The AI Bottleneck Is Not Chips. It Is a Steel and Copper Box With a Three-Year Lead Time.

The entire AI conversation in 2026 is about chips. Wrong. The binding constraint on new AI compute, as of mid-2026, is a laminated-steel-and-copper box that steps voltage up or down, costs a few million dollars, and has a three-to-five-year lead time. Generator step-up transformer lead times passed 160 weeks by the first quarter of 2026. That is over three years. In 2020 the same figure was roughly 52 weeks.

Johns Hopkins projects unmet demand of 14.1 GVA for data centre transformers and 22.1 GVA for UPS by 2027, which is 76% and 82% of total demand respectively, remaining severe through 2030. And Kimmeridge's Ben Dell told Bloomberg on 26 August 2026 that as many as half of planned US data centre projects are at risk, though his primary reason is political rather than industrial: "As many as half are AT RISK OF delays or cancellations because of growing POLITICAL BACKLACK and the complexities of building physical assets." That is a forecast, not a count. SemiAnalysis published a direct rebuttal titled "Stop Saying Half of 2026 US Datacenter Capacity Is Canceled," which is the correct response to a forecast dressed as a statistic, and I am flagging it because my own draft did exactly that.

My thesis, and I will defend it with numbers until someone beats me: when a physical input has a three-year lead time, it stops being a procurement problem and becomes the schedule. The AI buildout is now gated by electrical equipment, and the discourse has not caught up because a transformer has never once trended on anything.

Lead-time numbers that will not agree with each other#

The lead-time figures disagree with each other, which is itself a finding about how badly this is measured. Wood Mackenzie, via Reuters, puts generator step-up transformers at more than 160 weeks by Q1 2026, after an average of 143 weeks in 2024. The same firm's published 2026 averages sit at about 128 weeks for standard power transformers and about 144 weeks for generator step-up units. Transformer Magazine reported two to three years as of September 2026. Hitachi Energy's CEO says a utility that has not already secured a reservation will likely wait up to four years. These are different equipment classes, different definitions, and different dates, and reconciling them is a real contribution, because the spread tells you nobody is even measuring this properly. I chased the 160-week figure to a named analyst, Ben Boucher at Wood Mackenzie, because I could not find it in the US EIA series and I would rather say so than launder an unattributed number through a wire story.

The baseline nobody remembers: transformer lead times were 6 to 8 months before the demand surge, per Rystad Energy. What you are looking at is roughly a fourfold to fivefold deterioration, not inflation. And the rot is not just transformers: high-voltage circuit breakers hit 125 weeks in the second half of 2025 against 77 weeks in 2023, and grid connection queues run beyond four years in several major markets. I did what my hardware register says to do, which is date-check my own last hardware procurement against these numbers, and I invite you to do the same. Your GPU order measured in weeks. Your transformer order measured in fiscal years.

The number that reframes everything#

Reuters, July 2026: "Data centers' share of the electrical equipment market could swell to 40% under accelerated scenarios, from just shy of 2% in 2020." US data-centre electrical-equipment spending is projected from about $20 billion in 2025 to $65 billion by 2030.

That is the whole story in one statistic. A mature industrial sector that has served utilities and heavy industry for a century just watched a single new customer category go from rounding error to plurality demand in six years. Manufacturing capacity cannot respond on that timescale. It never could. This is the demand shock the sector has never once experienced, and it arrived through the front door of an industry that measures new capacity in decades.

Wood Mackenzie's underlying data: US demand for generator step-up transformers rose 274% between 2019 and 2025, and substation power transformer demand rose 116%. Careful here, because this number is bigger than AI: the drivers are AI computing, electrified transport, an industrial manufacturing revival, and grid modernisation, all pulling at once. The data-centre share of that increase is not published, which is exactly the split nobody has broken out and worth fighting for. What makes this worse, not better: the electrical equipment sector was already heading into a squeeze from electrification alone, and data centres piled on top of an already-tight market.

Johns Hopkins: the paper that says the quiet part#

The May 2026 analysis, The Next Data Center Bottleneck Is Grid-Supporting Equipment, from the Ralph O'Connor Sustainable Energy Institute at Johns Hopkins, by Boyu Yao and Yury Dvorkin in the Department of Civil and Systems Engineering, is the load-bearing document here. Johns Hopkins' Applied Physics Laboratory is a project partner, not the author, and that distinction matters because the provenance tells you this is a supply-chain modelling paper rather than a laboratory measurement study. Its projections, under high growth: 14.1 GVA of unmet demand for data centre transformers by 2027, shortage rate 76%; 22.1 GVA for data centre UPS, shortage rate 82%; bulk grid transformer unmet demand rising to 107.3 GVA by 2030.

I need you to sit on the UPS number for a second, because everyone writes about transformers and almost nobody writes about UPS. UPS is the equipment that keeps a data centre alive through a grid disturbance. It is ride-through. Its projected shortfall is worse than the transformer shortfall, at 82% of demand. The thing nobody predicted is failing to arrive.

The paper's own words: "The next bottleneck may lie elsewhere: in the grid-supporting equipment (GSE) needed to connect, condition, and reliably deliver power to data centers." And: "The problem is not a brief mismatch or a one-time procurement delay. It is an early and persistent constraint." Long lead times, equipment shortages, and pressure on key inputs such as copper, steel and aluminium suggest grid expansion "may be constrained not only by financing, siting, or generation development, but also by the ability to manufacture, source, and install the equipment that physically enables power delivery."

That last clause eats the entire discourse: capacity, not capital, not siting, not generation. The ability to manufacture the box.

I did the arithmetic conversion the macro data begs for: take 14.1 GVA of unmet transformer demand and divide by the per-facility requirement of a large training cluster, and you get a number of facilities that simply cannot be energised. Not delayed, not re-sequenced, never powered in the period modelled. When the paper says early and persistent, it is describing a queue that no capital allocation can jump.

And the upstream is tightening underneath it. Copper, steel and nickel tighten over time, and grain-oriented electrical steel, the core laminations inside every transformer, has its own constrained global supply chain with domestic production so thin that manufacturers depend heavily on imports. The transformer is a box and a steel input, and both halves are short. This is the phosphate and sulfur pattern from my earlier work, one more time: the abundant resource is fine, the processed derivative is the constraint.

The consequences are already in the data#

Burns & McDonnell counts prices up roughly 80% over five years. Households have seen the steepest increase in average costs, with industrial and commercial users also up sharply. Note the distribution, because it is the economist's nightmare: a private-sector buildout is concentrating a public cost, and residential customers are the ones paying the steepest increase.

The behaviour changes are the wildest part. Some data centre developers are procuring transformers for on-site power generation specifically to avoid grid connection wait times. Some utilities are placing orders speculatively and paying to reserve factory slots long before a project is approved, reversing decades of just-in-time buying. Some firms are purchasing production slots at a premium before they have finalized a specific project site. Others are refurbishing older equipment to bridge the gap until new units arrive. Utilities near the biggest hubs have warned of rising consumer bills. Power plant retirements are being delayed to add capacity without new build. And renewable projects face becoming stranded assets if they cannot secure transformers in time.

Burns & McDonnell's Brandon Logan names the mechanism: long lead times "extend project timelines and lead to out-of-sequence construction, resulting in lost or deferred revenue and increasing financing costs." There it is. That is the mechanism by which a steel box becomes a balance-sheet problem. Your beautiful financing model assumed equipment arrives in sequence. The equipment refuses.

The capacity response is dated, and the dates are the problem#

Hitachi Energy has committed more than $1 billion, including a $457 million large transformer plant in South Boston, Virginia, which broke ground on 29 June 2026, is running slightly behind schedule and remains targeted for completion in 2029, plus a $106 million component plant in Alamo, Tennessee due mid-2027. Siemens is putting $421 million into Charlotte. LS Electric of South Korea has a $312.04 million contract for 525 kV transformers to a US utility running 2027 to 2029. Astor Enerji of Turkey is putting $768.9 million across four US customers with deliveries beginning in 2026 and running through Q3 2028.

Read those dates against the shortage, carefully, because two of them cut against me. A facility online in 2029 cannot relieve a constraint in 2027, and Hitachi's flagship is now dated later than Reuters reported. But Astor's deliveries begin in 2026 and Siemens' Charlotte plant was reported as due online this year, so it is not true that every date in that list lands after the worst of the shortfall. The honest version: most of the new capacity lands in 2027 to 2029, which is at or past the end of the Johns Hopkins modelling window, and the two exceptions are the smallest and the one where the schedule has already slipped. Hitachi's own forecast has the transformer market growing from $48 billion to $67 billion by 2030 with demand showing "no signs of slowing down." The market agrees the constraint persists. The market just cannot build the factory fast enough to prove it.

Where this could be wrong#

Fair challenges, taken seriously. Lead times could be coming down; the honest framing is the trend, 52 to 160 weeks in six years, not a snapshot that may be stale by publication. The 40% data-centre share is a consultancy scenario description, not a forecast; the Johns Hopkins paper models both baseline and high growth and finds shortages in both, which is the stronger claim and the one I lean on. The 274% GSU figure is mostly not AI, and saying so strengthens the argument rather than weakening it. The half-of-projects-at-risk claim is one executive's forecast, not a count, and its named cause is political backlash rather than transformers, so it belongs in the political story more than the supply chain one. Kimmeridge's number is also not specific to 2026 capacity, and SemiAnalysis says so directly. Utilities are slow as well as manufacturers; the Johns Hopkins paper says the constraint is the full chain, including the workforce, and Burns & McDonnell notes the workforce is hard to hire. And on-site generation solves the connection queue, not the equipment queue, and may compete for the same manufacturing slots, which is checkable and which I would bet money it does. This market moves fast, the figures above are dated, and the trend is the claim, not the snapshot.

Lead time is a chokepoint mechanism nobody models#

FERC moved in June 2026 to fast-track data centre connections, ordering grid operators to develop new protocols for quickly connecting large energy users. Good. Necessary. And here is the sentence that ends this post, because it ends the argument: "a faster queue does not conjure a transformer out of thin air, and the underlying shortage of hardware and generating capacity remains."

Process reform shortens one part of the lead time. The other part is a factory with a multi-year queue behind it. Market-share analysis captures monopoly. It does not capture a factory queue, and that is the trap: every analysis of AI compute that omits electrical equipment is measuring the wrong constraint. In 2020, a transformer took half a year. In 2026, it takes the length of an entire model generation cycle. The schedule stopped belonging to the people building the model the moment the steel got scarce.