The aging US grid
Seventy percent of the transformers are past twenty-five. Replacement and storm hardening already run $51 billion a year — whether or not a single new campus gets built.
Robin Millican keeps a short list of how old the American grid is. Roughly 70 percent of power transformers are more than twenty-five years old. So are about 70 percent of the transmission lines. About 60 percent of the circuit breakers have passed thirty. He said this on Columbia Energy Exchange in June, from the Center on Global Energy Policy, and he was not talking about AI.
Alice Yake, who used to be chief planning officer at Xcel, arrived at the same fact from a planner’s desk. Average asset life around sixty years. Average age across the United States around fifty-five. “Even without data centers, even without load growth,” she said, a significant portion of the kit is due to be replaced. The grid, in her telling, is less reliable than it was five years ago, and five years ago it was less reliable than it was ten years ago. Aging parts. A supply chain that cannot replace them fast enough.
Even without data centers, even without load growth.
Alice Yake, formerly Xcel, July 2026
Share of US grid kit past the cited age
Millican, Columbia Energy Exchange, June 2026. Cited on the aging-grid replacement chain.
Fifty-one billion a year, before anyone adds a watt
Replacement, plus the storms and the fires, has already become a habit. Annual distribution capex rose about 160 percent from 2003 to 2023, to nearly $51 billion. That is about 43 percent of what investor-owned utilities spend. Millican’s point: most of it is replacing and hardening what is already there, not expanding capacity. Doug Arent named the weather leg on the same podcast — hurricanes in Florida, wildfires in the west, the cost of burying lines.
Utilities earn their return on capital they put in the ground. The incentive is to build. The check is the regulator, not a load forecast. That is why this page is not an AI-load story. If the campuses disappoint, the poles still have a birthday. Ramez Naam put the same layer in different words: the poles and wires are the limit, and they have not become an exponential technology. In territories where data centers can be built, the regional monopoly still collects.
The bill of materials has already moved. Wood poles are up 50 percent since 2019. Wires and cables are up 150 percent. Arent ranks that equipment inflation ahead of load growth as a driver of retail electricity prices. Some of it is copper, steel, and labor — input-cost push, not a clean supplier-margin story. Eaton’s Q1 print put the US data-center backlog at 228 gigawatts. The Q2 call restated it to 307 — fifteen years at 2025 build rates, only about 20 percent near-term, most of it 2028 and later. Laguna’s 342 gigawatts of announced projects is a different measure. Do not flatten 228, 307, and 342. Electrical Americas backlog printed $15.175 billion at June 30, book-to-bill 1.3, data-center orders about plus 85 percent. That is the AI-cycle book. The $51 billion distribution habit is the number that survives if that cycle cools.
The steel in the building is moving the same way. Construction Dive, citing Skanska’s Tom Park on a September 10 webinar, has steel prices up 25 percent over the past year and lead times more than doubled to 55 weeks — data centers and semiconductor plants, mills and fabricators running full. Skanska’s own summer note put structural packages at 40 to 50 weeks in some markets and HVAC at 52. Nucor’s second-quarter slides name data centers as a structural and joist-deck demand driver. They do not give a mix. That is not a mill-share story, and it is not a reason to re-rate Quanta, Eaton, or Vertiv. It is the same physical-build constraint the electrical book already has, one layer earlier in the bill.
The other reading of the same wires
Millican and Arent, on that same episode, also made the case against treating every new megawatt as a reason to pour new iron. Industry estimates put about 260 gigawatts of incremental load inside the grid that already exists — if you use dynamic line rating, reconductoring, demand flexibility. Average utilization of the US grid is around 40 percent. Gas turbines on the system run around 50 percent. Peakers sit at 1 or 2 percent. The rules, Arent noted, were written more than a hundred years ago.
PPL tried the cheap version. Dynamic line rating in its territory cut congestion 65 percent and saved customers $50 million. In Texas, a data center that agrees to be curtailable — to drop load when the grid needs it — buys a faster place in the queue. That reduces peak stress. It does not reduce average consumption. Yake adds a cousin of the same idea: behind-the-meter generation at a campus can be a grid asset in an emergency, if the operator will turn the hall down or fire the engines up. That takes a contract, not a press release.
When the queue moves, the orders follow
The hyperscaler buildout has its own regulatory gate. In October 2025 the Department of Energy invoked Section 403 of the Federal Power Act, asking every regional grid operator to justify how large loads interconnect. Robin Millican, on a Columbia episode recorded around June 30, said what landed was not a single national rule but show-cause orders to the RTOs — a staggered reform path, not a binary unlock. NARUC’s state-jurisdiction challenge is still live. New York’s July 2026 moratorium on large data-center permits is a separate gate: a favorable federal ruling does not clear a state environmental hold.
Where permits and equipment align, the order books are already enormous. Vertiv carried a record $15 billion backlog with book-to-bill near 2.9x. Eaton’s Q1 data-center orders rose 240 percent year on year against a 228-gigawatt book — about twelve years at 2025 build rates. The live figure is the Q2 restatement: 307 gigawatts, fifteen years, most of it not converting soon. Quanta Services sits at a record $53 billion backlog. Brannon McBee, CoreWeave’s chief development officer, put the binding constraint in plain English: “Having a powered shell is the bottleneck today.”
Transformers are the longest pole. Lead times stretch to three or four years, against about eighteen months in the old world. Earl Austin, Quanta’s chief executive, said the wait is on the order of thirty-six months — and Quanta is spending $500 million to $700 million to manufacture its own, plus a breaker deal with HICO because high-voltage breakers are “very difficult to get.” Whether that vertical integration prints manufacturing margin or only wins bids is still open on the wiki. SemiAnalysis reports some builders routing around large power transformers entirely by delivering medium-voltage power from on-site generation — a bypass that pressures the scarcity premium without changing the aging-fleet replacement math.