Power, HBM and cooling become bottlenecks

Artificial intelligence servers are running into a new limit: not just chips, but the electricity and cooling needed to keep them online. (datacenterknowledge.com) (iea.org) At Data Center World 2026, engineers from Oracle Cloud Infrastructure, Nvidia and Google described data centers as tightly integrated AI systems, with power architecture, cooling design and network layout all changing at once. (datacenterknowledge.com) The physics is simple: graphics processors draw more power than older servers, and more power turns into more heat. The International Energy Agency says servers account for about 60% of electricity demand in modern data centers, while cooling ranges from about 7% in efficient hyperscale sites to more than 30% in less-efficient enterprise facilities. (iea.org) That is pushing operators away from air cooling alone. Data Center World said access to power remains the top constraint in 2026, and 19% of surveyed operators already use liquid cooling, with more planning adoption over the next 12 to 24 months. (datacenterworld.com) Suppliers are reporting the same strain in their order books. Eaton said in February that Electrical Americas orders accelerated 16% on a rolling 12-month basis, driven by data-center momentum, while Electrical backlog across the company rose 29% year over year. (eaton.com) Vertiv, which sells power and thermal equipment for data centers, reported on April 22 that first-quarter 2026 diluted earnings per share rose 136% and it raised full-year guidance. The company’s investor materials still highlight a business built around powering and cooling digital infrastructure. (investors.vertiv.com) Memory is the other choke point. Micron said in its December 17, 2025 fiscal first-quarter call that its high-bandwidth memory supply for calendar 2026 was sold out and customer negotiations on 2026 volume and pricing were complete. (fool.com) High-bandwidth memory, or HBM, is the stack of very fast memory chips that sits next to AI processors and feeds them data. Without enough HBM, accelerator shipments can be delayed even if the graphics processor itself is available. (fool.com) Property and infrastructure advisers are now framing this as a site-selection problem as much as a semiconductor problem. JLL said in its 2026 Global Data Center Outlook that global capacity could nearly double by 2030, but the buildout will require energy innovation to ease grid constraints and about $3 trillion of investment. (jll.com) The result is that AI deployment timelines are increasingly set by transformers, switchgear, liquid loops and memory allocation, not just by how many accelerators cloud companies want to buy. (jll.com) (datacenterknowledge.com)