Power Magazine flags grid bottleneck

Data centers are turning into power projects with servers attached. That is the real shift in the new POWER Magazine reporting from May 1 — not that AI needs lots of electricity, which everyone already knew, but that grid access is now deciding who gets to build, where, and on what timetable. In other words, the bottleneck has moved downstream. Chips still matter, but megawatts, substations, transmission headroom, and interconnection approvals are starting to set the pace. ### What changed here? The fresh thing is how explicitly the power trade press is framing the constraint. POWER’s two May 1 pieces say the first question on many large AI projects is no longer “what GPUs can we buy?” but “where can we get enough reliable power, how fast, and under what conditions?” That sounds subtle, but it changes the whole build logic. Developers used to start with fiber routes, metro proximity, and tax breaks. Now they are often starting with firm generation, transmission capacity, water, and a believable path to energization — then fitting the campus around those limits. ### Why is the grid the choke point? Because hyperscale campuses have gotten enormous, fast. POWER describes 300 MW to 600 MW sites as part of the normal development conversation now — basically mid-size-city loads arriving on data-center timelines, not utility timelines. A 400 MW campus can need a dedicated high-voltage substation and multiple 230-kV or 345-kV feeds. That is not something you improvise after the land deal closes. Utilities can add generation, transformers, and transmission, but those are multi-year projects with permitting and equipment delays. The mismatch is simple — AI campuses want months, the grid often needs years. ### Why did AI make this worse? Rack density jumped. Traditional racks often ran around 5 kW to 10 kW. AI-oriented facilities are pushing 50 kW to 100 kW per rack, which forces changes in cooling, distribution, and monitoring. So the problem is not just “more buildings.” It is more power concentrated in less space, with less tolerance for interruption. That pushes campuses toward liquid cooling, heavier electrical infrastructure, and tighter reliability requirements. Basically, the compute boom is arriving in a form the existing grid was not planned around. ### Is this just a Virginia story? No, but Virginia shows the shape of it. Northern Virginia already exceeds 3 GW of data-center load, and Harvard’s Belfer Center notes a July 2024 voltage event there disconnected 60 data centers at once, creating a 1,500 MW power surplus that required emergency action to avoid broader instability. That incident became a kind of warning label — when huge, tightly clustered loads trip together, the grid does not just absorb it gracefully. Texas is seeing the same pressure from the other side, with ERCOT flagging large-load growth and the state changing rules for very large customers through SB 6 in June 2025. ### What are regulators doing? They are moving from cheerleading to traffic control. FERC opened a rulemaking docket on large-load interconnections after DOE pushed for action, with the proceeding focused on loads above 20 MW — squarely in data-center territory. The point is to make interconnection more timely and orderly, but also to sort out cost responsibility, queue treatment, and co-located generation. States are moving too. Texas’s SB 6 tightened planning and cost rules for giant loads, and POWER says broader state-level intervention is becoming more likely as reliability and ratepayer concerns rise. ### So what does this mean for AI buildouts? It means power realism is now part of product strategy. If a site cannot get energized on time, GPU deliveries, construction sequencing, and launch plans all have to bend around that fact. Some developers will chase regions with spare capacity. Some will pair campuses with on-site generation. Some projects will slip. And some announced campuses will turn out to be more like options than near-term builds. ### What is the bottom line? The AI race still looks like a software-and-silicon story from far away. Up close, it increasingly looks like transmission planning, substation engineering, and utility regulation. POWER’s warning matters because it captures that turn clearly: the limiting reagent is no longer just compute. It is electricity that can actually show up, at scale, on time.