The chips exist and the money exists — the electrons don't arrive on time. About half of the 12 GW of US data-center capacity planned for 2026 is delayed or cancelled, PJM wholesale prices ran +76% year over year in Q1, and 2,061 GW of generation sits stuck in interconnection queues. Power is now the binding constraint on the AI buildout.
Green means watching — we are tracking it and it looks normal. Amber means warning — it needs a closer look. Red means danger — time to act. The color never carries the meaning alone: every card says it in words.
Regime: load-growth supercycle, supply stuck in paperwork. After 15 years of flat US electricity demand, consumption has risen about 2% a year for five years — and EIA expects record generation of 4,368 TWh in 2026, up 2.2%, with another 1.7% in 2027. Data-center servers were 7% of commercial electricity use in 2025; EIA's outlook puts them at 22–33% by 2050.
The new capacity meant to serve that load is overwhelmingly trapped: 2,061 GW across ~8,200 projects, with a median 61-month journey from interconnection request to commercial operation — up from 22 months in 2008. Only about 13% of capacity that entered queues from 2000–2020 ever reached operation.
So the hyperscalers stopped waiting for the grid. Thirteen nuclear deals totaling ~10 GW in 18 months — restarts, uprates (squeezing more megawatts out of existing reactors), SMR fleets, and this week's Google–Constellation $4.3B uprate package — amount to tech companies building their own power system in parallel to the public one.
While Big Tech tries to buy its way out through private nuclear deals, the rest of the buildout is crashing directly into the regulated public grid — and retail consumers are picking up the tab. PJM's July 2025 capacity auction (fees paid to generators just to stand ready) cleared at a record $329.17/MW-day — the maximum regulators allow — and that single number is driving 8–15% retail rate increases across the 13 states PJM serves starting this year. Separately, PJM's market monitor found wholesale energy prices (what's paid for electrons actually consumed) averaged $136.53/MWh in Q1 2026, up 76% year over year, with data centers driving 63% of net load growth in the zone. Transmission congestion costs alone hit $6B in the first half, up 43%.
Regulators are scrambling to assign the bill. On September 29, FERC accepted PJM's reliability backstop procurement plan but suspended it until February 2027 — protections against shifting data-center costs onto existing customers remain unresolved, and PJM's July auction left a 6,831 MW shortfall for 2028–2029. TVA's separate data-center rate took effect October 1: new large loads face a capacity commitment charge of about $1.5M per MW.
The political question underneath: does AI load pay its own way, or do ratepayers subsidize the buildout? Every 2026 docket is a version of that fight.
There is no national power market — there are seven regional ones, plus western states where utilities negotiate power sales directly rather than through a centralized clearinghouse. The AI load story looks completely different in each:
| Region | Load pressure | Power cost indicator | The bottleneck |
|---|---|---|---|
| PJM (Mid-Atlantic/Midwest) | 165,563 MW all-time summer peak, Jul 2 | $136.53/MWh Q1 wholesale avg, +76% YoY | hottest — capacity market at ceiling; ratepayer pushback slowing grid expansion |
| ERCOT (Texas) | ~10%/yr summer peak demand growth (projected) | $34.85/MWh 2026 avg (North hub) | two-speed — midday solar glut vs. evening ramping risk; local transmission constraints |
| TVA (Southeast) | Industrial pipeline surging; new data-center rate Oct 1 | $1.5M/MW large-load connection surcharge | repricing — growth pre-funds its own dedicated capacity |
| Non-ISO West | 567 GW in queue (largest US share) | Bilateral premiums ~$15–30/MWh for firm green power | pipeline — no central clearing; multi-state permitting across federal land |
The queue's fuel mix tells the same story in aggregate: solar 773 GW, storage 749 GW, wind 220 GW, gas 253 GW. The future grid is queued — it just can't get connected.
The Iansiti & Lakhani lens, applied to electrons.
Value created: genuine and enormous. Every gigawatt of firm power delivered to a data center unblocks tens of billions in AI capex. The nuclear restarts, the uprates, the SMR order books — these are real electrons solving a real bottleneck.
Value captured: asymmetric — and it is not accruing to the electrons.
The tollbooths: (1) Regulated utilities — special contracts like DTE–Google and TVA's $1.5M/MW charge convert AI load into rate-base growth with captive customers. (2) PPA counterparties — Constellation and Talen locked in 20-year revenues at prices set before the shortage was fully priced. (3) Equipment makers — 144-week transformer lead times are pricing power by another name. (4) Infrastructure funds — LS Power's $6B raise is a bet that buying existing plants beats building new ones when the queue takes five years.
What this means for you: don't just track AI chip stocks — track who gets paid when the power flows. In the near term, the scarcest assets aren't GPUs; they're grid connections, transformers, and signed PPAs. Own the tollbooths: regulated utilities with AI-load rate base, nuclear operators with contracted output, and the equipment makers with two-year backlogs. Rent merchant power volatility; own contracted independent power producers that locked in long-term hyperscaler off-take before regulators close the gate.