Methodology & data sources

The single biggest error in “AI power gap” charts is treating interconnection-queue requests as if they were load on the meter. They are not.

• Requested load = gross interconnection / large-load queue inflow. It is enormous and heavily padded: the same project appears in multiple queues, speculative requests pile up, and historically only ~21% of queued projects ever reach commercial operation (LBNL). It is a leading indicator of intent.
• Energized load = capacity actually metered and online. This is physical reality, and it is far smaller. We ground it in utility/ISO metered filings — never the EIA/S&P “statewide forecast” headline numbers, which are themselves padded constructs.

Example: ERCOT’s large-load queue is 200+ GW, of which ~13.7 GW is approved-to-energize but only ~5.0 GW shows up as “observed energized.” Dominion shows ~70 GW requested vs ~4 GW energized. We display both, and never present the request figure as if it were a power deficit.

• Gen delivered / quarter — new generation reaching commercial operation (trailing-4Q avg). The deliverable supply, and the headline KPI.
• DC requested / quarter — data-center load entering the queue (padded; intent).
• DC energized / quarter — data-center load actually coming online (reality).
• Realization rate = energized ÷ requested. The honesty metric: how little of the pipeline is real. It runs ~0% (Louisiana, all-future) to ~60% (Memphis, self-supplied).
• Gen delivery rate = delivered ÷ planned generation (queue throughput health).

The per-region “static queue drain” is the time to build out the standing load queue at the current generation-delivery pace, assuming no new requests arrived. It is a deliberately static reference, not a forecast. In reality requests keep flooding in and most queued load withdraws — so the queue neither clears on schedule nor reflects real demand. The realization rate is the corrective: a 12-year “drain” against a queue with a 4% realization rate is not a 12-year shortage.

“Gen delivered” is total regional generation reaching operation — it serves all load growth, retirements and electrification, not only data centers, and much of it is intermittent (solar/wind nameplate) while data centers need firm 24/7 power. So a region adding more generation MW than it energizes in data-center MW is not necessarily “covered”: the binding constraints are queue-study throughput, firm capacity, and local transmission. That is precisely why the requested pipeline is worth tracking as a leading signal, with the realization rate keeping it honest.

The four ★ regions — Northern Virginia (PJM/Dominion), Texas (ERCOT), Arizona, and xAI Memphis (TVA) — are anchored directly on the AI-datacenter research vault, which flags power as “the fastest-growing constraint in the entire stack.” Four additional top-tier US hubs (Central Ohio, Louisiana, Georgia, Iowa) are included for national completeness and to span the full realization spectrum.

• LBNL “Queued Up” 2025 ↗ — aggregates >50 grid operators (~98% of US capacity); the ~21% historical realization rate comes from this work.
• Metered/energized anchors — ERCOT TAC report ↗, Dominion SCC filing ↗, AEP Ohio / PUCO 24-508 ↗.
• Per-ISO queue files — PJM ↗, ERCOT ↗, MISO ↗ — and the EIA Wholesale Market Portal ↗.

Queue stocks and the energized anchors above are directly sourced (each region page lists its sources). The per-quartersplits and the six non-anchor energized series are modeled to fit those anchors and the national ~+11 GW/yr energized total (S&P). Treat the quarterly flow figures as grounded estimates, not metered series. Refresh the public source files with npm run fetch.