The Substation Decade
On March 31, 2026, a company most investors have never heard of filed an earnings release containing the single most important number for understanding the next ten years of AI infrastructure.
The backlog stood at $1.98 billion — up 157 percent from the prior year.
The company is Forgent Power Solutions. It makes transformers.
If you have been following the AI trade, you have probably read fifty thousand words on chips, fifteen thousand on power, and zero on transformers.
That asymmetry is what this edition is about.
The bottleneck the market has agreed to look at is not the bottleneck that matters. The bottleneck that matters has a different shape, a different supplier base, and a different timeline — five years instead of eighteen months.
It has been hiding in plain sight, embedded in earnings releases from companies that do not appear in any consensus AI basket.
I will explain why this matters in a moment, and I will name the consequences for capital allocation in the section after that.
First, the character.
In September 2024, Joe Dominguez, the CEO of Constellation Energy, stood in front of reporters and announced that the Three Mile Island Unit 1 reactor — shut down in 2019 because the economics did not work — would be restarted to supply 835 megawatts of round-the-clock power to Microsoft data centers under a twenty-year PPA.
The capital cost was $1.6 billion. The restart timeline was originally 2028; it has since been pulled forward to 2027. Many of the workers laid off in 2019 were rehired within months.
Dominguez, in remarks I keep coming back to, said publicly: We made a mistake in shutting down this plant.
That sentence is, to me, the cleanest possible summary of the entire era of US energy and infrastructure policy from roughly 2010 to roughly 2023.
A consensus formed, it was wrong, and the cost of the reversal is being paid in 2026 dollars.
Nuclear was the first reversal. The transformer and substation supply chain is the second, and the reversal has barely begun.
The misread bottleneck
For two years, the consensus narrative on the AI buildout has moved in one direction.
First the bottleneck was GPUs, and Nvidia was the bottleneck supplier. Then the bottleneck was power, and the gigawatt became the new unit of measurement.
Both of these are partially true. Both are now priced in.
The chip stocks have absorbed twelve months of forward demand into multiples that assume execution. The power names — Constellation, Vistra, Talen — have rerated on the back of hyperscaler PPAs and nuclear restarts.
There is a third rung that almost no consensus AI note treats as a primary constraint.
It sits between generated megawatts and useful megawatts at the data center rack. It is the physical equipment required to step the voltage up at the plant, transmit it across the grid, step it down at the load, and route it through the substation interface to the customer.
It includes generator step-up transformers (GSUs), high-voltage substation power transformers, switchgear, breakers, and the interconnection infrastructure that ties any new industrial load above roughly 200 megawatts into the regional grid.
Without these — every one of these — a gigawatt of generation and a gigawatt of demand may as well exist on different planets.
They cannot meet. The market has agreed to assume they will.
Pre-2020, lead times on a US high-voltage power transformer were 24 to 30 months.
The industry has just normalized on four to five years.
Power transformers now average 128 weeks. Generator step-up units average 144 weeks.
Since 2019, US demand for GSUs has grown 274 percent. Demand for substation power transformers has grown 116 percent.
Domestic production capacity has not doubled. It has barely moved.
Roughly 80 percent of large power transformers used in the United States are imported — primarily from Mexico and South Korea, with the remainder split among European and Asian suppliers whose own backlogs are saturated by domestic and EU demand.
Sightline Climate, which tracks the industrial side of the AI buildout more carefully than almost any consensus source, published a report this spring documenting 12 gigawatts of new US data center capacity announced for delivery in 2026.
Of that, only 5 gigawatts is actually under construction.
The remaining 11 gigawatts exists on paper but has no visible construction progress. Note that figure quietly: it is the size of a mid-size country’s electricity grid.
The named bottleneck, in Sightline’s own words, is no longer GPUs or chip allocations. It is transformers, switchgear, and batteries.
When you read five-year lead time on a transformer, do not pass over it. Read it again.
A modern semiconductor fab takes roughly three years from groundbreaking to first production.
A nuclear restart, like TMI, takes three to four years from announcement to first MWh.
The transformer is the part of the AI infrastructure stack with the longest lead time.
Not the chip. Not the power plant.
The transformer.
Three companies sit on the constraint
The global power transformer market is one of the more concentrated industrial markets in the developed world.
Hitachi Energy holds roughly 21.7 percent of global share. Siemens Energy holds roughly 19.6 percent. GE Vernova holds roughly 16.2 percent.
Together, the three control approximately 57 percent of the market for the single piece of equipment that has become the binding constraint on the rollout of AI infrastructure in the developed world.
This is not a market structure that responds to demand shocks elastically.
Building a new high-voltage transformer factory takes — depending on jurisdiction and equipment — between three and five years.
Hitachi Energy announced in late 2025 that it would invest roughly $1 billion in expanded US manufacturing capacity, including a $457 million transformer plant in South Boston, Virginia, expected to begin operations in 2028.
Read that sentence carefully: the response to a transformer shortage that is binding now will not produce its first unit until three years from now.
And the additional supply, even at full ramp, will not change the structure of the global market. Hitachi is making the largest single domestic investment of the three majors, and it will move the needle by perhaps five percent of US annual demand by the late 2020s.
This is the supplier base on which the entire AI infrastructure trade depends.
The market is pricing the trade as if the bottleneck is power generation, which is wrong by one rung.
The bottleneck is the steel and copper between the power and the rack.
A detour, because the rhythm requires one
There is a quieter version of this story that lives inside the regional transmission operators — the entities like PJM and ERCOT that manage interconnection of new generation and new load to the grid.
As of the most recent figures, PJM and ERCOT together have 1,216 active interconnection projects representing 111.6 gigawatts of generation and 64.3 gigawatts of storage in queue.
Wait times to obtain a binding interconnection agreement run, in ERCOT, 17 months for natural gas, 23 months for storage, and 27 months for renewables.
PJM, after a recent reform cycle, expects to compress its IA processing to one to two years going forward — an improvement that still leaves it well outside any data center construction timeline.
The vast majority of projects in queue — 60 percent in ERCOT, 76 percent in PJM — will likely not come online for many years, or ever.
This is what the consensus AI trade is built on top of.
A queue.
A queue managed by entities that, until very recently, had not designed their processes for an industrial load shock of this magnitude.
The queue manager is the most consequential financial actor in the AI infrastructure story that no one has heard of.
The trade press has named a half-dozen speed-to-power solutions, but the underlying queue arithmetic does not bend.
A 200-megawatt industrial load — the smallest end of an AI campus — currently requires three to five years from interconnection application to energization in most US grid operators.
The data centers cannot wait. So they bypass.
BYOP — the operators have already given up on the grid
In December 2025, GE Vernova disclosed that it expected to end the year with an 80-gigawatt backlog of gas turbines stretching into 2029, with reservation slots potentially fully sold through 2030 by the end of 2026.
Most of those turbines are not headed to traditional utilities.
They are headed to behind-the-meter deployments — gas turbines installed on-site at data center campuses, with their own generation, their own substations, their own switchgear, contracted on a bring your own power basis.
In Texas alone, the largest single BYOP project — Pacifico Ranch’s GW Ranch — has secured permits for up to 7.65 gigawatts of on-site generation.
That is the entire power consumption of metropolitan Houston, deployed behind the meter, for a single AI campus.
When the largest data center operators conclude that the most reliable way to power a gigawatt of compute is to build their own power plant, what they are saying — without saying it — is that the grid will not be there when they need it.
BYOP is not a strategy. It is a verdict.
It is the operators of the AI infrastructure stack collectively concluding that the regulated utility model, and the transformer-and-substation supply chain it depends on, cannot deliver on the timeline AI demands.
What this means for you, and what I am calling it
The phone in your hand connects, increasingly, to inference servers that consume a measurable share of national electricity.
The model that wrote you a response this morning ran on a GPU that sits in a building that exists because someone — three years ago — placed an order with Siemens Energy for a 500-MVA transformer that arrived last quarter.
If you have bought any part of the AI trade in the last 18 months — directly, through an index, or through a fund — you have implicitly bet that this supply chain extends.
The bet most consensus narratives have ignored is the bet most likely to determine whether the trade works.
I have called this, for my own reference over the last several months, the Substation Decade.
The phrase is not subtle, and it is not meant to be.
The defining infrastructure bottleneck of the AI economy is not the silicon. It is the steel.
The next ten years of capital deployment will be governed by the lead time of equipment most investors have never priced, manufactured by three companies most investors have never owned, gated by regulatory queues most investors do not understand.
The trade that follows from that recognition is not the trade the consensus is doing.
The chip names will rerate down.
The power names — Constellation, Vistra, Talen — have already absorbed the first wave of repricing on hyperscaler PPAs, but they will rerate again on a second wave when the market figures out that contracted megawatts are not the same as deliverable megawatts, and that the gap between the two is filled with equipment that does not exist yet.
The equipment makers — Hitachi Energy, Siemens Energy, GE Vernova, Eaton — and the specialty contractors who install them — Quanta Services, MasTec, Primoris — are sitting on the actual rent.
They have been telling us this in their earnings releases for two years.
The consensus has been reading other earnings releases.
I am skipping the part of this edition I find most operationally important, because that part belongs behind the paywall.
What follows below the cut is the trade.
If you have read this far, you already know the most important thing about this story: the consensus AI trade is mispriced on the wrong rung of the supply chain.
You also know which kinds of companies sit on the actual constraint, and roughly when the constraint will start to bind.
The thing you do not yet have is the trade.
What is behind the paywall:
— The three names within the equipment-and-installation complex with the cleanest
exposure to the substation constraint, including the one that is mispriced because
the market is still tracking the wrong end of the value chain
— The catalyst calendar between now and Q4 2026: which earnings prints will mark
the trade, which order disclosures to watch, which queue reforms could compress
or extend the bottleneck
— The two scenarios where the thesis breaks — named explicitly, with the trigger
we are watching for each — including the one I assign the highest non-trivial
probability
— A hedge structure for readers who already have core exposure to the AI
infrastructure trade through the chip names or the power names, and want
isolated exposure to the constraint without doubling the AI beta of their book
Annual subscription remains at the same rate it has been for the last six months. That rate is not promised forever, but it is the rate today…