Schneider pushes liquid cooling metrics

Schneider Electric used a May 14 blog post to argue that liquid cooling is changing how data-center operators should measure efficiency in AI facilities. The company said power usage effectiveness, or PUE, no longer captures enough of the picture when high-density AI racks push sites against power and cooling limits. Christopher Leonard, writing on Schneider’s data-center blog, said operators should add power-chain efficiency, water usage effectiveness and workload measures such as “tokens per watt” and cost per token. The post frames the issue as a facility question as much as a server question. Schneider said AI planning increasingly depends on how much provisioned power reaches compute, how much water cooling systems consume and how much useful output a site can produce from fixed electrical capacity. ### Which metric is Schneider saying no longer tells the whole story? PUE remains the industry’s standard shorthand for data-center efficiency, and Schneider said it measures the ratio between total facility energy and IT equipment energy. (blog.se.com) Leonard wrote that the metric “only captures energy overhead” and does not show how effectively a site turns available power into AI computation. The May 14 post says that gap matters more as liquid cooling becomes common in AI deployments. (blog.se.com) Schneider wrote that liquid cooling is becoming the default for AI infrastructure and said the relevant question is shifting from how much energy a facility uses to how much compute that energy produces. ### What new measurements did Schneider put forward? Power Compute Effectiveness, or PCE, is the foundation Schneider highlighted in the post. (blog.se.com) Leonard defined PCE as a measure of how much available power is doing useful computational work rather than being stranded, underused or lost in system inefficiencies. Two outcome measures sit on top of that, according to Schneider: tokens per watt and cost per token. The company said tokens per watt measures how much AI work is generated per unit of energy, while cost per token measures how efficiently that energy becomes usable output or revenue. (blog.se.com) Water usage effectiveness also enters the calculation in Schneider’s framing. The post says liquid cooling changes the operating trade-offs, making water consumption and the design of the cooling loop part of the efficiency discussion alongside electrical losses. (blog.se.com) ### Why is liquid cooling central to this argument? Goldman Sachs was cited in Schneider’s post as estimating that 76% of AI servers deployed by the end of 2026 will be liquid-cooled. (blog.se.com) Schneider used that figure to support its case that the industry is moving into a regime where rack density and thermal design can no longer be treated as secondary engineering details. A separate Schneider blog post published on April 15 tied that cooling shift to changes in power architecture. (blog.se.com) Stuart Sheehan and Tuan Hoang wrote that 800 VDC distribution is aimed at racks approaching 400 kW and beyond, with AC-to-DC conversion moving out of the IT rack and into sidecar power systems. ### How does this change data-center planning? Capacity constraints are the operational backdrop in Schneider’s argument. (blog.se.com) Leonard wrote that power is finite, costs are rising and inefficiencies compound quickly as AI workloads scale, making it more important to know how much useful compute a site can deliver within fixed utility and cooling limits. That means site selection and expansion decisions increasingly depend on power delivery and cooling infrastructure, not only on server specifications. (blog.se.com) Schneider said higher tokens per watt can translate into more compute within fixed power limits, while lower cost per token can make AI scaling more predictable. That is Schneider’s interpretation of how operators should evaluate capacity, and it is stated in the company’s post. (blog.se.com) ### Where does Schneider place this within its broader product push? Schneider has published a string of data-center posts this year around liquid cooling, brownfield AI retrofits and high-density cooling distribution. A February 26 post by Andrew Whitmore, vice president of sales at Motivair by Schneider Electric, argued that existing facilities can be modernized for AI with liquid cooling rather than rebuilt from scratch. The company’s data-center blog listed additional May 2026 pieces on direct-to-chip liquid cooling and NVIDIA-linked AI factory infrastructure. (blog.se.com) Those posts place the efficiency argument alongside Schneider’s broader effort to position cooling, power distribution and facility design as a combined offer for AI data centers. May 2026 entries on Schneider’s data-center site show the company continuing that campaign with posts on direct-to-chip cooling, liquid-cooling choices and AI-factory infrastructure. (blog.se.com) The May 14 article by Christopher Leonard remains available on Schneider’s blog, where the company laid out the metrics it says operators should track next. (blog.se.com 1) (blog.se.com 2)