Data centers hit power bottleneck

Data centers have a new bottleneck, and it is not chips. It is electricity — getting enough of it, in the right place, on the right timeline. That sounds obvious, but the shift matters because it changes what counts as progress on an AI build. This week, industry coverage from POWER Magazine and Data Center Knowledge made the same point from different angles: power availability is now deciding where projects go, how fast they move, and sometimes whether they happen at all. (powermag.com) ### Why is power suddenly the hard part? AI facilities are much hungrier than the older cloud buildings people got used to. POWER Magazine says hyperscale campuses in the 300 to 600 MW range are now part of normal planning conversations, and rack densities in AI environments are running around 50 to 100 kW per rack instea(powermag.com)smission access, substations, step-down transformers, switchgear, backup systems, and utility commitments that can support a small city’s worth of load. (powermag.com) ### What changed this week? The notable thing is the convergence. POWER Magazine framed power as the “binding constraint” in hyperscale development, while a separate POWER commentary said utilities’ long-term expansion plans are being outpaced by data center requirements. Data Center Knowledge pushed the same idea from the (powermag.com)capacity. When multiple corners of the industry start saying the same thing at once, that is usually the signal that a background issue has become the main one. (powermag.com) ### Why can’t utilities just hook them up? Because the queue is not one wire and one switch. A big AI campus often needs dedicated high-voltage substations and multiple 230-kV or 345-kV transmission feeds, plus N-1 reliability so a single equipment failure does not take the site down. Those assets take time to permit, eng(powermag.com)data centers could rise past 9,000 units by 2030 from roughly 1,500 today. (powermag.com) ### How big is the demand shock? Very big. Data Center Knowledge says US data center capacity could scale from about 24 GW in 2026 to 100 GW by 2030, and that data centers may account for roughly 68% of US load growth through 2030. One speaker at Data Center World put the broader US challenge even more starkly: about 200 GW(powermag.com) estimate is directional rather than precise, the point lands — demand is rising faster than the system was built to absorb. (datacenterknowledge.com) ### So what are developers doing instead? They are moving toward “speed to power.” Basically, if the grid cannot deliver fast enough, owners look for behind-the-meter generation, fuel cells, storage, or sites with existing industrial infrastructure and better utility relationships. POWER Magazine says on-site generation is b(datacenterknowledge.com)less for square footage and more for how quickly they can unlock electricity. (powermag.com) ### Does this change where data centers get built? Yes — and probably more than any server roadmap does. SmartBrief notes that the debate has shifted from how many data centers to build to who will power them and who will allow them. That means the winning regions are not just cheap-land regions. They are places with tran(powermag.com)eed it. (smartbrief.com) ### Why does this matter beyond tech? Because the constraint spills into the whole electrical system. Transformers, switchgear, feeders, and substations are not data-center-only products. If hyperscalers become dominant buyers, they can reshape production capacity, pricing, and lead times for everyone else. That is why this no longer looks like a niche cloud story. It looks like an infrastructure story with AI attached. (datacenterknowledge.com) ### Bottom line? The old critical path was pouring concrete and filling halls with servers. The new one is energization. In 2026, the scarce input for AI buildouts is not ambition. It is grid-ready power. (powermag.com)