GridLeads research · October 2026
ERCOT is tracking about 474 GW of large loads that want to connect, around 90% of them data centers1. The generator queue is about 463 GW1. Whoever supplies that load is mostly decided when the site is chosen.
We asked one question of every Texas large-load site in the GridLeads book: where does its power come from? 73 of 378 have generation of their own or beside them, 115 use the grid only, and 188 have no public answer.
Population: all 426 demand-side rows in the GridLeads project book, 378 in Texas and 48 in other states. The 106 supply-side rows are power plants and were left out.
Each site got one label: on-site (its own plant on the property), adjacent plant (a separate plant next door built or contracted to serve it), mixed (grid plus on-site or adjacent generation), grid only, or unknown. Battery-only sites are shown but left out of the percentages, because a battery moves power in time and does not make it. Back-up or emergency generators never count as on-site power.
A free keyword scan sorted every site into evidence groups. Research agents then read the sources for each site and had to give two word-for-word quotes per source, one naming the site and one stating the power arrangement. A separate checker re-fetched each page and ran our quote-grounding code on every answer that was not unknown: 105 were accepted, 39 downgraded (almost all to unknown) and none rejected.
One in five Texas large loads has generation, and half have no public answer.
Among the 188 sites we could settle, 39% have generation (95% interval 32 to 46). The unknowns are not random: two thirds of sites under 50 MW are unknown, against about one in four at 500 MW and over. Small plain data centers rarely say how they are fed, so the 39% is a better guide to large projects than to small ones.
Industrial plants nearly always bring their own power (26 of 29), which fits steel mills and petrochemical works that already run on-site utilities. AI and hyperscale campuses come next at 22 of 48. Crypto mines (12 of 39) and plain data centers (21 of 96) mostly buy from the grid.
The share with generation climbs from 10% under 50 MW to 48% at 1,000 MW and over, and the unknown share falls as sites get bigger. Across data centers and AI campuses in every state, sites of 200 MW and above were 2.4 times as likely to have generation as smaller ones (36% against 15%; odds ratio 3.1, Fisher exact test p = 0.014). The loads that matter most to a generator are the ones most likely to arrive with supply attached.
| Stated size | On-site | Adjacent plant | Mixed | Grid only | Unknown | With generation, of settled |
|---|---|---|---|---|---|---|
| Under 50 MW94 sites | 2 | 1 | 0 | 27 | 64 | 10% 3 of 30 |
| 50 to 199 MW58 sites | 2 | 2 | 1 | 19 | 34 | 21% 5 of 24 |
| 200 to 499 MW56 sites | 3 | 6 | 3 | 25 | 19 | 32% 12 of 37 |
| 500 to 999 MW37 sites | 4 | 3 | 1 | 21 | 8 | 28% 8 of 29 |
| 1,000 MW and over31 sites | 3 | 3 | 5 | 12 | 8 | 48% 11 of 23 |
276 Texas sites that state a size. Cells are site counts and shading is each cell's share of its row. Two battery-only sites are not shown. Stated size is a soft number.
Sites still in development (39%) look much like live ones (45%), and the difference is not significant (p = 0.52). Self-supply is not only a legacy pattern: new projects choose it at a similar rate.
Generation-paired load clusters in a few counties: Brazoria, Calhoun, Harris, Pecos, Hood and Reeves. The four Dallas-Fort Worth core counties have 89 sites and none with generation. Two thirds of those are still unknown, so the 0 of 30 settled sites there (against 73 of 158 elsewhere in Texas, p below 0.001) shows where the unknowns need work, not that metro campuses cannot self-supply.
| County | Sites | With generation | Grid only | Unknown |
|---|---|---|---|---|
| Brazoria | 7 | 5 | 2 | 0 |
| Calhoun | 6 | 5 | 0 | 1 |
| Pecos | 6 | 4 | 2 | 0 |
| Hood | 6 | 4 | 2 | 0 |
| Harris | 32 | 4 | 9 | 19 |
| Reeves | 6 | 3 | 1 | 2 |
| Bexar | 24 | 3 | 3 | 18 |
| Dallas, Collin, Tarrant and Denton | 89 | 0 | 30 | 59 |
The Texas sites fall into three groups. Large and early means 200 MW or more and announced, permitted or under construction.
Already paired
73
median stated size 400 MW, 39 of the 73 state one
These sites have a supplier, and in the on-site cases they are one. For the owner of an adjacent plant they are a relationship to protect. For others they show how much of the large-load market closes before a sales conversation begins. The 14 mixed sites also still buy grid power, so they are customers for firming, backup capacity and expansion.
Grid only, large and early
42
median stated size 500 MW
Projects of real size whose sources show a grid connection and nothing else. They are the clearest targets for a power purchase agreement, a co-location offer on an existing plant, or bridging supply. A co-location with a plant already registered as stand-alone needs PUCT approval, and the two approvals so far carry a 30-minute recall condition on the plant6, so price that condition before pitching. Sites that need power sooner than about three years out may lean on smaller machines1011.
Unknown, large and early
25
of 188 unknown sites, median stated size 400 MW
Our sources are silent on how these sites will be fed. Silence is a sales opening as much as a research gap, and a first call is also the check, because silence in public sources does not prove a site has no supplier.
Of 378 Texas data centers, crypto mines and industrial plants in the GridLeads book, 73 (19%) have generation of their own or beside them (behind-the-meter or co-located): 44 on-site, 15 with an adjacent plant and 14 mixed. 115 (30%) use the grid only, 2 have only a battery, and 188 (50%) could not be settled from public sources. Among the 188 settled sites, 39% have generation (95% interval 32 to 46%).
Industrial plants (26 of 29 settled sites), then AI and hyperscale campuses (22 of 48), crypto mines (12 of 39) and plain data centers (21 of 96). Bigger sites are more likely to: 10% of settled Texas sites under 50 MW have generation, against 48% at 1,000 MW and over.
Research agents read public sources for each site and gave two word-for-word quotes per source. A separate checker re-fetched each page and ran a quote-grounding check on every answer that was not unknown. 105 were accepted and 39 downgraded, almost all to unknown. Labels are research grade, not audited, and no site reaches our verified grade.
No. It covers the 378 Texas sites in the GridLeads book. ERCOT reported about 474 GW of large-load requests in June 2026. The percentages describe the book and should not be scaled up to that queue.
GridLeads has the details on all the large loads in this study and who to contact for each. Try the live demo at demo.gridleads.co or start at gridleads.co.
GridLeads has the details on all the large loads in this study and who to contact for each. Start with the live demo or book a call.