Energy & Power

This rung asks where the electricity for AI comes from and what it costs: generation (gas, nuclear, renewables), power-purchase deals, the queues to connect a site to the grid, and the long wait for transformers and switchgear. The buildings themselves belong to datacenters, and the cost of a token belongs to inference-economics. This rung is the feed into both.

What this rung prices

A chip that cannot be switched on serves no tokens. So before any question of what a token costs, there is a question of how many chips can be powered at all, and that number is set here. The rung's sources speak to the how many side more than the how much side. They count power that has been promised but not delivered, measure how long a new plant waits to connect (a median of about five years), and show a state government stopping the queue. The rung now holds one price: PJM's capacity auction, the yearly price for power plants promising to be there at peak, cleared at $325 per MW-day for 2028/29, at its agreed cap (PJM 2028/29). That is roughly $119,000 per MW-year (our arithmetic). It is a price for availability, not for the electricity used, and none of the sources gives a price per megawatt-hour for a data centre or says how much reaches a token price.

The load-bearing findings

• About 12 GW of US data-centre power was announced for 2026, and about 5 GW was being built. That leaves a gap of about 7 GW, with only about a third of announced 2026 sites having broken ground. The same source puts AI data-centre load at about 10 GW by the end of 2026, and shows one utility (AEP) with 24 GW of committed demand by 2030 against a raw incoming queue of 190 GW (7 GW shortfall). This is the weakest source on the rung: it was ingested from a search snippet because the page could not be fetched, and it is flagged for a full re-read.

• In the biggest US load-growth market, the queue was stopped by a governor, not by physics. On 2026-08-03 ERCOT said Governor Abbott had directed it to verify data-centre "Large Loads" before moving them through interconnection, that it would miss its 2026-08-07 deadline for classifying those loads, and that it would ask the Texas utility commission for a good-cause exception ahead of the 2026-08-20 meeting (ERCOT notice). The notice does not say how much load is affected or how long the check takes, and a missed deadline is not a cancellation. What it does undo is the easy assumption that if connecting is slow elsewhere, you build in Texas.

• The federal forecast absorbed the Texas pause and still named that region the leader. The EIA's September 2026 outlook forecasts US generation of 4,368 billion kWh in 2026 (+2.2%) and +1.7% in 2027, and says that "despite a pause in new data center projects in Texas", the West South Central region remains the largest contributor to growth in electricity sales, driven by data centres and manufacturing (EIA STEO). The national growth figure is modest; the regional concentration is the point. The release gives energy, not gigawatts of data-centre load, and no data-centre share of demand.

• Power is a binding constraint, but no longer the only one. The concept page carried over from datacenters sets out the case: 36+ projects worth $162B stalled or delayed, transformers and switchgear as the long-lead items, and a 3–4 year wait to connect in data-centre load zones in PJM and ERCOT, where the queue holds enough generation "on paper" and the shortage is in connecting it. It also records that advanced packaging and memory were binding at the same time in mid-2026. See Power: The Binding Constraint.

• The national queue is about 2,000 GW on paper, and most of it will never be built. At the end of 2025 about 8,200 projects, 1,312 GW of generation and about 749 GW of storage, were waiting to connect to the US grid. The queue shrank 10% in a year, but gas rose 86% to 253 GW, the only large type that grew. The median project built in 2025 took 61 months from request to switching on (22 months in 2008), only 2% of the 2024 queue came online in 2025, and only 13% of what was requested in 2000–2020 was ever built. The count excludes data-centre load queues, so it measures supply waiting, not AI demand (LBNL Queued Up 2026). This census also corrects the Carbon Direct queue size carried on the concept page: LBNL counts ERCOT at 408 GW and PJM at 144 GW, well above that source's "more than 300 GW" for the two combined.

• PJM's grid, where data centres drive most of the demand growth, is short, and its price is pinned at a ceiling. PJM's capacity auctions ran at $329.17 (2026/27), $333.44 (2027/28) and $325 (2028/29) per MW-day, under a price collar agreed with the 13 state governors and the federal regulator; the last two cleared at the cap. The last two were also the first in PJM's history to come up short of the reliability requirement for the whole region, by 6,623 MW and 6,831 MW. PJM attributes nearly 5,100 MW of the 5,250 MW rise in its 2027/28 peak forecast to data centres (PJM 2028/29). The two earlier prices are kept as dated, superseded prints. A capped price understates scarcity: it says the market wanted at least this much.

Concept map

Concepts

Concept	Sources	Evidence	Last Updated
Power: The Binding Constraint	11 (0 papers)	Moderate to strong	2026-09-24

The concept page was carried over from datacenters and still cites mostly sources held there (links repointed 2026-09-24). It now folds in the LBNL census and the PJM price series, and records where they correct its earlier figures.

Entities

No entity pages exist on this rung yet.

What we do not know yet

• What the electricity costs. The PJM capacity price is a price for being available at peak, not for energy used. No source gives a data-centre electricity price, a PPA price, or how much of either passes through to a token. The EIA release carries a Henry Hub gas price assumption of $3.43/MMBtu, which is an input to its outlook, not a power price.
• What an uncapped PJM auction would have cleared at. The last two hit the cap; an uncapped figure appears only in secondary coverage and is not used here.
• How much load the Texas check affects, and what the commission decided on 2026-08-20. Not in the notice. A figure of about 474 GW of large-load queue in ERCOT is explicitly flagged in the source as unverified and is not used here (ERCOT notice).
• Gigawatts of data-centre demand from a primary source. The EIA gives energy, not load; LBNL excludes load queues; the 7 GW gap comes from a snippet-only analysis. PJM's ~5,100 MW of data-centre forecast growth (one region, one year) is the closest primary figure held.
• Most of the scope. Nuclear and gas generation deals, power arbitrage, energy ownership as a moat, and transformer lead times all lack a source on this rung.
• Standing: read it. Every number here is read from filings, forecasts or analysis.

Timeline

See timeline.md for chronological developments, and frontier.md for the open questions.

Sources

#	Title	Type	Date	Status
1	U.S. AI Data Center Delays: 7 GW Capacity Crisis (2026)	analysis	2026-Q1	compiled
2	ERCOT Market Notice M-A080326-01	technical-report	2026-08-03	compiled
3	EIA September 2026 Short-Term Energy Outlook	technical-report	2026-09-09	compiled
4	LBNL, Queued Up: 2026 Edition	technical-report	2026-06	compiled
5	PJM 2028/2029 Base Residual Auction results	technical-report	2026-07-14	compiled

Read next

• pjm-capacity-auction-2028-2029 — the rung's first price for power, and three auctions of it.
• lbnl-queued-up-2026-edition — the size of the wait, and how little of the queue gets built.
• ercot-batch-zero-timeline-pause-2026 — the newest constraint on the rung, and a political one.
• eia-steo-record-generation-2026-2027 — read it with the ERCOT notice: the pause, absorbed into a federal forecast.
• us-ai-dc-7gw-power-shortfall-2026 — the gap between announced and built, with its weak provenance kept in view.