Article
← All posts
tech

2,000 Gigawatts Stuck in Line

August 3, 2026 — Darwin Frederic

Reading time: 8 minutes

TL;DR

  • Roughly 2,060 gigawatts of generation and storage were actively waiting in US grid interconnection queues at the end of 2025, per Lawrence Berkeley National Laboratory. The entire installed US power fleet is about 1,300 GW
  • The median project that reached commercial operation in 2025 spent more than five years in the queue. Two decades ago it was under two years
  • Historically, only 13% of the capacity that requests connection ever gets built. Three quarters gives up and withdraws
  • The equipment is a second wall behind the paperwork: high-power transformer lead times have stretched from 24 to 30 months before 2020 to as long as five years today
  • The result, per Bloomberg and Sightline Climate: of the roughly 12 GW of US data center capacity slated for 2026, only about a third is actually under construction, and nearly half of planned builds may be delayed or canceled
  • Hyperscalers are responding by going around the grid entirely: on-site turbines, fuel cells, and self-manufactured switchgear

1. The Line Is Longer Than the Grid

In an earlier article, The Power Wall, I covered the generation crunch: the gas turbines, the nuclear deals, the raw supply of electrons. This article is about the less famous half of the problem. Even when the electrons exist, getting permission and equipment to move them is now its own multi-year constraint.

Start with the headline number from Lawrence Berkeley National Laboratory’s Queued Up report, the definitive annual census of the US interconnection backlog. At the end of 2025, roughly 8,200 projects were actively seeking permission to connect to the US grid, representing 1,312 GW of generation and 749 GW of storage. That is about 2,060 GW in line, versus roughly 1,300 GW of everything currently running in America.

The queue is not a line in any normal sense. Before a power plant or large load can connect, grid operators must run impact studies to figure out what upgrades the connection requires and who pays. Those studies are slow, sequential, and frequently restarted when other projects drop out, and projects drop out constantly.

Interestingly, the queue actually shrank in 2025, down about 10%. That was not because projects got connected. It was because a record wave of them gave up and withdrew, while gas projects surged 86% as developers chased the data center boom.

2. Five Years, 13%

Two LBNL statistics tell the whole story.

First, the wait. For projects that reached commercial operation in 2025, the median time from interconnection request to switching on exceeded five years. For projects built between 2000 and 2007, it was under two.

Second, the survival rate. Of all the capacity that requested interconnection between 2000 and 2020, just 13% had reached commercial operation by the end of 2025. 75% withdrew. The queue is less a pipeline than a filter that discards most of what enters it.

For AI, the mismatch is brutal. An AI deployment cycle runs 18 months or less. The grid connection process runs five years or more. Per Bloomberg’s reporting on Sightline Climate data, about 12 GW of US data center capacity was slated to come online in 2026, but only about one third is under active construction, and close to half of planned US builds may be delayed or canceled. The industry announced more than $650 billion in 2026 AI infrastructure spending, and a meaningful fraction of it is currently waiting on paperwork and hardware.

The Power Chain, and Where It Chokes

  FROM ELECTRON TO GPU: THE DELIVERY STACK
  ──────────────────────────────────────────────

  [ GENERATION ]      turbines, nuclear, solar
       │              (covered in The Power Wall)
       ▼
  [ INTERCONNECTION ] ◄── CHOKE POINT 1
       │              ~2,060 GW waiting
       │              median wait: 5+ years
       │              survival rate: 13%
       ▼
  [ TRANSMISSION ]    lines + substations
       │
       ▼
  [ TRANSFORMERS ]  ◄── CHOKE POINT 2
       │              lead time pre-2020: 24-30 months
       │              lead time now: up to 5 years
       │              GSU demand since 2019: +274%
       ▼
  [ SWITCHGEAR &      breakers, batteries,
    DISTRIBUTION ]    site electrical
       │
       ▼
  [ DATA CENTER ]     <10% of project cost,
                      ~100% of the schedule risk

3. The Transformer Wall

Behind the administrative queue sits a physical one: the machines that step voltage up and down at every stage of the grid.

The numbers, per Wood Mackenzie’s 2025 supply analysis: demand for generator step-up transformers has grown 274% since 2019, and substation power transformers 116%. Standard power transformers averaged 32 months for delivery as of mid 2025, generator step-up units three years, with specialized orders stretching to four years. Sightline Climate, cited by Bloomberg, puts high-power units at up to five years, versus 24 to 30 months before 2020. Wood Mackenzie models a 30% supply shortfall for power transformers.

Three demand waves hit simultaneously: AI data centers, industrial and vehicle electrification, and the replacement cycle for aging equipment. That last one is bigger than most people realize. More than half of America’s roughly 40 million distribution transformers are already past their expected service life.

And the bottleneck has a bottleneck. Transformer cores require grain-oriented electrical steel, a specialized material with exactly one domestic producer, Cleveland-Cliffs. Every US transformer maker drawing on domestic steel draws from a single source. China remains the world’s largest producer of the electrical equipment category overall, which means the US buildout is exposed to tariffs and trade policy on top of everything else.

The industry likes to point out that electrical equipment is less than 10% of a data center’s cost. True, and irrelevant. A missing transformer stops a $2 billion project as completely as a missing building.

4. Bring Your Own Power

When the queue is five years and the transformer is four, the rational move is to stop asking. That is exactly what is happening.

The trend has a name in industry reports: BYOP, bring your own power. Per HSBC’s analysis of AI bottlenecks, the high voltage substation wait of 3 to 5 years is driving developers to generate on-site at medium voltage, skipping the biggest equipment entirely. The examples are piling up: Oracle’s multi-gigawatt Project Jupiter campus in New Mexico announced it will run on Bloom Energy fuel cells with no grid connection for primary power. Brookfield and Bloom signed a partnership worth up to $5 billion for AI factory fuel cells. Some developers, like Crusoe Energy, have started manufacturing their own switchgear rather than waiting in the vendor queue.

Some of these operators plan to connect to the grid eventually and treat on-site power as a bridge. Others intend to bypass it indefinitely. Either way, the meaning is the same: the grid connection process has become slow enough that billion-dollar projects find it cheaper to build their own miniature utility than to wait in line.

5. What Actually Fixes This

Three things are in motion, all on long fuses.

Process reform. FERC Order 2023 pushes grid operators from first-come-first-served study queues toward cluster studies with readiness requirements, designed to filter speculative projects earlier. LBNL notes the reforms are real but too new to measure. Process reform also cannot manufacture a transformer.

Factory capacity. Nearly $2 billion in North American transformer manufacturing expansion is underway: Siemens Energy’s large power transformer plant in Charlotte targets 2027 production, Hitachi Energy’s Virginia facility aims to be the largest in the US by 2028, plus expansions from GE Vernova, Eaton, ERMCO, and Hyundai Electric. As with memory fabs, the relief arrives in 2027 and 2028, and projects executing before then face the market as it is.

Load flexibility. The fastest lever is making data centers easier for grids to say yes to: curtailable loads, on-site batteries, and consuming power behind the meter. Every megawatt that does not require a new high voltage substation skips both chokepoints at once.

The Honest Summary

The queue reforms filter the line, the factories lengthen it more slowly, and the workarounds route around it. Nothing on the table shortens the line dramatically before 2028.

Sources

  1. Lawrence Berkeley National Laboratory, “Queued Up: 2026 Edition,” June 2026 (2,061 GW active at end of 2025: 1,312 GW generation, 749 GW storage; 8,200 projects; median 5+ year duration; 13% completion, 75% withdrawal; gas capacity +86%; 10% queue shrinkage)
  2. Lawrence Berkeley National Laboratory, “Queued Up: 2025 Edition,” December 2025 (~2,290 GW at end of 2024; ~55 month average queue duration; 700+ GW withdrawn in 2024)
  3. Bloomberg, citing Sightline Climate, March to April 2026 (12 GW planned 2026 US data center capacity, ~one third under construction; nearly half of planned builds delayed or canceled; transformer lead times from 24 to 30 months pre-2020 to as long as 5 years; $650B+ hyperscaler capex; via Tom’s Hardware and Yahoo Finance syndication)
  4. Wood Mackenzie transformer supply analysis, via POWER Magazine, January 2026 (GSU demand +274% since 2019; substation transformers +116%; 30% power transformer shortfall; 40 million distribution transformers past service life; ~$1.8B manufacturing expansion; Cleveland-Cliffs as sole domestic GOES producer)
  5. IndustrialSage / PwC via pv magazine USA, May 2026 (128-week average power transformer delivery, 144-week GSU, four-year specialized orders)
  6. HSBC AI bottlenecks report, via Tech Fund analysis, May 2026 (3 to 5 year high voltage substation lead times; BYOP trend)
  7. Company announcements: Oracle Project Jupiter with Bloom Energy fuel cells; Brookfield-Bloom partnership up to $5B (October 2025); Siemens Energy Charlotte plant (2027); Hitachi Energy Virginia facility (2028); Crusoe Energy in-house switchgear manufacturing
  8. FERC Order 2023 and LBNL commentary on interconnection reform implementation