Analysis · EIA · NRC · DOE · company filings & PPAs

Powering around the queue: how AI is buying its way past the grid

If the interconnection queue is a five-year line, the fastest way through it is not to stand in it. Three workarounds are already reshaping who gets power — and when.

A Gridlas analysis · built from public data · figures 2024–2026, primary sources cited throughout

Our companion analysis made the case that the real limit on the AI buildout isn't generation — it's the interconnection queue, where the median wait has doubled to about five years and roughly 77% of proposed capacity is eventually withdrawn. So the hyperscalers stopped waiting. Faced with a queue measured in years and GPUs measured in months, they are routing around it three ways: burn gas on-site now, restart the reactors we already shut down, and bet on small modular reactors for the 2030s. Here is what each route actually delivers, in megawatts and dates, from the public record.

1. The five-year wait created a workaround market

The logic is simple arithmetic. A new grid interconnection now takes about five years at the median — and that is if the project survives the queue at all. A GPU cluster depreciates on a two-to-three-year clock. When the power path is slower than the asset it powers, operators pay almost anything to shorten it. That gap is the entire market for the workarounds below, and it explains why each one trades a premium — higher cost, more emissions, or execution risk — for one thing: time to power.

0 yr 2 yr 4 yr 6 yr 8 yr On-site gas turbines already on site On-site gas (existing turbines): power in months to ~1.5 years ~1 yr Nuclear restart TMI, Duane Arnold Nuclear restart: ~3–4 years from deal to grid ~3–4 yr New grid interconnection median, LBNL New grid interconnection: ~5-year median wait (LBNL) ~5 yr New nuclear / SMR first units ~2030+ New nuclear or SMR: ~6+ years — first units around 2030 at the earliest 2030+
Rough time from decision to delivered power, by route. On-site gas is fastest only when turbines are already in hand; restarts and new build are multi-year. Sources: LBNL "Queued Up" (grid median); company restart timelines (Constellation, NextEra); IEA "Energy and AI" (SMR timing).
The catch on "one year" — On-site gas is fast only if you already have turbines. New orders now hit a wall: GE Vernova ended 2025 with an ~80 GW gas-turbine backlog stretching into 2029, and expects reservations sold out through 2030 by the end of 2026 — 3–4 year lead times on a new machine. The fast lane fills up too.

2. Route one — burn gas on-site, now

The bluntest workaround is to build your own power plant behind the meter and skip the queue entirely. The signature case is xAI's Colossus in Memphis: with only an 8 MW grid connection at first, xAI wheeled in 35 gas turbines (~422 MW) — more than a conventional peaker — to energize the site while the utility caught up. Aerial imagery caught the full count; the site was permitted for just 15, drawing a Clean Air Act notice from the Southern Environmental Law Center, and xAI has since pared back and permitted the remainder. A second site in Southaven, Mississippi added 27 more turbines (~495 MW).

The utilities are doing the same thing at far larger scale. For Meta's Richland Parish campus in Louisiana, Entergy broke ground on 2.26 GW of new gas generation dedicated to a single customer — regulator-approved, online from 2028. Behind-the-meter turbines or utility-built plants, the move is identical: attach firm gas to one campus rather than wait in the shared line.

0 800 1600 2400 xAI Colossus 1 Memphis TN · 35 turbines xAI Colossus 1, Memphis: ~422 MW from 35 on-site gas turbines (later permitted to 15) ~422 MW xAI Colossus 2 Southaven MS · 27 turbines xAI Colossus 2, Southaven MS: up to ~495 MW from 27 on-site gas turbines ~495 MW Meta · Richland Parish LA · Entergy, online 2028 Meta Richland Parish, Louisiana: 2.26 GW of new Entergy gas generation, online from 2028 2,260 MW
Dedicated gas generation built for single AI campuses (MW). xAI's turbines are behind-the-meter and running; Meta's is utility-built, front-of-the-meter, online from 2028 — different mechanisms, same idea. Sources: DCD, SELC, Governors' Wind Energy Coalition (xAI); DCD, POWER, S&P Global (Meta/Entergy).
The queue rations connections. On-site gas opts out of the queue — and, for now, out of the wait. The trade is carbon and controversy for speed.

The speed is real, and so is the cost. On-site gas re-carbonizes the near-term buildout exactly as utilities were retiring coal, and the Memphis fight shows the permitting and environmental-justice backlash it invites. It is a bridge, not a destination — which is why the same companies are simultaneously buying the cleanest firm power on the board.

3. Route two — restart the reactors we already shut down

The most striking signal that AI has changed the power market: plants the market had left for dead are coming back, paid for by hyperscaler contracts. Constellation is restarting Three Mile Island Unit 1 — rechristened the Crane Clean Energy Center — on the strength of a 20-year Microsoft PPA, targeting 835 MW by 2028 at roughly $1.6B of capex. In Michigan, Holtec's 800 MW Palisades plant became the first U.S. reactor ever approved to return from decommissioning, backed by DOE loans. And Google is funding NextEra's restart of Iowa's ~615 MW Duane Arnold under a 25-year PPA, targeting early 2029.

Where a full restart isn't on the table, the deals reallocate output from existing reactors. Amazon's expanded, 17-year, $18B agreement with Talen draws up to 1,920 MW from the Susquehanna plant; Meta signed for the full output of Constellation's Clinton plant in Illinois. FERC has pushed back on the pure behind-the-meter version of these co-location deals — rejecting an expanded Susquehanna interconnection in a 2024 order — which is exactly why Talen and Amazon restructured into a grid-connected retail PPA.

0 500 1000 1500 2000 Susquehanna Amazon · Talen PPA Susquehanna (PA): up to 1,920 MW to Amazon under a 17-yr Talen PPA — existing plant 1,920 MW Crane / TMI Microsoft · Constellation Crane Clean Energy Center / Three Mile Island Unit 1: 835 MW restart, 20-yr Microsoft PPA, target 2028 835 MW Palisades Holtec · DOE-backed Palisades (MI): 800 MW restart, NRC-approved — first US reactor brought back from decommissioning 800 MW Duane Arnold Google · NextEra Duane Arnold (IA): ~615 MW restart, 25-yr Google PPA, target early 2029 615 MW Kairos Power Google Google–Kairos Power: up to 500 MW of SMRs, first unit ~2030, fleet by 2035 ≤500 MW Oklo Aurora Equinix Oklo Aurora powerhouses: 500 MW to Equinix, early 2030s 500 MW X-energy Amazon Amazon–X-energy: 300 MW of XE-100 SMR capacity (Pacific NW & Virginia), early 2030s 300 MW Existing-plant PPA Restart New SMR (2030+)
Nuclear capacity contracted or restarting to serve AI data centers, by type (MW). Restarts and existing-plant PPAs deliver this decade; SMRs are announced, with first units targeted around 2030 or later. "Up to" figures reflect phased or fleet totals. Sources: Utility Dive, World Nuclear News, ANS, NRC, company releases.
835 MW
Three Mile Island restart for Microsoft — target 2028
1,920 MW
Susquehanna nuclear to Amazon (Talen, 17-yr PPA)
~80 GW
GE Vernova gas-turbine backlog into 2029 — the fast lane, filling

4. Route three — the SMR bet: real money, later power

The third route is the one everyone announces and no one has yet plugged in. Small modular reactors are the hyperscalers' hedge for the 2030s: Amazon led a $700M round in X-energy and contracted for ~300 MW of its XE-100 design; Google agreed with Kairos Power for up to 500 MW across seven units, first online targeted for 2030; and Oklo signed Equinix for 500 MW of its Aurora powerhouses. The dollars are committed and the ambition is real — Oklo alone claims agreements totaling ~12 GW through 2044.

But the honest caveat is the timeline. Not one commercial SMR is operating in the U.S. today. First units are targeted for the end of this decade, and first-of-a-kind nuclear projects have a long history of slipping. SMRs answer the 2030s demand curve, not the cluster a hyperscaler is energizing this year — which is why the same companies are burning gas and restarting old reactors in the meantime.

Announced ≠ online — The nuclear chart mixes three very different certainties. Restarts (TMI, Palisades, Duane Arnold) are funded and NRC-track, delivering 2026–2029. Existing-plant PPAs (Susquehanna, Clinton) reallocate power that already exists — they add contracts, not electrons. SMRs are pre-construction announcements with 2030+ targets. Treat the dates, not just the megawatts.

5. What the workaround actually changes

Put the three routes together and a pattern emerges. The buildout isn't waiting for grid reform — it's pricing around it, and the bill comes in three currencies. Gas buys speed with carbon and local backlash. Nuclear restarts buy clean firm power but only for those who can sign billion-dollar, decade-long PPAs — concentrating the scarcest resource with the deepest pockets. And SMRs buy optionality on a timeline that may slip. None of them make the queue faster for everyone else.

And notice that every fast lane is itself congesting: turbine slots sold out through 2030, NRC restart reviews measured in years, SMR supply chains yet to be built. The workarounds relieve the pressure on the queue precisely by creating new queues of their own. The underlying bottleneck — too few ways to connect firm power where the load is landing — is what all three are really paying to avoid.

The AI buildout has decided it can't wait for the grid. What it's building instead — gas, restarts, reactors — is the real map of where U.S. power is heading.

The full picture, mapped. The Gridlas report puts the demand forecasts, the interconnection queue, and five regional deep-dives onto the grid — with high-res maps and the underlying dataset (CSV/GeoJSON), built entirely from public data.

Get the report →
Sources: See also the grid-bottleneck analysis, the water problem, the full methodology & sources, and the regional breakdowns.
This is the supply-side companion to the grid bottleneck. The visual analysis plots demand over the U.S. grid, and the regional pages break it down market by market.

Gridlas · independent & unaffiliated · built from public data.