Data centers hit power bottlenecks

AI data centers are starting to look less like a software story and more like an electricity story. The bottleneck is no longer just GPUs, financing, or land. It’s whether a developer can get enough power to a site — soon enough, cheaply enough, and without tripping over the local grid. That’s why this has become the real constraint on AI expansion in 2026: demand is exploding, but the wires, substations, turbines, and approvals are not moving at AI speed. (iea.org) ### Why is power suddenly the limiter? AI training clusters and inference farms draw far more electricity than older cloud workloads, and they want it in concentrated bursts. The IEA said data-center electricity demand rose 17% in 2025, with AI-focused facilities growing even faster, and it expects total data-center consumption to double by 2030 while AI-focused demand triples. That sounds abstract, but it means the industry is trying to add giant new loads to grids that were planned years in advance. (iea.org) ### Why can’t utilities just add more power? Because grid buildout is slow in all the boring, physical ways. Interconnection queues are long. Transformers and gas turbines are tight. Permitting is slow. Transmission upgrades take years. The IEA says those supply-chain and regulatory bottlenecks are now directly holding up grid connections and approvals for new facilities. Basically, the compute industry can order servers in quarters, but the power industry often builds in multi-year cycles. (iea.org) ### What changed this year? The language got blunter. Bloom Energy’s 2026 data-center power report says power availability has moved from a planning issue to a “defining boundary” on growth. Its survey work points to capital shifting toward power-advantaged regions, widening gaps between what developers expect and what utilities can deliver, and a bigger role for onsite generation as more than a temporary patch. That matters because it suggests this is no longer a temporary crunch — it’s becoming the operating model. (bloomenergy.com) ### Why are AI racks making this worse? A single rack is becoming a small power plant. NVIDIA says next-generation AI infrastructure is moving toward 1 MW IT racks starting in 2027, and that the old 54-volt approach simply does not scale cleanly to that level. Once a rack gets that dense, power delivery inside the building becomes a design problem too — copper, heat, conversion losses, cooling, all of it. So the bottleneck isn’t only “Can the grid serve the campus?” It’s also “Can the building distribute that power efficiently?” (developer.nvidia.com) ### Where does this hit first? It hits the biggest hubs first, because that’s where the queue is already crowded. Northern Virginia is the clearest warning sign. A July 2024 voltage event there disconnected 60 data centers at once and created a sudden 1,500 MW surplus that required emergency action. That episode didn’t prove data centers are uniquely fragile, but it did show how concentrated these loads have become — and how grid reliability starts to look different when dozens of giant facilities move together. (belfercenter.org) ### So what are companies doing? They’re going around the bottleneck where they can. The IEA says developers are pushing more onsite gas generation in the U.S., and the tech sector signed about 40% of all corporate renewable PPAs in 2025. It also says conditional offtake deals tied to small modular reactor projects grew from 25 GW at the end of 2024 to 45 GW by April 2026. In plain English — if the grid won’t show up fast enough, companies are trying to bring their own power. (iea.org) ### Does that solve the problem? Not really. It buys time, but it creates new tradeoffs — higher capex, fuel risk, emissions questions, and more uncertainty about who pays for oversized grid upgrades if projected demand never fully arrives. Harvard’s Belfer Center notes that utilities and regulators could end up stuck with stranded costs, while consumers face affordability and reliability concerns if planning goes wrong. The catch is that everyone agrees more power is needed, but nobody wants to be the one left holding the bill for the wrong forecast. (belfercenter.org) ### What’s the bottom line? AI is still scaling, but the industry has run headfirst into physics and permitting. The winners won’t just be the companies with the best models or the most GPUs. They’ll be the ones that can secure megawatts, navigate interconnection queues, and redesign data centers around power as the first constraint, not an afterthought. (iea.org)