- AI data centers are hitting a power wall, but Airsys argues operators should focus less on PUE and more on a new metric
- Power compute effectiveness, or PCE, could help data centers unlock stranded capacity
- Higher chip temperatures, dry coolers and smarter retrofits could boost AI compute capacity without requiring more power
Data center operators are scrambling to solve the power equation to meet a wave of AI demand for new compute. But demand, Airsys Chief Strategic Relations Officer Paul Quigley argued the industry is looking at the wrong metric.
Rather than looking at how efficiently data centers use power, they need to be looking at how much of the power allocated to a facility is actually going into servers, he said.
That gap is the difference between the traditional yardstick of power usage effectiveness (PUE) and the new power compute effectiveness (PCE) benchmark that is rapidly gaining traction across the industry. Developed by Airsys, PCE has garnered the attention of the Uptime Institute and Open Compute Project, as well as rival vendors like Schneider Electric. The metric also has potential utility for telcos.
“Physics never separated power and cooling, businesses did,” Quigley said. “Data centers bring it all back together.”
Quigley noted that data centers, by design, are provisioned for their highest-demand operating conditions, even if those rarely occur. That means a data center provisioned with 100 MW of power may have to reserve a large portion of that power for cooling capacity only needed on peak days. Quigley likened it to a restaurant holding every table for a crowd that shows up once a year.
In a world where power has become a key bottleneck for AI factories, that kind of gap has become untenable. And it is particularly egregious in brownfield data centers.
Quigley said many older data centers may only deliver 50% to 60% of their allocated power to compute because they were designed around legacy cooling assumptions. PCE calculations – coupled with retrofitting projects – could help those facilities capture capacity they’re leaving on the table.
“As AI deployments scale, sustainability is no longer defined by reducing energy use—it’s defined by how efficiently that energy is converted into compute. Metrics like tokens per watt and cost per token are becoming more meaningful, shifting the focus from consumption to output,” Schneider Electric noted in a recent blog about PCE.
A hot new opportunity
Earlier this year, Nvidia CEO Jensen Huang noted its next-generation Vera Rubin systems can run at 45 degrees Celsius, marking a stark departure from historic assumptions that chips needed to be cooled to much lower temperatures. This temperature shift might seem small, but it has radical implications for data center design and efficiency. It’s also a perfect example of where PCE applied in the real world.
As Fierce noted at the time, the ability to use high-temperature water for cooling (it’s a bit counterintuitive, we know) eliminates the need for chillers and their power-hungry compressors. This, in turn, enables a shift to dry cooler systems, which use ambient air to cool the water running through their pipes.
Quigley argued that retrofitting existing sites with higher-temperature loops and dry coolers could free up power previously reserved for chillers — potentially allowing a facility to increase compute capacity without increasing its total power allocation.
But Airsys isn’t arguing chillers disappear. Rather, Quigley said its vision of the future includes a split thermal architecture in data centers that includes three cooling loops rather than a single loop for the whole facility. These tiers would comprise a high-temperature loop for chips, a mid-temperature loop for everything outside the rack and a low-termperature loop for loads (like UPS and BESS systems) that still require chillers.
He argued breaking out different cooling loads will become even more critical as chip operating temperatures increase in the coming years. Quigley said operating temperatures are currently limited by manufacturers rather than the silicon itself and predicted they will increase from 45C to 60C in the coming years.
In that kind of scenario, it won’t make any sense at all to run power-hungry chillers for chips when they’re not needed.
The telco angle
The PEC pitch is also beginning to resonate beyond data center operators. As AI moves to the edge, Quigley said PEC can serve as a tool for operators evaluating their existing facilities before pursuing costly rebuilds or new sites.
“Latency won’t be tolerated with a lot of where technology is going,” he concluded. “Somebody’s gonna own it, and whoever decides today that they need a use case before they jump in is already left behind.”
Read more about data center power and cooling issues here:
Trane buys LiquidStack to bolster AI data center cooling portfolio