moving air can carry away. Liquid cooling is no longer an efficiency play. It has become a functional requirement for large-scale AI deployment, with a market on pace to approach $7 billion by 2029. More than half of AI chips shipping this year are liquid-cooled.

What the industry has not yet fully reckoned with is what it traded away.

The stakes, in numbers

Cooling failure has climbed to become the second-leading cause of impactful data center outages, behind only power, accounting for 14 percent of incidents in Uptime Institute’s most recent global survey, up from 7 percent two years earlier. And outages have grown more expensive as the assets behind them have. For the second consecutive year, one in five operators reported a major outage costing more than $1 million.

Insurers noticed first, as they usually do. Allianz Commercial’s 2026 analysis of 221 data center claims worth more than €700 million found water damage to be the single most frequent cause of loss (ahead of fire and ahead of equipment breakdown). Swiss Re, drawing on FM Global’s fifteen-year review of data center losses, puts liquid-related damage at roughly 24 percent of total loss costs, second only to fire. Underwriters have begun writing liquid-cooling risk controls into new policies.

The uncomfortable part is that those figures largely predate the current generation of high-density AI racks.

Harder than it looks

A modern liquid-cooled rack is not one pipe. An NVL72-class rack contains on the order of 125 to 270 fluid connection pairs (cold plates, manifolds, quick-disconnects), and upcoming generations add more. The reliability engineering behind them is genuinely excellent; couplings are rated for thousands of mating cycles and near-zero spillage. But reliability compounds. A very small per-connection failure probability, multiplied across hundreds of connections per rack and thousands of racks per site, still produces a meaningful number of events per year. The industry is scaling the number of things that can leak faster than it is reducing the odds that any one of them does.

And the leaks that matter most are the ones hardest to see. A burst line is easy to spot. A slow weep at a manifold seal, perhaps a few microliters an hour, evaporating almost as fast as it appears, can persist for days beneath a rack before enough fluid pools to trip a conventional sensor. By then the damage is usually done.

The detection dilemma

Here is where operators get stuck.

The instinctive response to leak risk is caution: lower the alarm threshold, shut down at the first signal, and ask questions afterward. But precaution is not free. An idle GB200-class cluster represents enormous foregone revenue every minute it sits dark, especially at a time when compute demand is outstripping supply industry-wide. A detection system prone to false alarms does not eliminate cost. Rather, it relocates cost from hardware damage to lost compute revenue, with reputational damage layered on top. And lost revenue is frequently the larger number.

So operators face a choice that they otherwise should not have to make: absorb the cost of false alarms, or raise the threshold and accept the risk of missing a real leak. The right answer is not to pick a side. It is to make the choice unnecessary. This means detection is precise enough, and fast enough, that caution and uptime stop working against each other.

What’s deployed today, and what it misses

The current toolkit was largely designed for a different problem.

Conductive cable and spot sensors represent the incumbent approach. Two-wire lines run along rack bases or beneath raised floors, triggered when fluid completes a circuit. They work within three structural limits. (1) Fluid must physically reach the cable, often with help from drip pans and trays engineered to guide it there. (2) Enough fluid must accumulate to register, which means detecting pools rather than weeps. And (3) the fluid must conduct electricity. That last constraint matters more each year: manufacturers’ own documentation states plainly that these cables will not detect deionized water or dielectric coolants, which are increasingly being recommended for the newest rack designs.

Optical point sensors detect liquid at a specific location and respond quickly once wetted there. Their constraint is geometric: coverage is discrete, so someone has to correctly predict in advance where a leak will occur. In dense deployments, instrumenting every junction and manifold densely enough to close the gaps becomes cost-prohibitive, not to mention that optics deployed in real data halls contend with dust and debris that clean-room testing never surfaces.

Flow and pressure monitoring watches the loop for anomalies and reliably catches gross failures. Early-stage leaks perturb flow and pressure so slightly that they hide inside normal operating variation.

Tracer gas testing, such as helium or hydrogen sniffing, validates seals rigorously during manufacturing and commissioning. It is not a live-operation tool, and it says nothing about the connection that begins seeping in month seven.

AI and machine learning forecasting is the most discussed and least proven of the group. Indirect telemetry is no substitute for physical detection; a model inferring a leak from humidity and pressure signatures is working several steps removed from the event itself. As a complement, though, it is genuinely promising. Machine learning is well suited to act as a risk predictor that dynamically raises sampling rate or sensitivity at a particular node when conditions warrant it, thereby catching events earlier without flooding operators with false alarms.

What good looks like

The requirements follow directly from those failure modes. Detection should be continuous across an area rather than concentrated at points, so nobody has to guess where the leak will be. It should register microliters rather than milliliters (or more), and seconds rather than minutes (or longer). It should identify what the fluid actually is, namely distinguish coolant from condensation, and work across water, glycols, oils, and dielectric cooling fluids alike. And it should interface with the monitoring systems operators already run rather than demanding new ones.

That is the problem that San Diego-based Serinity was founded to solve: a printed, flexible sensing film providing continuous two-dimensional coverage, paired with wireless mesh nodes and analytics. Current sensors detect under five microliters, which is roughly a tenth of a droplet, in under a second, across every coolant chemistry currently used in the data center. The form factor matters as much as the sensitivity. A conformal film fits dimensionally constrained installations and retrofits without the drip pans and trays that conventional sensors depend on to bring the leak to them.

The mandate is already written

The buyer community codified this need itself. The Open Compute Project’s leak detection guidance addresses cold plates, manifolds, CDUs, and rack piping. NVIDIA’s GB200 and GB300 reference documentation defines three distinct levels of leak detection: tray, rack, and data center; automated power and coolant-valve response are also included. Insurers are pricing the exposure. The requirement is not in question.

What remains is meeting the requirement sufficiently to minimize costly downtime. AI infrastructure is fast becoming the backbone of the global economy, and it now runs on water. The technology protecting it should be as advanced as the technology it protects.

Return to: 2026 Feature Stories