This claim comes as the gap between average rack density and AI-cluster density widens into what the company calls a “discontinuity” rather than a gradual shift.
Uptime Institute’s 2026 Global Data Center Survey of more than 800 operators puts modal rack density at 11kW, with the average (excluding facilities purpose-built above 30kW) at 7.8kW, up only 0.3kW on the year before. NVIDIA’s GB300 NVL72 reference architecture, by contrast, reaches up to 142kW per rack, with HPE specifying 132kW for the same system. Supply chain reporting places the next-generation VR200 rack between 190kW and 230kW, and NVIDIA has outlined a roughly 600kW figure for Rubin Ultra.
“The installed base is creeping along at roughly three tenths of a kilowatt a year while rack densities at the frontier are increasing dramatically,” said Paul Quigley, chief strategic relations officer at Airsys. “That is not an incremental shift but a discontinuity.”
Not all surveys agree on the scale. AFCOM’s 2026 report puts average rack density at 27kW, up from 16kW the year before, a materially different figure to Uptime’s. Quigley said the two surveys measure different populations, “but either way, the direction is the same.”
Constraint is allocation, not heat, Airsys argues
Airsys’s central argument is that the binding constraint on AI buildout is less about cooling capacity than about how provisioned power is allocated. The company uses a metric it calls Power Compute Effectiveness (PCE), provisioned power for compute divided by total provisioned power for the site, which it says moves from 0.73 to 0.95 when a facility converts to a two-loop architecture. That represents roughly 30% more available compute from the same grid connection, Quigley said, “with no new interconnect and no new site.”
The efficiency argument lands as Capacity has separately reported that Texas data centre demand could lift ERCOT’s peak load by 125%, and that fibre has overtaken power as data centres’ top constraint, both signals that grid capacity, rather than raw construction pace, is becoming the sector’s binding limit.
On liquid cooling adoption, Quigley said the technology is “already standard practice at the top of the density curve and still niche everywhere else.” There is, he said, a rack density at which air stops making engineering sense, and that threshold, roughly 40kW per rack at the point air is no longer a realistic choice, has been moving down steadily for three years. He also flagged agentic inference as a wildcard: unlike steady training loads, agentic workloads “behave more like city traffic… with none of them necessarily adhering to previous trends,” which could shift where the liquid-cooling threshold sits on a given site. In Europe, he noted, Germany’s energy efficiency legislation is setting mandatory heat reuse thresholds that pull design decisions toward liquid cooling ahead of the economics alone.
Airsys pushes for PCE reporting language in Liquid Cooling for AI Act
Airsys is engaging directly with congressional offices behind the Liquid Cooling for AI Act, including Senator Dave McCormick’s, and Quigley was in Washington in September for those conversations. The bill, introduced in November 2025 and still in committee following an April 2026 Energy Subcommittee hearing, would direct the GAO to assess liquid cooling R&D needs and have the DOE report recommendations to Congress.
Airsys wants power-allocation measurement written into that assessment. “Nothing in it asks the question that matters most to an operator, which is how much of the power that better cooling designs give back reaches compute,” Quigley said, drawing a comparison to how PUE became embedded in procurement standards over roughly a decade. “We do not have a decade.”
On bottlenecks, Quigley pointed away from the cold plate itself toward the supporting infrastructure: coolant distribution units, manifolds, hoses, quick disconnects and leak detection, alongside a lack of agreed interface standards and a workforce still built around air-cooling expertise. “Everyone is shopping for the engine, and nobody is shopping for the fuel system.”
Asked whether Airsys’s policy engagement amounts to a sales pitch, Quigley acknowledged the commercial overlap but argued the metric cuts against the company as much as for it: “If an Airsys cooling architecture does not increase the proportion of provisioned power available for compute, PCE will show it.”
He also pointed to Airsys’s advocacy for converting existing sites over building new ones as a position that runs against its own commercial interest, since greenfield builds represent a larger equipment opportunity than retrofit conversions.
Related coverage:
Metro Connect USA 2027
Metro Connect USA is the largest executive-level digital infrastructure event in the U.S. The only one of its kind, this 25-year-strong gathering is where decision makers come together to make deals happen.
