AI's growth trajectory for the next three to five years is set by transformer purchase orders, not model capability. Investors pricing AI stocks on software multiples are mispricing the physical constraint that is currently binding.
The AI investment thesis runs on a software clock. Foundation models improve on eighteen-month cycles. Benchmark scores double. Revenue projections extend the curve. Wall Street prices the sector as if compute scales on the same schedule.
It does not. Every new datacenter requires high-voltage power transformers to connect to the grid. A standard power transformer takes 128 weeks to build. Generator step-up units take 144 weeks. Substation transformers take more than 160 weeks. These timelines have roughly doubled since 2020, when lead times ran 24 to 30 months.
Of the twelve gigawatts of datacenter capacity expected to come online in the United States in 2026, only a third is under active construction. The rest waits for electrical equipment that does not yet exist. Half of planned U.S. datacenter builds have been delayed or cancelled. A single missing transformer can hold up a two-billion-dollar project for a year.
The gap between announced AI capacity and deliverable AI capacity is the most significant mispricing in the sector.
The dead zone
The average high-voltage transformer in the U.S. grid is 35 years old, past the 30-year design peak. Only twenty percent of domestic demand for large power transformers is met by U.S. manufacturing. The rest is imported, primarily from Mexico and South Korea. Demand for generator step-up transformers has grown 274 percent since 2019, far outpacing manufacturing capacity.
Manufacturers are responding. Hitachi Energy committed more than a billion dollars to new facilities in Virginia, Tennessee, and Montreal. Siemens Energy is spending $150 million on its first U.S. large power transformer plant in Charlotte, North Carolina, with production starting in 2027. Eaton committed $340 million to a three-phase facility in South Carolina.
Roughly two billion dollars total has been committed to new manufacturing capacity since 2023. Nearly all of it delivers in 2027 or 2028.
The structural dead zone runs from now until that capacity arrives. Industry analysts cite four-year transformer lead times for large-scale infrastructure. A campus that secures land and financing in May 2026 without a locked-in transformer order cannot power on before 2030. The backlog grows faster than capacity because every new project enters the queue while the manufacturing base remains fixed.
Money cannot solve a capacity constraint. It can only wait.
Who benefits
GE Vernova is the most direct play on the bottleneck. The company posted $18.3 billion in orders during Q1 2026, up 71 percent organically. Its electrification segment booked $2.4 billion in datacenter orders in a single quarter, more than all of 2025. The record backlog reached $163 billion, roughly seventeen times quarterly revenue. Gas turbine slots have grown from 83 to 100 gigawatts, with the company targeting at least 110 gigawatts by year end. Analysts price the stock between $1,144 and $1,400. It trades near $1,039 after rising 124 percent in the past year.
The backlog is the signal. Years of locked-in revenue at rising prices, because the customer cannot source the equipment from anyone else on a shorter timeline.
Siemens Energy offers similar exposure at a lower valuation. In January, UBS raised its price target from €38 to €175, a 4.6-fold increase reflecting the market still repricing this name after years when wind turbine losses suppressed the stock. The Charlotte plant will be its first U.S. large power transformer production facility. Grid infrastructure has become the dominant revenue driver.
Eaton operates one layer further down the stack. Every datacenter that does get built needs Eaton's switchgear, power distribution units, and uninterruptible power supplies. Q1 datacenter orders rose 240 percent. Total electrical segment backlog grew 48 percent year over year. The company raised its organic growth guidance from 8 to 10 percent. Eaton estimates 228 gigawatts of total datacenter capacity in the pipeline, twelve years of backlog at 2025 build rates. Seventy percent of the 32 gigawatts currently under construction is AI-related.
Beyond these three, Hubbell and Powell Industries serve the enabling infrastructure: connectors, enclosures, and power distribution for the substations and switchyards that connect transformer to datacenter.
Who is overpriced
Datacenter REITs trade on announced capacity additions. Equinix and Digital Realty price expansion timelines that assume electrical equipment arrives on schedule. When thirty to fifty percent of planned capacity slips by two years, the revenue projections underlying those valuations slip with it.
The hyperscalers committed more than $700 billion in combined capital expenditure projections. Those projections assume build timelines the transformer supply chain cannot support. The spending is real. The delivery schedule is aspirational. Any investor pricing AI revenue growth on announced capex is implicitly assuming the transformer bottleneck does not bind. It does.
The structural mispricing
The AI investment narrative assumes growth is gated by three things: model capability, capital availability, and customer demand. All three are abundant. The narrative misses a fourth constraint that is currently binding: the physical capacity to build the infrastructure.
Software multiples applied to hardware-gated companies produce the wrong answer. A SaaS company with 40 percent growth and no physical constraint earns a premium multiple. An AI infrastructure company with 40 percent growth and a five-year transformer backlog does not face the same ceiling risk. The transformer manufacturer has a structural floor under its revenue for years. The SaaS company has none.
The contrarian position is straightforward. The companies that make, install, and enable the electrical equipment the AI buildout requires are underpriced relative to the visibility of their revenue. The companies that depend on that equipment arriving on time are overpriced relative to the probability that it will.
The model gets faster every quarter. The transformer does not.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)