Everyone's crowded into the obvious AI-power trade: nuclear plants, gas turbines, generation. They're one layer too high. The actual chokepoint is getting the electricity from the source into the chip — and that part has a 4-year waiting list almost nobody's pricing. Hover or tap any underlined term.
AI is electricity-hungry on a scale the grid has never seen. A single large data center campus can now demand 1–2 gigawatts — roughly a million homes' worth. So the market did the obvious thing: it bid up everything that makes electricity. Nuclear (CEG, OKLO, VST), gas turbines (GEV), the independent power producers. That trade is real — and it's also fully discovered. When a thesis is on the cover of every outlet, the easy money is already made.
Making electricity and delivering it are two completely different industries. Here's the plain-English split:
Generation = the power plant. Turns fuel or fission into electricity. (Crowded trade.)
Delivery = everything between the plant and the chip: high-voltage lines, switchgear, substations, transformers, and the power distribution inside the building. (Almost nobody's trade.)
You can flood the country with power plants. If the delivery layer can't keep up, none of it reaches a GPU. And right now, the delivery layer is the binding constraint — for a reason most investors have never had to think about.
Large grid transformers are custom-built, one at a time, from materials only a handful of factories produce. Demand — from AI data centers, grid upgrades, and electrification all at once — has blown lead times out to about 128 weeks (~2.5 years) on average, with the largest units running up to 4 years (Wood Mackenzie 2025; PwC). US transformer demand is up ~119% since 2019, and roughly 80% of large units are imported. This isn't a rumor; it's the single most-cited physical constraint by the utilities and hyperscalers actually building.
And getting a finished transformer is only half the battle — you still have to plug the project in. The US grid-interconnection queue has swelled to roughly 2,600 gigawatts — more than double the entire existing US power capacity — with median waits to come online approaching 5 years, and some data-center projects facing far longer (Lawrence Berkeley National Lab). Capital is everywhere; the wait is what's scarce.
While we're talking about things hiding in plain sight — the AI headline is GPUs (NVDA). But every GPU is starved without HBM (high-bandwidth memory), and that's a three-player oligopoly (Micron, SK Hynix, Samsung) reportedly sold out into 2027. "Memory" sounds boring and cyclical, so generalists skip it — which is exactly why it's worth watching. MU is the cleanest US-listed way in.
| Name | Ticker | The role |
|---|---|---|
| Eaton | ETN | Electrical distribution, switchgear, transformers — the broadest delivery-layer pure-play. |
| Quanta Services | PWR | The contractor that physically builds transmission + connects projects to the grid. Owns the interconnect bottleneck. |
| GE Vernova | GEV | Both sides — gas turbines (generation) and grid/electrification gear (delivery). |
| Vertiv | VRT | Power distribution + liquid cooling inside the data center. The "last 10 feet." |
| Micron | MU | The memory chokepoint — HBM sold out into 2027. |
| nVent | NVT | Electrical connection + cooling enclosures — smaller, more concentrated bet. |
None of that kills the thesis — it sizes it. The point stands: when a trade gets crowded, look one layer down the supply chain for the part that physically can't scale. Right now, that's the wire, not the watt.
Same discipline as our orbital-compute scenarios: not one prediction — dated tracks with markers, because the trade's exit matters as much as its entry.
| Scenario | Bottleneck binding until | What has to be true | You'll know when… |
|---|---|---|---|
| Long squeeze | ~2032+ | Factory expansion stays slow (skilled labor + grain-oriented electrical steel are their own bottlenecks); AI buildout keeps compounding; permitting reform stalls. | Lead times still >3yrs in 2028; backlog growth keeps outpacing capacity adds. |
| Base case | ~2029–2031 | Announced factory expansions land 2027–28; demand keeps growing but supply finally compounds faster; premium pricing fades gradually, volumes stay strong. | Lead times plateau 2027, shrink visibly by 2029; equipment-maker margins peak before volumes do. |
| Fast unwind | ~2027–28 | An AI-capex bust (see the malinvestment scorecard) cuts data-center demand just as new factory supply arrives — the classic cyclical double-hit. | Hyperscaler capex guidance falls; cancelled interconnection requests rise; backlogs get revised down. |
A 4-year queue doesn't make demand wait politely — it makes demand route around the queue. Each bypass is its own investable lane, and several are earlier (less crowded) than the equipment names:
| The problem | The bypass it's creating | How to watch it |
|---|---|---|
| Can't connect to the grid for years | On-site / behind-the-meter generation — gas turbines at the data center, and the nuclear-SMR pipeline as its long game | Gas-turbine order books (GEV again — both sides of the trade), SMR names, utility PPA announcements |
| New plants take a decade to permit and build | Nuclear-in-a-can: factory-built microreactors. Radiant's Kaleidos — 1.2 MW in a shipping container, air-cooled (no water), 5 years per fueling — targets first criticality at Idaho National Lab July 4, 2026, with a Tennessee factory built to ship 50 reactors/yr from 2028. The anchor customer says everything: Equinix pre-ordered 20 for data centers. Power becomes a manufactured product, not a construction project. | The July 2026 criticality test; factory output rate; who orders after Equinix |
| The grid's spare capacity is stranded in homes | Distributed residential compute. The average US home uses only ~40% of its electrical service (~80 idle amps). Nvidia + Span's XFRA nodes (16 Blackwell GPUs, the size of an AC unit) mount on houses: the homeowner pays ~$150/mo and gets full electricity + internet bills covered; Span sells the compute. Targeting 1 GW/yr from 2027 — harvesting grid capacity that's already built and connected. | The 100-home trial (2026); noise/liability complaints vs expansion; homebuilder partnerships (PulteGroup already in) |
| New transformers take 4 years | Refurbish, rewind, and life-extend old units — the "used car market" of the grid, plus standardized/modular transformer designs | Refurb specialists' pricing; DOE standardization push |
| Can't build new lines fast | Grid-enhancing technologies — dynamic line ratings, advanced reconductoring, power-flow software that squeezes 20–40% more from EXISTING wires with no new permits | The quiet software layer; mostly private/small-cap — the least crowded shelf on this entire map |
| Peak demand breaks the local grid | Batteries as grid infrastructure (BESS) — which feeds straight into the lithium chain we map separately | Storage attach-rates on new DC builds |
| Eventually: the grid itself is the constraint | Compute leaves the grid — undersea (China, live now) and orbital (AI1, attempting 2027+) | The relief valve is real but arrives AFTER this thesis's 2026–2030 window — sequence, not contradiction |
Every bypass above routes around the shortage. These would erase it — the "we don't have the materials yet, but it's physically allowed" tier. Clearly speculative, and included for a reason: the prize is enormous, and the first credible breakthrough in any of them is the signal that the scarcity premium has a ceiling.
Room-temperature superconductors moonshot — the holy grail. Zero-resistance wire would gut transmission losses, shrink transformers, and multiply line capacity — dissolving most of the delivery bottleneck at once. The world wants it so badly that every rumored candidate goes viral. Not here, repeated false alarms — but civilization-scale if it ever lands.
Solid-state transformers emerging — replace the steel-and-copper unit (the 4-year backlog, the scarce grain-oriented steel) with power-electronics. Smaller, faster to build, far more controllable. Expensive and early today — at scale it could break the very factory/steel constraint that makes transformers so slow.
Beamed / wireless power moonshot — orbital solar microwaved to the ground, or point-to-point transmission, sidesteps the wire entirely. Decades off and lossy, but it's the ultimate route-around-the-grid — and it ties straight into the orbital-compute thesis.
Dragonfly Lens looks one layer below the headline — where the real scarcity (and the real edge) tends to hide. Plain English, every claim sourced and flagged.
Join the Lens →Sources: transformer lead times — Wood Mackenzie 2025 transformer survey (~128 weeks) and PwC (up to 4 years), via pv magazine USA and POWER Magazine; interconnection queue (~2,600 GW, ~5-year median wait) — Lawrence Berkeley National Laboratory; June 2026 status (128/144-week lead times, GSU demand +274% since 2019, >50% of data centers may be delayed, on-site-generation bypass) — pv magazine USA, Energy News Beat, Smartland Energy; equipment-maker backlog $1.98B (+157% YoY, Mar 2026) — Forgent Power 8-K (SEC); Radiant Kaleidos (1.2MW container microreactor, Jul 2026 criticality target, 50/yr Tennessee factory, Equinix 20-unit preorder) — US Department of Energy, Radiant/Equinix announcement, Radiant ($300M Series D); Nvidia/Span XFRA residential compute nodes (bills covered, 1 GW/yr target 2027) — CNBC, Fortune.
Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Facts as of June 2026, drawn from the public reporting and industry sources cited above — verify against primary sources before acting. Tickers are named to illustrate the supply chain, not as buy recommendations. Past performance does not guarantee future results.