Air cooling was designed for a world of 8 to 12 kW racks, but that world no longer exists. Today’s AI training racks routinely draw more than 100 kW, with roadmaps pushing toward 250 kW and beyond, and no amount of moving air can pull that much heat out of that small a footprint. That’s why liquid cooling has moved from a niche technique to the backbone of AI infrastructure almost overnight, with liquid-cooled server deployments more than tripling in a single year. But as coolant has moved from the room’s perimeter deep into the rack itself, running through manifolds, cold plates, and dozens of wetted quick-disconnect points inches from live electronics, it has quietly changed the risk profile of the entire data center.
VMblog sat down with Joe Arena, Chief Commercial Officer at Parameter Technologies, to unpack what that shift means for designers, operators, and OEMs building the next generation of AI infrastructure. Arena explains why a single high-end AI training rack can represent upward of $4 million in compute hardware, why leak detection can no longer be an afterthought bolted on at the end of a project, and what it actually takes for a detection system to keep pace with racks that are denser, more expensive, and more tightly integrated than ever before. Read on for his take on rack-level resolution, modular design, cybersecurity, and why protecting the fluid side of the hall deserves the same rigor as the cooling system itself.
++
VMblog: Why has liquid cooling become a foundational technology for modern AI data centers?
Joe Arena: Air cooling was built for racks pulling 8 to 12 kW, and for decades that was enough. But with a single AI training rack today drawing over 100 kW and roadmaps pointing toward 250 kW and beyond, air cooling simply cannot remove that much heat. No volume of moving air can extract that much heat from that footprint.
Liquid-cooled AI servers more than tripled from 2024 to 2025 to over half of those installed, and the share continues to climb. Market researchers have estimated that the liquid-cooling equipment market will nearly double from last year. Liquid cooling is where the industry already is for AI infrastructure.
VMblog: How has the architecture of liquid cooling systems changed over the past few years?
Arena: Not long ago, liquid cooling in most facilities was a rear-door heat exchanger and not much else. Now the coolant path runs deep into the rack. Coolant distribution units feed into manifolds, which feed cold plates bolted directly to the processors, with flexible hoses routing between them. Quick disconnects are installed at every service point, so a technician can pull a tray without draining a loop.
Everyone is aware that those interconnections are also places where fluid can escape. A modern high-density rack can carry dozens of wetted connections inside the enclosure, inches from live electronics. Liquid cooling has enormous cooling capacity. But it also shifted the risk from the room’s perimeter to the rack’s interior, and protection has to follow the fluid to where it actually is.
VMblog: Why should leak detection now be considered an integral part of the cooling system design rather than simply an accessory or “add-on” monitor?
Arena: Two reasons. The odds of a leak went up because there’s more fluid, more connections and more pressure inside the rack than ever before. And the cost of a leak went up because of what the fluid now sits next to.
A single high-end AI training rack today represents upward of $4 million in compute hardware. A few milliliters of coolant in the wrong place can take it down in the middle of a job that’s been running for weeks or even months, stalling thousands of synchronized GPUs and erasing the compute already sunk into the run. Also, coolant on energized electronics is a safety event as much as an availability one. When the asset is that valuable and the fluid is that close, detection can’t be a sensor someone remembers to add at the end. It belongs in the design of the cooling system itself, sized and placed with the same stringency as the manifolds and the CDU.
VMblog: What characteristics define a leak detection system that is truly designed for today’s liquid-cooled environments?
Arena: A room-level alarm telling you there’s liquid somewhere is almost useless when the floor holds hundreds of racks. Operators should expect leak detection that resolves to the individual rack and localizes to the nearest meter, so responders go straight to it.
The window between a drip and damage is short at these power densities, so the system has to flag and place a leak in seconds.
Then there’s sensing. It has to be installed without a forklift and with a two-week commissioning schedule. It has to scale as the hall grows without a re-architecture. It has to speak to the building management and DCIM platforms operators already run, over secure connections. And it has to trigger an automated response, isolating a zone or signaling a CDU, rather than lighting up a dashboard and waiting for a human to notice.
VMblog: How can leak detection be designed to simplify installation and future expansion?
Arena: Design it as modular, like the racks it protects. A detection platform should let you instrument the racks you have today and add more by clicking in another module, not by pulling the whole system apart. That keeps commissioning to hours rather than days because most of the work is connection.
The sensing cable should follow the existing trays and cable paths. This saves installation time and labor every time the layout changes. A hall that starts at 20 racks and grows to 200 shouldn’t need a second detection architecture at rack 65. Lifecycle costs drop because you’re extending one system instead of ripping and replacing.
VMblog: What role do system integration and cybersecurity now play in modern leak detection systems?
Arena: Leak detection used to be an island. These systems now sit on the same building management, SCADA and DCIM networks as everything else in the facility and that changes the security conversation entirely.
A sensor with a network port is an entry point. Any monitoring device connecting to operational networks has to meet the same security bar as the rest of the infrastructure, which means encrypted communications, modern industrial protocols and a design that satisfies the cybersecurity requirements that enterprise customers now write into their contracts. The easier a device is to fold into existing platforms through standard, secure protocols, the less tempted an operator is to bolt on an insecure workaround just to get data out of it.
VMblog: As AI infrastructure scales, what operational benefits come from designing leak detection into the project from the beginning?
Arena: A platform installed with the cooling is faster because it goes in as part of the build rather than as a retrofit around live equipment. It gives operators continuous visibility into the fluid side of the hall, which is a blind spot in many facilities today. When something does go wrong, technicians see exactly which rack and roughly where it is, so there are fewer nuisance trips, faster recoveries and less unplanned downtime. Detection is insurance against leaks and subsequent consequences. But considering metrics like lifecycle costs, leak detection is more than just insurance.
VMblog: What should designers, operators, and OEMs be thinking about as the next generation of liquid-cooled AI infrastructure is deployed?
Arena: The industry took liquid from a niche technique to the backbone of AI infrastructure, and the CDUs, cold plates and manifolds shipping today are sophisticated equipment. The protection around them hasn’t always kept pace.
Designers should treat leak detection as part of the cooling system’s specification, not a line item added if budget allows. Operators should ask for rack-level resolution and secure integration before signing. And OEMs building these platforms have a chance to make detection native to the rack rather than something the customer must handle. Do that, and the whole system gets more resilient, easier to run and ready to scale with workloads that show no sign of slowing. The money going into these facilities is enormous. What protects it should be built with the same sense of rigor.
##
