The binding constraint is a transformer, not a chip

TL;DR. Roughly 2,300 gigawatts of generation and storage sit in US interconnection queues, which is more than the country's entire installed capacity. Grid connection waits in Northern Virginia, Phoenix and Dallas run 4 to 7 years, and that applies to campuses with full capital, allocated GPUs and broken ground, because the queue is set by the utility rather than the operator. Large power transformers run 3 to 5 year lead times, up from 24 to 30 months before 2020. Medium-voltage switchgear is effectively sold out through 2028. The World Resources Institute finds lead times extending construction timelines by 24 to 72 months in the most affected markets. The scarce input is no longer capital or silicon. It is energised power, and the response, building generation on site to skip the queue, solves a schedule problem by unsolving an emissions one.

---

Status: established, with a source-quality warning. The underlying data comes from Lawrence Berkeley National Laboratory's interconnection queue work, Wood Mackenzie's transformer market survey, Sightline Climate's project tracking, the World Resources Institute and FERC. Several of the most quotable secondary sources sell solutions to the problem they describe, including off-grid campus operators and turbine vendors, and their framing is weighted accordingly.

---

The queue

Around 2,300 gigawatts of generation and storage capacity is waiting in US interconnection queues. For scale, that is more than the entire installed generating capacity of the United States.

Globally, roughly 1,650 gigawatts of renewable generation sits in connection queues, unable to reach consumers because the wires and substations to carry it do not exist yet.

The queue is a waiting list maintained by grid operators for new generation and large-load connections, and it has become the rationing mechanism for the AI buildout.

In the three highest-density US data centre markets, Northern Virginia, Phoenix and Dallas, waits run 4 to 7 years.

That figure applies to campuses that already have everything else. Full capital, allocated chips, permits, broken ground. The queue position is set by the utility, not the operator, which is why grid power binds across the pipeline regardless of how well funded a project is.

The equipment

Even an approved interconnection agreement does not deliver electrons. A facility needs substation transformers, generator step-up transformers and switchgear.

Large power transformers now run 3 to 5 year lead times, up from 24 to 30 months before 2020.

Medium-voltage switchgear is effectively sold out through 2028.

Wood Mackenzie recorded demand for generator step-up transformers rising 274% between 2019 and 2025, with substation transformer demand up 116% over the same period. Supply did not follow, because transformer manufacturing is capital-intensive, skill-constrained and sized against decades of flat demand.

The World Resources Institute finds these lead times extending data centre construction timelines by 24 to 72 months in the markets most affected.

And an irony worth stating plainly

US imports of medium-voltage switchgear from China went from about 1,500 units in all of 2022 to more than 8,000 in the first ten months of 2025.

The same policy environment that restricts advanced chip exports to China is importing from China the equipment that connects American data centres to the grid.

That is not a contradiction in any strict sense; export controls and import dependence are different instruments in different directions. It is a reminder that supply chain policy focused on the most visible component leaves the less visible ones untouched, which is the chokepoint problem in reverse: there the narrowest point was upstream and unwatched, here it is downstream and unwatched.

What this does to the buildout

The constraint has moved. Earlier in the cycle, packaging capacity at foundries lagged chip demand and a facility with power to spare could sit half empty waiting for silicon. Advanced packaging capacity has since expanded repeatedly and GPU shipments scaled with it.

Grid capacity did not. Interconnection queues, transformer manufacturing and utility capital planning operate on cycles measured in years, and none of them accelerated to match.

The consequence is a repricing of physical assets. Sites that can deliver substation capacity quickly command premiums with no precedent in industrial real estate. Farmland next to high-voltage transmission corridors in Indiana and Ohio trades at multiples of agricultural value, not because anyone wants to farm it, but because the queue for those corridors is years shorter than for Northern Virginia.

And projects are failing. Sightline Climate tracked 9 cancelled projects in its 2026 dataset as of May. Maine voted 82 to 62 for a state-level moratorium through 2027. Industry estimates suggest a substantial share of planned 2026 openings will slip or be cancelled, though those estimates come from parties with an interest and should be read as such.

The response, and what it costs

Operators are building generation on site to skip the queue entirely.

Behind-the-meter generation converts a five-to-seven-year utility wait into a 12-to-18-month equipment delivery and commissioning schedule. Gas turbines and reciprocating generators sized for the full IT load, owned by the operator, with no interconnection request at all.

As a schedule decision this is rational and it is happening.

As an emissions decision it is a step backwards, and the framing rarely says so. A data centre drawing from a grid receives that grid's mix, which in most markets is decarbonising. A data centre burning gas on site has the emissions of gas, and it has them for the twenty-year life of the equipment, locked in by a decision made to solve a queueing problem.

Which is the scope problem again: the timeline improves, the emissions move from someone else's ledger to your own, and the aggregate gets worse while every individual decision is defensible.

Three things this establishes

Capital is not the scarce input. Announced hyperscaler spending runs to hundreds of billions and cannot buy a transformer that has not been built or a queue position that does not exist. A constraint that money cannot relieve behaves completely differently from one that money can, and most forecasting treats them the same.

The bottleneck moved and the discussion did not. Coverage of AI infrastructure still centres on chip allocation, which was the binding constraint two years ago and is now substantially eased. The current one is a utility waiting list, which is less interesting to write about and more determinative of what actually gets built.

And who holds the scarce input holds the pricing power. Utilities with available capacity, owners of permitted sites and electrical equipment makers with full backlogs are price-makers. The AI companies are price-takers, which inverts the usual assumption about where leverage sits in this industry.

What it does not establish

That the buildout stops. Projects are being delayed and relocated more than abandoned, and secondary markets in Texas, Georgia and Indiana are absorbing displaced demand.

That the queue figures mean what they appear to. Interconnection queues contain speculative projects that will never be built, and a queue of 2,300 gigawatts is not 2,300 gigawatts of real intent. Queue length overstates genuine demand and the degree of overstatement is disputed.

That behind-the-meter generation is universally worse. Where it displaces marginal gas on a fossil-heavy grid the difference is small, and some deployments are paired with capture or intended for hydrogen conversion.

And nothing about whether any specific project should proceed. That depends on local grid conditions this article has not examined.

What is unresolved

Whether reform accelerates connections. FERC Order 2023 restructured queue processing, and whether it materially shortens waits is not yet observable in the data. Analysis suggests structural relief is unlikely before the end of the decade.

How much queue volume is real. No published method separates speculative requests from committed projects, and the ratio determines whether the backlog is a crisis or an artefact.

Whether transformer capacity expands in time. Manufacturing investment is underway and the equipment takes years to build, so relief arrives after the current capex wave rather than during it.

And what behind-the-meter generation does to emissions in aggregate. Nobody has published a figure, because it requires knowing how much capacity moved off-grid and what it displaced.

The counter-argument

Queue length is a famously bad indicator. Interconnection queues have been described as backlogged for a decade, they contain large volumes of projects that never proceed, and citing the total as though it represented demand is exactly the error this corpus criticises elsewhere. A queue of 2,300 gigawatts against installed capacity is an arresting comparison and a weak one.

The four-to-seven-year figure is market-specific. It describes the three most congested markets in the United States. Elsewhere connections are faster, which is why the buildout is relocating rather than halting, and quoting the worst markets as the general condition overstates the constraint.

Many sources here sell the alternative. Off-grid campus operators, turbine vendors and industrial property analysts all benefit from the perception that grid connection is hopeless. Their data may be sound and their framing is not disinterested, which this article states and does not fully escape.

And the emissions objection may prove temporary. Behind-the-meter generation is being deployed with an expectation of grid connection later, in which case it is bridge power rather than a twenty-year lock-in, and treating it as permanent assumes a fact not in evidence.

The short version

Roughly 2,300 gigawatts sits in US interconnection queues, more than the entire installed capacity of the country. Grid connection waits in Northern Virginia, Phoenix and Dallas run 4 to 7 years, and that applies to campuses with capital, chips, permits and broken ground, because the queue belongs to the utility.

The equipment is the other half. Large power transformers on 3 to 5 year lead times, up from 24 to 30 months before 2020. Switchgear sold out through 2028. Generator step-up transformer demand up 274% since 2019 against supply that did not move. The World Resources Institute puts the resulting delay at 24 to 72 months in the worst markets.

And US switchgear imports from China went from about 1,500 units in 2022 to more than 8,000 in ten months of 2025, while the same policy environment restricts chip exports in the other direction. Supply chain policy aimed at the visible component leaves the invisible one untouched.

The constraint has moved and the conversation has not. Chip allocation bound two years ago and has eased. What binds now is a waiting list, and the people holding scarce capacity, utilities, permitted sites, equipment makers, are the price-makers while the AI companies are price-takers.

The response is to leave the grid. On-site generation turns a five-to-seven-year wait into twelve to eighteen months. It is a rational schedule decision and an emissions decision nobody frames as one: grid supply is decarbonising, gas on site is not, and the choice locks in for the life of the equipment.

Common questions

What is an interconnection queue? A waiting list maintained by regional grid operators for new generation and large-load connections. Around 2,300 gigawatts of generation and storage capacity currently sits in US queues, more than the country's entire installed capacity, and roughly 1,650 gigawatts of renewable generation waits in queues globally. It has become the effective rationing mechanism for the AI buildout.

How long are the waits? In Northern Virginia, Phoenix and Dallas, the three highest-density US data centre markets, 4 to 7 years. That applies to campuses that already have full capital, allocated GPUs, permits and construction underway, because queue position is determined by the utility rather than the operator.

Why can't money solve this? Because the scarce inputs are physical and time-bound. Large power transformers run 3 to 5 year lead times, up from 24 to 30 months before 2020, and medium-voltage switchgear is effectively sold out through 2028. Demand for generator step-up transformers rose 274% between 2019 and 2025 while manufacturing capacity, which is capital-intensive and was sized against decades of flat demand, did not follow. A constraint capital cannot relieve behaves differently from one it can.

What is the switchgear point about China? US imports of medium-voltage switchgear from China rose from about 1,500 units in all of 2022 to more than 8,000 in the first ten months of 2025. The same policy environment that restricts advanced chip exports to China depends on Chinese equipment to connect American data centres to the grid. These are different instruments pointing in different directions rather than a strict contradiction, and the point is that supply chain policy focused on the most visible component leaves less visible ones untouched.

What are operators doing about it? Building generation on site. Behind-the-meter gas turbines and reciprocating generators sized for the full IT load convert a five-to-seven-year utility wait into a 12-to-18-month equipment delivery and commissioning schedule, with no interconnection request at all. As a schedule decision it is rational and it is happening at scale.

What does that do to emissions? It moves them and usually increases them, and the framing rarely says so. A facility drawing from a grid receives that grid's mix, which in most markets is decarbonising over time. A facility burning gas on site has the emissions of gas for the life of the equipment, locked in by a decision made to solve a scheduling problem. The counter is that some deployments are intended as bridge power pending later grid connection, which would make the lock-in temporary.

How reliable are these figures? The underlying data is solid: Lawrence Berkeley National Laboratory on queues, Wood Mackenzie on transformers, Sightline Climate on project tracking, the World Resources Institute on timelines. The secondary framing is less disinterested, since off-grid campus operators, turbine vendors and industrial property analysts all benefit from the perception that grid connection is hopeless. Queue length in particular is a weak indicator, because queues contain large volumes of speculative projects that never proceed.

Does this mean the AI buildout is stalling? Not stalling, relocating. Projects displaced from capacity-constrained markets are moving to Texas, Georgia and Indiana, and farmland adjacent to high-voltage transmission corridors is trading at multiples of agricultural value because the queue there is years shorter. Nine cancellations were tracked in one 2026 dataset and Maine voted for a moratorium through 2027, so some projects do die, but the dominant effect so far is delay and geographic resorting rather than abandonment.