The constraint on AI expansion in 2026 is not chips and increasingly not even generation. It is the queue to connect to the grid. This page tracks how long that queue is and what it means for the buildout.
Queue Size
| Metric | Value |
|---|---|
| Total US interconnection queue backlog | ~2,600 GW |
| Projects actively seeking interconnection (end 2025) | ~8,200 |
| Generation capacity in queue | ~1,312 GW |
| Storage capacity in queue | ~749 GW |
| ERCOT large-load queue (April 2026) | ~410 GW |
| Data center share of ERCOT large-load queue | ~73% |
| New large load applying in ERCOT, Q1 2026 alone | 198 GW |
The ERCOT figures are the most striking. A single quarter brought 198 GW of large-load applications, which exceeds total global data center power demand for 2026. Queue applications are not commitments, and a large fraction never get built, but the ratio of applications to deliverable capacity tells you how far demand has outrun supply.
Wait Times by Region
| Region | Typical wait to commercial operation |
|---|---|
| ERCOT (large load above 75 MW) | 3-4 years |
| MISO | 4-5 years |
| SPP | 4-5 years |
| CAISO (California) | 5-6 years |
| PJM (largest AI campus filings) | Approaching 7 years |
| US median across projects | Approaching 5 years |
| Worst-case data center scenarios | Up to 12 years |
ERCOT is fastest because it runs a separate large-load process outside the generation queue, which is the main structural reason Texas has attracted disproportionate AI campus development.
Why This Is Structural
Interconnection queues were designed for generators joining a grid, processed serially with study after study. They were not designed for gigawatt-scale loads arriving faster than transmission can be planned. Reform is underway across most operators, but transmission construction runs on decade timescales and queue reform does not add wires.
The practical consequence is that announced capacity and deliverable capacity have decoupled. A 2026 announcement with a 2028 target date is, in PJM, more likely to be a 2031 facility.
Brand Visibility Implications
Interconnection delay is the strongest argument that inference capacity will stay tight and inference pricing will not fall as fast as the last three years suggested. Tight capacity favours efficient models, sparse architectures, and cheap open weights, all of which sit outside the monitoring surface most brand teams currently use. See the inference and training split.
Methodology
Figures compiled from Gartner, IDC, LBNL, grid-operator filings, utility rate cases, and press reporting through July 2026. Forecasts are cited to the forecaster because independent projections in this area diverge widely, and several of the underlying quantities are estimates rather than measurements. Where sources disagree, ranges are given rather than a single number. Updated quarterly.
How Presenc AI Helps
Presenc AI measures brand representation across cheap and premium model tiers, so shifts caused by inference economics are visible rather than inferred.