At the end of 2025, more than 2,000 gigawatts of generation and storage capacity was sitting in American interconnection queues, waiting for permission to connect to the grid. That is a vast amount of electricity nobody can use.
Meanwhile, the grid operator serving thirteen states and 67 million people held an auction in December 2025 that cleared at its regulatory price cap and still came up short of what it needs to keep the lights on reliably.
Those two facts are the whole story. The United States is not running out of electricity. It is running out of the ability to connect things to the grid at the speed AI is being built. A data center takes eighteen to twenty-four months to construct. Everything it needs to plug into takes considerably longer, and no amount of capital compresses the difference.
Understanding which bottleneck is actually binding matters, because the three proposals you hear most — build more generation, use AI less, wait for the market to catch up — address different ones, and two of them barely touch the problem.
The Demand Is Real, and It Is Concentrated
Start with the scale, because it is genuinely large.
Lawrence Berkeley National Laboratory, in the assessment Congress mandated under the Energy Act of 2020, put US data center electricity use at 176 TWh in 2023 — about 4.4% of national consumption, up from 58 TWh in 2014. Its projection for 2028 is a range rather than a number: 325 to 580 TWh, or 6.7% to 12% of national electricity.
Globally, the International Energy Agency's Energy and AI report puts data centers at about 415 TWh in 2024, roughly 1.5% of world electricity, rising to about 945 TWh by 2030 in its base case — more than double, and comparable to Japan's entire national consumption today.
But national percentages understate the problem badly, because the load does not arrive evenly. It arrives as single facilities drawing hundreds of megawatts, clustered into a handful of regions with cheap land, fibre and favourable tax treatment. A grid can absorb 12% more demand spread across a decade and a continent. It cannot absorb a 600 MW customer appearing in one county in three years.
That concentration is what turns a demand story into an infrastructure story. (The per-query side of the argument is a different question entirely, and a much smaller one — we cover it in how much energy one AI query actually uses.)
Bottleneck One: There Are Two Queues, and Coverage Keeps Merging Them
This is the most common error in reporting on the subject, and it makes the numbers incoherent.
The generator interconnection queue is where power plants wait. If you want to build a solar farm, a battery installation or a gas plant and sell into the grid, you file a request, and the operator studies what your project does to system stability and who pays for the upgrades it requires.
The large-load queue is where big customers wait. It is a separate process with separate rules and separate thresholds — ERCOT runs one for loads of 75 MW and above, and PJM defines a large load as 50 MW or more at a single point of interconnection.
They are not interchangeable, and they do not offset each other. A gigawatt of solar clearing the generator queue in Texas does nothing for a data center waiting in Virginia. Yet headlines routinely add the two totals together, or quote a load-queue figure as though it were generation about to arrive.
Load-queue figures are also unusually unreliable. Reported ERCOT large-load totals through 2026 range from roughly 240 GW to over 400 GW depending on the source and the date, partly because a single prospective data center can file in several places at once while it shops for the fastest connection. Treat any load-queue number as an expression of interest, not a pipeline.
What a Queue Number Does Not Mean
Even for generation, where the data is good, the headline figure is widely misread.
Berkeley Lab's Queued Up series tracks this directly. Its 2026 edition, covering data through the end of 2025, found 2,061 GW of generation and storage actively seeking interconnection — down 10% from the year before, as withdrawals outpaced new requests. Natural gas was the exception, rising 86% year on year to 253 GW, which is itself a signal about what developers expect the next few years to need.
Then comes the number that reframes everything else. Of all the capacity that requested interconnection between 2000 and 2020:
- 13% had reached commercial operation by the end of 2025.
- 75% had been withdrawn.
- More than 40% withdraw even after signing an interconnection agreement.
So "2,061 GW is waiting to connect" does not mean two thousand gigawatts are coming. It means most of those projects will never be built, and nobody can tell you in advance which ones. Developers file speculatively because a queue position is cheap and valuable; the queue clogs; studies slow down; more developers file early to hedge against the delay. The congestion is partly self-inflicted.
And it is slow regardless. Berkeley Lab found the median time from request to a signed interconnection agreement was well over three years in 2025, with the full timeline to commercial operation exceeding five years in the regions where data is available.
Bottleneck Two: The Equipment Has Its Own Clock
Suppose you clear the queue. You still have to buy the machine.
Heavy-duty gas turbines — the technology most utilities are turning to for firm capacity that can follow demand — have been effectively sold out. Wood Mackenzie found the average lead time for delivery had stretched to nearly five years from receipt of order, and the major manufacturers have publicly described directional lead times of three years or more on new orders. Manufacturers are expanding production, and lead times are forecast to ease from 2027, but a factory expansion is itself a multi-year project.
This is the bottleneck least responsive to money. A hyperscaler with effectively unlimited capital cannot buy a turbine slot that does not exist, and the queue for the turbine runs in parallel with the queue for the interconnection rather than after it. Transformers, switchgear and high-voltage cable face similar constraints for similar reasons: a decade of flat demand gave nobody a reason to build spare manufacturing capacity.
What It Looks Like When the Clocks Don't Line Up
PJM, the grid operator for thirteen states and the District of Columbia, is where this has become most visible — and its own numbers tell the story without editorialising.
In the December 2025 capacity auction for the 2027/2028 delivery year, PJM procured 134,479 MW and fell 6,623 MW short of its reliability requirement. It was the first auction in which the entire footprint came up short. The clearing price hit the FERC-approved cap of $333.44 per MW-day across the whole region; PJM estimated that without that cap — negotiated with Pennsylvania's governor and due to expire — the price would have been close to $530.
The following auction, held in July 2026 for 2028/2029, procured more capacity and came up 6,831 MW short. The gap widened.
PJM's own February 2026 analysis names the condition directly: a "transition gap" in which demand growth has temporarily outpaced infrastructure development. That is an unusually plain description from a system operator, and it is the correct one. Prices are not high because generators are scarce in some absolute sense. They are high because the thing being auctioned — capacity available on a specific date — cannot be manufactured on the timescale demand is arriving.
The cost of that gap does not fall only on data centers. It falls on every customer in the region through capacity charges, which is why this has stopped being an energy-industry story and become a political one.
The Fix That Gets Misquoted
There is a genuinely promising line of research here, and it deserves care because it is routinely overstated.
Researchers at Duke University's Nicholas Institute asked what would happen if new large loads agreed to curtail occasionally rather than demanding firm power every hour. Their February 2025 study, covering the 22 largest US balancing authority areas — about 95% of national load — found that 76 GW of new load could be integrated at an average annual curtailment rate of about 0.25%, rising to 98 GW at 0.5% and 126 GW at 1%.
That is roughly 10% of US peak demand, available without building anything new. It is a real and important finding.
Two caveats travel badly and usually get dropped:
- The 0.25% is a share of maximum uptime, not a simple count of hours offline. Trade coverage translates it to roughly 85 hours a year, and the study notes that for 88% of curtailment time, half the new load would keep running. "Switched off for 85 hours" is not what it says.
- It is a system-level result, not a per-project entitlement. The study estimates headroom across a whole balancing area. It does not establish that any individual data center should receive reduced network upgrade obligations — an inference some advocates have drawn, and one its authors did not make.
Flexibility is probably the single highest-leverage lever available in the next five years, precisely because it works on the timescale the problem has. It is not a reason to skip building.
How to Read a Data-Center Grid Headline
Four questions, which will resolve most of the confusion in circulation:
- Is this generation or load? Two different queues. Numbers from one say nothing about the other.
- Is the figure a request or a commitment? Historically, 13% of requested generation capacity gets built and 75% is withdrawn.
- Is the constraint energy or capacity? Total annual terawatt-hours and firm megawatts available at 6pm on a still August evening are different problems with different solutions.
- What is the clock? A proposal that needs a new transmission corridor and a proposal that needs a demand-response agreement sit a decade apart.
The Bottom Line
The grid is not short of electricity, and AI is not going to exhaust the nation's generating potential — there is more capacity queued up than will ever be built. What the grid is short of is time: interconnection studies that take over five years, turbines that take three to five, and transmission that takes longer than both, against data centers that go up in under two.
That diagnosis points somewhere specific. Faster interconnection processing, serious penalties for speculative queue positions, expanded equipment manufacturing, and flexible-load agreements all attack the binding constraint. Building more generation that then waits five years to connect does not, and neither does arguing about individual prompt usage.
It also means the energy transition and the AI build-out are now competing for the same scarce resource — not fuel, not land, but a place in line. Whether that pushes investment toward firm low-carbon sources like the nuclear fusion programmes still some way from the grid and today's fission fleet, or toward gas with carbon capture bolted on, is probably the most consequential open question in climate and energy right now. The queue will decide a lot of it.
Frequently Asked Questions
Is AI actually going to cause blackouts?
No serious grid operator is forecasting that, and the mechanisms that would prevent it — demand response, emergency procedures, interconnection conditions — are well established. What is already happening is cost: PJM's capacity auctions have cleared at their price cap twice running while falling short of the reliability requirement, and those costs reach ordinary customers through capacity charges. Reliability margins narrowing and prices rising is the realistic risk, not the lights going out.
Why can't data centers just build their own power plants?
Many are trying, through on-site generation, co-location at existing plants and long-term nuclear contracts. The constraints are the same ones everyone else faces: gas turbines have multi-year lead times, new nuclear has much longer ones, and a plant that connects to the public grid still enters an interconnection queue. Co-locating at an existing plant avoids some of that, but raises an unsettled regulatory question about whether the data center is then using grid services it is not paying for.
How long does it take to connect a new power plant to the US grid?
Berkeley Lab's 2026 analysis found the median time from interconnection request to a signed agreement was well over three years, with the full timeline to commercial operation exceeding five years where data is available. That is substantially longer than a decade ago. The delay is driven by study backlogs, the cost allocation of network upgrades, and the volume of speculative requests clogging the process.
Does building more solar and wind help with this?
It helps with energy, which is total terawatt-hours over a year, and less with capacity, which is firm megawatts guaranteed at a specific hour. Data centers run at high utilisation around the clock, so they need capacity. Paired with storage the gap narrows considerably — storage made up 749 GW of the 2,061 GW in queues at the end of 2025 — but solar on its own does not resolve a capacity shortfall.
Where is the strain worst?
In the regions where data center development has concentrated: PJM's footprint, especially northern Virginia, and ERCOT in Texas. Both have seen large-load interconnection requests arrive far faster than infrastructure can be added. The national percentages — around 4.4% of US electricity in 2023 — considerably understate what is happening in those specific places.
Will this get better on its own?
Partly. Gas turbine lead times are forecast to ease from 2027 as manufacturing capacity expands, interconnection reform is underway at several operators, and the withdrawal rate from queues is high enough that the backlog is already shrinking — total queue volume fell 10% in 2025. What will not resolve on its own is the mismatch between an eighteen-month build and a five-year connection, which needs a process change rather than time.
Sources
- Queued Up: 2026 Edition — Lawrence Berkeley National Laboratory, July 2026
- 2024 United States Data Center Energy Usage Report — Lawrence Berkeley National Laboratory, December 2024
- Energy and AI — International Energy Agency, April 2025
- PJM Auction Procures 134,479 MW of Generation Resources — PJM Interconnection, December 2025
- 2027/2028 BRA Reserve Target Shortfall Report — PJM Interconnection, February 2026
- Rethinking Load Growth — Nicholas Institute for Energy, Environment & Sustainability, Duke University, February 2025
