Why do wires take 15 years when chips take 18 months?
Transmission is the slowest layer in any modern economy. AI's planning horizon is colliding with permitting reality, and so far the permitting reality is winning.
In 2026, the binding constraint on AI scale-up is not chips, not capital, not models — it is high-voltage transmission and substation capacity. Generation can be added in months. Wires can not.
The interconnection queue, in one number
As of mid-2026, US grids have roughly 2,600 GW of generation capacity sitting in interconnection queues — projects that have applied to connect to the grid and are waiting. The total operating US generation fleet is about 1,200 GW. So the queue is more than twice the existing fleet. Average wait time from queue entry to interconnection: 5+ years. In PJM (the mid-Atlantic grid covering Virginia and Northern Virginia data-center alley), waits exceed 7 years.
The queue is not full of fake projects. It is full of real generation — solar, wind, batteries, gas — that wants to come online and cannot, because the wires to move that energy to load centres do not exist and cannot be built fast enough. That is a different problem from "we don't have enough generation." We have plenty of generation. We don't have the wires.
Why HVDC takes so long
A new high-voltage transmission line — say, a 500-kV interstate — passes through five doors before it operates:
- Planning and routing (1–3 years). Where the line goes determines who fights it.
- Federal environmental review (NEPA) (2–5 years). Required for any line crossing federal land or affecting protected species. The 2023 NEPA reforms compressed but did not eliminate this.
- Eminent domain and state-level permitting (1–4 years, often parallel). Multi-state lines need approval from each state PUC the line crosses; one holdout state can block the whole project.
- Equipment manufacturing and lead time (2–4 years). Large transformers are built to spec by a small global supplier base — Hyundai Heavy, ABB, Siemens. The current lead time on a 500-kV transformer is approximately 3 years, up from 12 months in 2020.
- Construction (1–2 years). The fast part. Once everything else is settled, the actual stringing of wire is fast.
Total: 10–15 years from concept to commercial operation. That is the number that does not move on AI's planning horizon.
The chips-vs-wires asymmetry
For comparison, the time from a new GPU architecture entering tape-out to commercial availability is roughly 18 months. From announcement of a new fab capacity expansion to first wafer is 2–3 years. From announcement of a new data-center building shell to first server racked is 2–3 years.
Every layer of the AI stack moves on a 1–3 year cycle except transmission, which moves on a 10–15 year cycle. Compounded over a full 2025–2035 buildout, this means the layer that determines the maximum capability frontier is the wires layer. Whatever you can plug in by 2030 was effectively decided by what was already in the transmission queue in 2024.
What is being tried
- FERC Order 2023 (2023). Reformed the interconnection queue process to use cluster studies and "first-ready, first-served" instead of strictly first-come-first-served. Compressed nominal wait times by 1–2 years where applied; uneven adoption across regions.
- The 2023 NEPA reforms. Capped federal environmental review at 2 years and 150 pages for most categories. Some categorical exclusions for transmission upgrades. Marginal but real.
- Behind-the-meter buildouts (the practical answer). Skip the grid for the first 1–3 years of cluster operation; let on-site gas and fuel cells carry the load while the interconnection queue clears.
- Reconductoring with HTLS conductors. Replacing existing line conductors with high-temperature low-sag designs increases transmission capacity 50–100% on existing rights-of-way, with no new permitting required. Underbuilt in the US relative to potential.
- HVDC links across constrained regions. Direct high-voltage DC links from generation-rich areas to load centres bypass intermediate AC capacity. The SunZia line (2025), Champlain Hudson (2026), and Pacific Connector are the named projects.
- Federal authority preemption over state permitting, on the model of natural-gas interstate pipeline rules. Discussed; not yet enacted.
The strategic punchline
"Power" as the AI bottleneck is shorthand for two distinct problems with very different solve times. Generation — adding gigawatts of new electricity production — is solvable in 1–3 years with the right contract structure. Transmission — moving those gigawatts to where the chips are — is solvable in 10–15 years on the current process.
That is why the actual bottleneck inside the bottleneck is wires, not watts. The hyperscalers that win the 2027–2030 capacity race will be the ones whose wire planning was done in 2022–2024. The ones that did not start early enough will pay it back in delayed clusters or behind-the-meter premiums.
And once the wires problem is solved — which it eventually will be, by some mix of the mechanisms above — the binding constraint will rotate again. Back to chips, back to packaging, back to model architecture, back to data quality. That is the loop. The only durable edge is knowing where it sits today and being three years early on where it sits next.