Accelerators have lead times measured in months. Grid connections have lead times measured in years, and the grid interconnection queue is where that mismatch becomes a number you can look up. At the end of 2025 it held more than 2,060 gigawatts of proposed generation and storage in the United States alone.
For anyone planning where large-scale compute will exist in 2030, this is the constraint that binds. Not fabrication capacity, not capital, not model architecture — the availability of a site where enough power can actually be delivered, on a timeline that a procurement cycle can absorb.
The encouraging part is that this is one of the best-documented bottlenecks in infrastructure. The data is public, current, and specific.
The quick answer
An interconnection queue is the study process a generator must complete before connecting to the transmission grid. Berkeley Lab's Queued Up: 2026 Edition, covering data through the end of 2025, finds roughly 8,200 active projects representing 1,312 GW of generation and 749 GW of storage. The median time from interconnection request to commercial operation, for projects completed in 2025, exceeded five years. And of the capacity that requested interconnection between 2000 and 2020, only 13% had reached commercial operation by the end of 2025 — 75% was withdrawn. A queue position is not capacity.
Key takeaways
- Over 2,060 GW sat in US interconnection queues at the end of 2025 — vastly more than will ever be built.
- Median request-to-operation duration exceeded five years for projects completed in 2025, and is increasing.
- Only 13% of 2000–2020 request capacity had reached commercial operation by end-2025; 75% was withdrawn.
- 549 GW already holds a draft or executed interconnection agreement but has not reached commercial operation.
- Active natural gas capacity in the queue rose 86% in 2025 to 253 GW, while solar, wind and storage each fell around 16–19%.
- The IEA notes that "long grid connection queues" push post-2030 data centre supply toward fossil fuels in its high-growth case.
What an interconnection queue actually is
Before a power plant can connect to the transmission system, the operator — an ISO, RTO or utility — must study what the connection does to the grid. As Berkeley Lab describes it, that process "establishes what new transmission equipment or upgrades may be needed before a project can connect to the system and assigns the costs of that equipment".
Two things about that sentence do the damage. The studies are sequential and interdependent, because what one project requires depends on which other projects proceed. And the cost allocation is only known at the end — a developer can spend years in the process and then discover that the required network upgrades make the project uneconomic.
The result is a queue that behaves less like a line and more like a lottery with a multi-year entry period. And because withdrawal by one project changes the cost allocation for the others, withdrawals cascade, forcing restudies that lengthen the queue for everyone remaining.
The numbers
Berkeley Lab compiles this data across all seven ISOs and RTOs plus 50 non-ISO utilities, which together represent about 98% of installed US generating capacity. The 2026 edition, published in June 2026 with data through the end of 2025, reports the following.
| Measure | Value (end of 2025) |
|---|---|
| Active projects | ~8,200 |
| Generation capacity in queue | 1,312 GW |
| Storage capacity in queue | ~749 GW |
| Total active queue | Over 2,060 GW |
| Change vs prior year | Down 10% |
| Holding a signed or draft interconnection agreement, not yet operating | 549 GW |
| Median request to commercial operation (projects built 2025) | Over 5 years |
| Completion rate, 2000–2020 requests | 13% |
The completion figure deserves emphasis because it is so often misread. Of the capacity that entered queues between 2000 and 2020, only 13% had reached commercial operation by the end of 2025. Seventy-five per cent had been withdrawn, and 10% was still active — some of it after more than a decade.
Announced capacity and connected capacity differ by roughly a factor of eight. Any forecast built on queue volume is forecasting a number that historically shrinks by 87%.
The 549 GW figure is the more actionable one. That capacity has completed the studies and holds a draft or executed interconnection agreement, and it still is not operating — 256 GW of solar, 161 GW of storage, 76 GW of wind and 45 GW of gas. Those projects have cleared the process and are held up by construction, equipment, financing or transmission build-out. It is the closest thing to a pipeline of capacity that will genuinely arrive.
The composition shift
The most interesting movement in the 2025 data is not the total, which fell 10%. It is what grew while everything else shrank.
Active natural gas capacity in queues rose to 253 GW, an increase of 86% in a single year. Over the same period solar fell 19% to 773 GW, storage fell 16% to 749 GW, and wind fell 19% to 220 GW.
Read that alongside the IEA's projection and the picture is consistent. In its analysis of energy supply for AI, the IEA expects renewables to meet "nearly half of the additional demand, followed by natural gas and coal, with nuclear starting to play an increasingly important role towards the end of this decade" — while noting that in its higher-growth Lift-Off Case, "long grid connection queues mean that most of the additional increase" beyond 2030 is met by fossil fuels.
The mechanism connecting those observations is dispatchability and siting flexibility. A load that must run continuously wants generation that can run continuously, sited where the load is. When the queue is the constraint and the demand is urgent, the projects that advance are the ones whose economics survive a long wait and whose output matches a 24-hour industrial profile. That is a structural pressure, not a policy preference, and it operates regardless of anyone's procurement commitments.
How the queues became congested
The backlog is not the result of a single failure, and understanding how it accumulated explains why it is slow to clear.
Entering a queue was, for a long time, cheap relative to the option it created. A developer could submit a request with modest commitment, hold a position while assessing the project, and withdraw at little cost if the numbers did not work. Rationally, developers submitted more requests than they intended to build, and submitted them earlier.
That behaviour interacts badly with sequential, interdependent studies. Each application had to be studied in relation to the ones ahead of it, so speculative applications imposed real analytical cost on serious ones. When a speculative project withdrew, the cost allocation for everything behind it changed, requiring restudy. The queue was not merely long; it was self-lengthening, and its length was substantially composed of projects that were never going to be built.
This is why the reforms focus on readiness requirements and cluster studies rather than on adding staff. Studying projects in groups removes the sequential dependency, and making a queue position expensive removes the incentive to hold one speculatively. Whether the effect is large enough is the open question the data cannot yet answer.
The workarounds, and what they cost
Operators facing a five-year wait have not simply waited, and the alternatives are worth understanding along with their limits.
Co-location with existing generation. Siting load next to a plant that is already connected sidesteps the queue for new generation entirely. It is the fastest route available, which is why sites near existing plants have repriced so sharply. The limit is supply: there are only so many such sites, and the regulatory treatment of arrangements that bypass the transmission system is genuinely contested rather than settled.
On-site generation. Building dedicated generation at the facility avoids the transmission queue, though not permitting, equipment lead times or fuel supply. It converts a queue problem into a construction problem, which is an improvement in predictability more than in speed.
Using existing headroom. Many connected sites have capacity that is not fully used, either because the original load has shrunk or because the connection was sized generously. Finding that headroom is unglamorous and it is the only option that delivers capacity in months rather than years.
Flexible load. Committing to reduce consumption when the system is stressed can secure a connection that firm load could not. For training workloads, which tolerate interruption far better than serving does, this is a genuine and under-used option — and one of the few places where the shape of the AI workload works in the operator's favour rather than against it.
What this does to compute siting
The consequence for AI infrastructure is that geography stops being a preference and becomes the primary design variable.
A five-year median from request to operation means that generation capacity for a data centre being planned now, if it depends on new generation, arrives around 2031. Chips ordered today arrive within the year. Those two clocks cannot be reconciled by working harder on the faster one, which is why capacity strategy has shifted toward places where the power already exists.
That produces the behaviour now visible across the industry: siting next to existing generation, co-locating with retiring or under-utilised plants, acquiring sites for their grid position rather than their land, and signing long-duration power agreements well before there is a building. A grid connection has become an asset with a value largely independent of what is built on it.
It also changes the internal economics of a deployment. As set out in how to cost an AI rack end to end, energy and facility overhead are terms in the cost model that operators partly control — but only if they have a site at all. Scarcity of connected sites turns a variable cost into a constraint on capacity, and constraints on capacity are what eventually reach customers. That transmission mechanism is the subject of why token pricing is a facilities question. UniverseBlend's survey of the hidden limits on AI compute covers the site-level version of the same problem.
Reform, and why it is too early to score
Regulators have not been idle, and the two most significant interventions are worth knowing by name.
FERC Order No. 2023, published in the Federal Register on 6 September 2023 as Improvements to Generator Interconnection Procedures and Agreements, targets the queue directly. It addresses "interconnection queue backlogs" and aims to "improve certainty, and prevent undue discrimination for new technologies", with the requirement that processes be "just, reasonable, and not unduly discriminatory or preferential". The central structural change is a shift from studying projects one at a time toward studying them in clusters, with readiness requirements and deadlines intended to deter speculative applications.
FERC Order No. 1920, published on 11 June 2024 as Building for the Future Through Electric Regional Transmission Planning and Cost Allocation, addresses the upstream problem: transmission planning and how its costs are allocated. Interconnection queues are congested partly because the network they connect to has not been expanded to accommodate them, and a queue reform that does not build wires eventually re-congests.
Berkeley Lab's own assessment is appropriately cautious: these reforms "are important measures to reduce interconnection bottlenecks and enhance grid system reliability, but it is too early to measure and assess their full impact". Note the mechanism by which improvement would show up — the 10% decline in queue volume during 2025 came from high withdrawal rates and fewer new requests, which is what a functioning readiness requirement should produce. Whether that is reform working or demand shifting cannot yet be distinguished from the data.
What actually arrives by 2030
Putting the demand and supply sides together gives a usable planning picture without requiring anyone to forecast.
On demand, the IEA estimates data centre electricity consumption at around 415 TWh in 2024 — about 1.5% of global consumption — rising to roughly 945 TWh by 2030, just under 3% of global demand. Consumption has grown around 12% a year since 2017, more than four times faster than total electricity consumption, and accelerated servers are projected to grow at 30% a year against 9% for conventional servers.
On supply, the IEA puts generation for data centres at 460 TWh in 2024, rising to over 1,000 TWh in 2030 and 1,300 TWh in 2035 in its Base Case, with a Lift-Off Case reaching nearly 2,000 TWh by 2035. Small modular reactors are expected to contribute meaningfully only after 2030.
The reconciliation is the 549 GW already holding interconnection agreements, plus efficiency. Every point of PUE improvement is capacity that does not need a queue position at all — which is the strongest practical argument for the efficiency work covered in the rack costing article, quite apart from its effect on any single operator's bill.
What follows for anyone planning capacity
- Treat announced capacity as an option, not a plan. The historical conversion rate from queue entry to operation is 13%.
- Track interconnection agreements, not queue entries. The 549 GW with signed agreements is a far better predictor than the 2,060 GW total.
- Site for the connection. Existing capacity beats new capacity by roughly five years, and that gap is not closing quickly.
- Price power as a long-term contract. A twelve-month view of electricity prices is the wrong instrument for an asset with a multi-year life.
- Count efficiency as capacity. Reducing consumption is the only capacity increase available on a software timescale.
Frequently asked questions
How long does it take to connect a power plant to the grid?
For projects that reached commercial operation in 2025, the median duration from interconnection request to operation exceeded five years in the regions with available data, and Berkeley Lab reports that duration is increasing. That is the median for projects that succeeded; the majority of requests are withdrawn before reaching operation at all.
How much capacity is waiting in US interconnection queues?
More than 2,060 GW at the end of 2025 — about 8,200 projects, comprising 1,312 GW of generation and roughly 749 GW of storage. Total active volume fell 10% from the prior year, driven by high withdrawal rates and fewer new requests.
Why do most queued projects never get built?
Because the cost of required network upgrades is only established at the end of the study process, and it frequently makes a project uneconomic. Withdrawals then change the allocation for remaining projects, triggering restudies. Of capacity requesting interconnection from 2000 to 2020, 13% reached operation and 75% was withdrawn.
Does the queue explain why data centres are being built next to power plants?
It is a large part of it. If new generation takes over five years to connect and has an unfavourable completion rate, then existing connected capacity is worth a substantial premium. Co-location, site acquisition for grid position, and long-term power agreements are all rational responses to that arithmetic.
Has FERC Order 2023 fixed the problem?
Too early to tell, and Berkeley Lab says so explicitly. The reforms move toward cluster studies and readiness requirements, which should reduce speculative applications. Queue volume did fall 10% in 2025, but distinguishing reform effects from shifting demand is not yet possible from the published data.
Final takeaway
Where large-scale compute exists at the end of this decade is being decided now, by a process most people in the industry have never read about, in proceedings that predate the current build-out by decades.
The numbers are unusually clear for an infrastructure question: 2,060 GW queued, five years median to connect, 13% historical completion. Those three figures explain the siting behaviour, the power agreements, and the sudden strategic value of unremarkable land near a substation better than any narrative about chip supply. Chips are the fast part of this system. The queue is the slow part, and the slow part sets the pace.
Sources and further reading
- Lawrence Berkeley National Laboratory, Queued Up: 2026 Edition — interconnection queue data through the end of 2025
- FERC Order No. 2023, Improvements to Generator Interconnection Procedures and Agreements, September 2023
- FERC Order No. 1920, Building for the Future Through Electric Regional Transmission Planning and Cost Allocation, June 2024
- IEA, Energy and AI — energy supply for AI, generation mix and grid connection queues
- IEA, Energy and AI — executive summary, data centre consumption and projections
