OCP updates open AI specs

AI data centers are becoming facility problems, not just server problems. That is the real shift here. The Open Compute Project — OCP — has been pushing deeper into “open” designs for the building itself: power rooms, liquid-cooling loops, row layouts, and the cluster shapes that sit around the racks. Last week that work got more concrete, with OCP surfacing its Open Data Centers for AI initiative and pairing it with updated design guidance aimed at 2026 and 2027 deployments. ### What actually changed? The headline change is not one single product spec. It is a bundle. OCP now has an umbrella effort called Open Data Centers for AI, and inside it are reference designs for facility layouts, cooling distribution, power delivery, telemetry, and operating agreements between IT operators and facility operators. The goal is to make AI sites less custom and more repeatable. ### Why is that a big deal? Because AI racks are getting absurdly dense. OCP’s own materials talk about roadmaps pointing toward 1 MW racks in the next few years, plus GW-scale facilities and even multi-GW campuses. At that point, the old model breaks. You cannot just drop hotter servers into a normal hall and call it a day. Cooling, power distribution, and building layout become the bottleneck. ### What is in the new spec? The clearest document is OCP’s Open Data Center Spec Version 0.5.0, dated February 2026. It says it updated minimum row length, row-quantity guidance, row power-density requirements, indoor humidity guidance, and liquid-to-air ratio assumptions. It also includes sections for CDU and sidecar specifications and layout examples at both cluster and campus scale. In plain English, just how to mount servers. ### Why does liquid cooling keep showing up? Because air is running out of room. OCP’s cooling work has been moving in this direction for a while, with rack-level blind-mate liquid interfaces, manifold specs, and Google’s Project Deschutes CDU contribution all feeding the same ecosystem. The idea is to standardize the plumbing the way earlier OCP work standardized racks and power shelves. ### What about power? Power is the other half of the story. OCP’s Open Data Centers for AI materials explicitly call out higher-voltage LVDC, power estimation, and grid interconnect as core priorities. The community is basically admitting that AI buildouts now hit the utility edge as much as the server edge. If cooling is the heat problem, power is the time-to-deploy problem. ### Is this only for hyperscalers? No — but hyperscalers are clearly driving it. OCP says the work is meant to span hyperscale, neocloud, colocation, enterprise, and smaller deployments. The cluster-design initiative makes the same point, saying today’s AI systems are too often one-offs and that the new designs should be detailed enough to support procurement and deployment across different scales. ### So what is the real point? Basically, OCP wants to do for AI facilities what it once did for servers and racks — turn tribal knowledge into shared building blocks. That does not mean every operator will build identical AI halls. But it does mean fewer bespoke decisions around cooling loops, sidecars, row geometry, and facility interfaces. That can shorten design cycles, broaden the supplier base, and make capacity easier to replicate. ### Bottom line? This is infrastructure standardization disguised as plumbing. The flashy part of AI is chips and models, but the gating factor is increasingly the building wrapped around them. OCP’s update matters because it pushes “open” AI beyond the rack and into the facility itself — right where the next deployment bottlenecks are forming.