Research note (plain English) · 2026-04-24
11 GW of ERCOT's incoming data-center queue is piling into the Texas Panhandle — and the bottleneck there is local power lines, not statewide generation
ERCOT is the grid operator for most of Texas. A widely-cited Duke University paper (Tyler Norris, February 2025) calculated that ERCOT could absorb roughly 10 GW of new round-the-clock load — picture ten large data centers running 24/7 — if that load agreed to be turned off about 0.5% of hours per year. Cliff's public dataset of announced ERCOT data centers shows a different problem: out of a 17.6 GW pipeline of new sites under construction or planned, 11 GW is bunched in two sites in the geographically isolated Panhandle. Whether those projects can actually plug in depends on building new transmission lines into that corner of the state, not on whether ERCOT has enough generation overall.
- Cliff dataset compiled
- 2026-04-24
- ERCOT sites analysed
- 30
- Total announced MW
- 22.6 GW
- Incoming (under construction + planned)
- 18.1 GW
Glossary
The terms used on this page
Quick reference for the acronyms and grid-jargon that appear in the table, the chart, and the methodology footnotes below.
- ERCOT
- Electric Reliability Council of Texas. The grid operator for most of Texas; runs largely independent of the rest of the U.S. grid.
- MW / GW
- Megawatts and gigawatts of electrical capacity. 1 GW = 1,000 MW, roughly the output of one large power plant or the demand of a city of about a million people.
- Load (vs. generation)
- "Load" is electricity demand (data centers, factories, homes). "Generation" is supply (power plants). The grid has to balance them every second.
- Curtailment
- Turning a customer or generator down or off when the grid is constrained. A data center that accepts X% annual curtailment is agreeing to be powered off for that share of hours per year.
- Headroom
- How much new load the grid could absorb without breaking. The Duke paper's contribution is that headroom rises a lot if new load accepts even small amounts of curtailment.
- Load pocket / load zone
- A geographic region of the grid where transmission limits trap load and supply together. ERCOT's administrative zones are coarse; the Panhandle is a textbook load pocket because transmission into it is constrained.
- Interconnection queue
- The waiting list of projects (loads or generators) that have asked to plug into the grid. Years long; a project's position in the queue is itself a tradable asset.
- Firm interconnection
- A grid contract where the customer has the right to full power 24/7 with no curtailment. The default for most data-center deals today.
- Controllable Load Resource (CLR) / PGRR145 / NPRR1325
- A new ERCOT designation, recently codified in protocol revisions PGRR145 and NPRR1325, that lets a large load be dispatched up and down by ERCOT like a power plant. Required if a load wants to accept curtailment in exchange for faster interconnection.
- SPP seam
- The boundary between ERCOT and the neighboring Southwest Power Pool grid. The Panhandle sits near it; cross-seam transmission is one of the main ways more power can be moved into Panhandle data-center loads.
The headline finding
Most of ERCOT's incoming data-center load is geographically clustered
The 11 GW in the Panhandle across just two announced sites is the most extreme geographic concentration in Cliff's public ERCOT dataset. That kind of clustering is invisible to the Duke paper, which models ERCOT as one big bucket of supply and demand. To group sites into regions we used a simple nearest-city rule (Houston, Dallas–Fort Worth, San Antonio, West Texas / Permian Basin, and Panhandle), which tracks where the transmission pinch points actually live more faithfully than ERCOT's formal four administrative zones. The table below is the resulting tally of announced data-center capacity per region, broken down by project stage. "MW" means megawatts of nameplate electrical capacity; 1,000 MW is one gigawatt (GW), roughly the output of a single large power plant.
| Region | Already operating (MW) | Under construction (MW) | Planned (MW) | At risk (MW) | Total (MW) | Sites |
|---|---|---|---|---|---|---|
| Panhandle | 200 | 0 | 11,000 | 0 | 11,200 | 2 |
| North (DFW) | 1,980 | 2,230 | 1,500 | 0 | 5,710 | 16 |
| West | 300 | 2,200 | 500 | 0 | 3,000 | 4 |
| Houston | 1,352 | 0 | 200 | 0 | 1,552 | 4 |
| South | 700 | 485 | 0 | 0 | 1,185 | 4 |
Putting 11 GW of new load into one geographically isolated corner of the state isn't the same as spreading it evenly across ERCOT. It's a classic transmission bottleneck: the projects can only connect to the grid as fast as new transmission lines into the Panhandle (and across the seam with the neighboring Southwest Power Pool grid) can be built. The Duke paper's "ERCOT can absorb 10 GW" statement is correct as a system-wide average; it just doesn't tell you that the headroom isn't in the place where the load is going.
Statewide upper bound
Even if every project agreed to be turned off sometimes, the queue still overshoots
The Duke paper's key idea: if a new data center will accept being curtailed (i.e. powered down) for some small percentage of hours each year, the grid can fit much more of it without building new generation. The chart below asks the most-favorable version of the question: assume every single MW of Cliff's 17.6 GW pipeline agrees to that deal. In reality, almost none of these projects are signed up to be curtailable today — they are on firm interconnection contracts — so this is a generous upper bound. Even so, 7.6 GW doesn't fit under the 0.5%-curtailment ceiling, which is more evidence that the binding problem is geographic concentration, not statewide capacity. Headroom numbers come from Figure 8 of Rethinking Load Growth, Duke Nicholas Institute, February 2025.
Curtailment-enabled headroom
Norris (Duke, Feb 2025) curve · Cliff queue overlay
At 0.25% curtailment
6.5 GW headroom
6.5 GW of Cliff's queue fits · 11.6 GW overshoots
At 0.5% curtailment
10.0 GW headroom
10.0 GW of Cliff's queue fits · 8.1 GW overshoots
At 1.0% curtailment
14.7 GW headroom
14.7 GW of Cliff's queue fits · 3.4 GW overshoots
If load accepts being curtailed 0.25% of hours
6.5 GW fits · 11.6 GW overshoots
Duke estimate: 6.5 GW absorbable · share that fits 36%
If load accepts being curtailed 0.5% of hours
10.0 GW fits · 8.1 GW overshoots
Duke estimate: 10 GW absorbable · share that fits 55%
If load accepts being curtailed 1.0% of hours
14.7 GW fits · 3.4 GW overshoots
Duke estimate: 14.7 GW absorbable · share that fits 81%
At a 0.5%-of-hours curtailment ceiling — roughly the budget today's utility "interruptible" programs are designed around — only 10 GW of the 17.6 GW pipeline fits, and that's with the generous assumption that every project is willing to be interrupted. The remaining 7.6 GW would need to either (a) accept being switched off more often, (b) wait years for new transmission, or (c) sign up under ERCOT's newer rule changes (PGRR145 / NPRR1325) that formally register a data center as a "controllable load resource" (CLR) — meaning ERCOT can dispatch it down like a power plant in reverse. Cliff's dataset only captures sites that have been publicly disclosed, so ERCOT's full confidential interconnection queue is even larger; the comparison here is conservative.
What this means for site developers
How much downtime your data center is willing to tolerate is now a site-selection variable
If you can tolerate 1.0% downtime
83% of the announced ERCOT data-center pipeline fits inside the grid's headroom. Roughly 2.9 GW still has to wait for new transmission or move to a less-constrained region.
If you can tolerate 0.5% downtime
57% fits. This is the budget most existing utility "interruptible" tariffs are designed around. Above this threshold, you have to formally register the site as a controllable load under ERCOT's new PGRR145 process, which is still a young regulatory program.
If you can tolerate only 0.25% downtime
37% fits. In practice, this is roughly "only the projects already running plus the very earliest under-construction sites clear." Most planned MW would have to redesign their load shape, move to a different region, or accept more downtime.
Methodology
How the numbers were built
Duke side (the paper)
Norris and co-authors used 9 years of hourly load data published by the federal Energy Information Administration (the EIA-930 dataset, 2016 through 2024) and ran a numerical solver to find the largest round-the-clock new load that could be added on top of the existing ERCOT load while staying inside each curtailment ceiling. They deliberately treat ERCOT as one big bucket: no transmission constraints between regions, no plant-by-plant constraints over time. That's what makes their answer a system-wide ceiling rather than a site-by-site forecast.
Cliff side (the dataset)
Curated from official press releases, ERCOT meeting materials, SEC filings, and major-trade-press coverage. Lat/lon are city/county centroids unless a specific facility coordinate was disclosed; precision is sufficient for state-wide visualization, not for parcel work. MW values are public nameplate or disclosed targets, not metered demand. Compiled 2026-04-24. This is a leadgen visualization, not an authoritative datacenter registry. Expect entries to be ±15% on MW, missing recent disclosures, and stale on stage transitions. Refresh quarterly.
Where this analysis is honest about its limits
- Cliff's dataset only includes data-center sites that have been publicly disclosed (press release, county filing, news report). ERCOT's confidential Large Load Interconnection queue is larger, so the "overshoot" numbers here are a floor, not a ceiling.
- The Duke paper's headroom is computed for ERCOT as a whole. We argue that local transmission bottlenecks bind before the system-wide capacity ceiling does — which lines up with what is publicly known about Panhandle transmission — but we have not built a formal regional model in this note.
- The MW numbers in Cliff's dataset are "nameplate" (maximum theoretical capacity), not the coincident peak that actually shows up at the same hour as the rest of the grid's peak. Realistic utilization factors would let more projects fit than these bars suggest.
- The Duke model assumes new load runs flat 24/7. Real AI training workloads are bursty — they don't run at full power every hour — so a more realistic load shape would create more effective headroom than the chart shows. We plan to extend the note with that sensitivity.
We'd welcome a methodology read from the Duke authors before publishing more broadly
Every quantitative claim on this page extends the Duke paper's methodology using Cliff's public dataset of announced ERCOT data-center sites. Before pushing this in front of a wider audience, we'd value a sanity check from the paper's authors — in particular: does the system-wide headroom framing still hold once you zoom into individual regions, and does the Panhandle concentration deserve its own separate treatment?
Suggested citation: Cliff Research, “ERCOT's local transmission bottleneck: 11 GW concentrated in the Panhandle,” 2026-04-24, cliffcenter.com/research/ercot-curtailment-headroom. cliffcenter.com/research/ercot-curtailment-headroom.
Built on Tyler H. Norris et al., Rethinking Load Growth, Duke Nicholas Institute, February 2025. https://nicholasinstitute.duke.edu/publications/rethinking-load-growth.