PEoperator: 30–40% data‑center capex

A 1 GW AI data center is starting to look less like a warehouse full of servers and more like a private power system with compute attached. That is the real story behind the recent PEoperator post tying electrical infrastructure to roughly 30–40% of total capex. The exact share will move by design, density, and how much utility work sits inside the project boundary. But the broader point holds — once campuses get this large, the electrical stack stops being a supporting line item and becomes the project. ### What changed in the conversation? The trigger was David Sacks’ $50 billion figure for bringing 1 GW of capacity online. That number has been circulating in AI infrastructure discussions for months, and it is now getting paired with a more grounded construction question: if the whole campus costs around $50 billion, how much of that is tied up in substations, transformers, switchgear, UPS systems, generators, cabling, and commissioning? (datacenterdynamics.com) That is the gap PEoperator was pointing at. ### Why does the electrical share get so big? Because a gigawatt is enormous. At that scale, you are not just feeding racks. You are stepping down utility power across multiple voltage levels, building redundancy into every path, backing up critical loads, and protecting everything with layers of controls and distribution gear. Alvarez & Marsal’s data-center breakdown explicitly groups substations, transformers, switchgear, UPS, PDUs, and generators as essential operating infrastructure — and on AI-heavy campuses that package grows fast. (tradersunion.com) ### Is 30–40% a crazy number? Not really. It is hard to pin down a universal percentage because public cost breakdowns are all over the place and many mix IT hardware with shell-and-core construction differently. But McKinsey’s 2030 view is directionally useful — more than $4 trillion of almost $7 trillion in global data-center capex goes to computing hardware, leaving a massive remainder for power infrastructure, real estate, cooling, and everything else. (alvarezandmarsal.com) In other words, even if chips dominate, the non-chip share is still gigantic in absolute dollars. ### So what sits inside “electrical”? The boring stuff that turns out not to be boring at all. Think utility interconnection, on-site substations, medium-voltage distribution, transformers, switchgear lineups, busway, backup generation, battery systems, protection relays, and the controls needed to make all of it fail safely instead of catastrophically. On a 1 GW campus, that starts to resemble grid infrastructure more than traditional commercial construction. (mckinsey.com) ### Why are developers suddenly obsessed with this? Because the lead times are ugly. Wood Mackenzie said power-transformer lead times were still around 128 weeks in Q2 2025, with GSUs around 144 weeks, and switchgear around 44 weeks. Separate industry coverage in 2026 says shortages of transformers, switchgear, and backup components are delaying a growing share of U.S. data-center projects, with only about one-third of planned 2026 capacity under active construction. (alvarezandmarsal.com) ### Why does commissioning matter so much here? Because buying the gear is only half the trick. A gigawatt campus has thousands of failure points across utility feeds, protection settings, backup systems, and control logic. If procurement, electrical design, and commissioning are not integrated early, the project can be physically built but still not ready to energize. That is the catch — the last mile is not cosmetic. It decides when revenue starts. (eepower.com) ### Does this change who wins? Yes. It favors developers, EPCs, and operators who can lock in electrical gear early, standardize designs, and sometimes pre-buy or vertically integrate around scarce components. Recent reporting says some builders are already stockpiling equipment or leaning into modular approaches to dodge bottlenecks. That is a very different competitive edge from simply getting access to GPUs. (rcrtech.com) ### Bottom line? The AI buildout is still a chip story — but at 1 GW scale, it is just as much a transformer-and-switchgear story. If the industry really is going to spend tens of billions per campus, the constraint is moving from silicon alone to the electrical spine that makes the silicon usable. (datacenterdynamics.com) (rcrtech.com)