Texas AI Data Center Intelligence Hub

Executive summary

A Texas AI Data Center Intelligence Hub flagship report — rev. 3, a full rebuild. Derived from dataset v2.25 (as of 2026-07-25); supersedes rev. 2 (2026-07-12, dataset v2.8) in full. Every figure was re-derived from the current dataset or re-sourced from the 2026-07-25 briefs; nothing was carried over from rev. 2 unverified.

Two sourcing tiers. Dataset records are cited inline by project id (fermi-matador), event id (2026-07-minor-air-permit-lawsuit) or report id (midland-county-2026-06). External web research comes from the market-analyst refresh sweep of 2026-07-25 and the project-scout sweep of 2026-07-25, each claim carrying its brief's [disclosure] / [reported] / [estimate] label with source and date. Dataset figures are never silently overridden by external ones; where they differ, both appear with attribution.

For two years the Texas behind-the-meter (BTM) gas model rested on two legs: an ERCOT interconnection queue nobody could get through, and Permian gas so trapped it was sometimes free. The second leg broke in June 2026. Waha exited a record 89-session negative streak, and by the week to 2026-07-24 was trading around $2.00/MMBtu [estimate — dated hedging-advisor brief] — a discount to Henry Hub, not a giveaway. The first leg still holds, and it is doing all the work: ten Texas projects carrying 35,500 MW of disclosed capacity are being built around self-supplied gas, and none of the pressures that produced them — the queue, the permitting shortcut, the turbine shortage — reversed.

What did change is that the model now has to survive on speed alone, at exactly the moment its regulatory cover came under direct legal attack. On 2026-07-22 three environmental groups served a Clean Air Act notice of intent to sue over the methodology Texas BTM campuses use to get permitted — a theory that reaches roughly 38 sites, not one (2026-07-minor-air-permit-lawsuit). Three weeks earlier, on 2026-07-01, EPA had proposed to make that same shortcut easier, and it held the public hearing on that proposal the same day the notice was served (2026-07-epa-minor-air-permit-rule). And 13,000 MW — 37% of the BTM gas total — sits in projects this tracker classes as at-risk.

Key insights

1. BTM gas is 35,500 MW across ten projects — and two of the twelve off-grid records aren't gas at all. Sorting the dataset's grid.strategy field by fuel matters: the full BTM/off-grid bucket is 12 projects / 36,155 MW, but last-energy-haskell is 600 MW of nuclear microreactors and sellers-street-lufkin is 55 MW of captive biomass. Netting those 655 MW out leaves 10 projects / 35,500 MW of behind-the-meter gas — still larger than the hybrid cohort (13 projects / 16,140 MW) and larger than every grid-tied project combined (33 projects / 15,199 MW), out of 72 projects / 78,119 MW in the dataset.

2. The paid-to-take-gas era ended in June 2026, about a year earlier than we said in July. Waha traded negative for 89 consecutive sessions from 2026-02-04, bottoming at a record –$9.53/MMBtu on 2026-04-15 [reported — Hart Energy 2026-04-17; PGJ; RBN], then turned positive in mid-June — the first positive print since 2026-02-03 — as Kinder Morgan's Gulf Coast Express expansion (+570 MMcf/d) entered service and Energy Transfer's Hugh Brinson Phase I began intrastate flows 2026-06-13, ahead of schedule [reported/disclosure — BOE Report/Reuters 2026-06-16; Energy Transfer; OGJ]. Blackcomb (2.5 Bcf/d) is still not online as of 2026-07-25. Permian siting still confers a structural basis discount — August basis –$0.91/MMBtu, balance-of-summer –$1.28/MMBtu [estimate — AEGIS, as of 2026-07-24] — and supply adjacency. It no longer confers free fuel. Any BTM cost stack built on fuel ≈ $0 is now wrong.

3. Thirty-seven percent of the ledger is at-risk, and it is the two biggest names. fermi-matador (11,000 MW) and poolside-horizon (2,000 MW) together are 13,000 MW — 37% of the BTM gas total. Fermi restated its ramp on 2026-07-21 from “first 1 GW by late 2026” to “approximately 1.5 gigawatts through 2027,” expressly conditioned on binding tenant agreements it does not have (2026-07-fermi-siemens-turbines-houston), while landing 780 MW of Siemens turbines at the Port of Houston and disclosing over $1.4B invested. Poolside lost its CoreWeave anchor and ~$2B raise in April 2026 (2026-04-coreweave-poolside-exit) and its developer newsroom has published nothing since July 2025. Excluding both still leaves 22,500 MW of anchored or advancing BTM gas.

4. At Nexus Hubbard, the buildings are outrunning the power island. The record moved to under-construction this month on a TDLR filing recording a $400M, 491,378 sq ft first building with a Jan 6, 2026 start and an Oct 29, 2027 completion (2026-06-nexus-hubbard-tdlr-filing) — while the TCEQ air permit for the on-site gas plant that is supposed to power it is still under review (nexus-hubbard). Its 7,200 MW headline (Cleanview, April 2026) sits unreconciled against the 600–612 MW trade coverage consistently describes. That gap is 20% of this report's headline number.

5. The legal risk changed category: from one site to one method. The 2026-07-22 notice of intent (Environmental Integrity Project, Sierra Club, Public Citizen v. Vantage Data Centers and VoltaGrid) alleges a “piecemeal approach to Clean Air Act permitting” — permitting the generation and the data center separately so each registers as a minor source instead of aggregating into a major one requiring PSD/BACT review and public participation (2026-07-minor-air-permit-lawsuit). Because it attacks the method, it reaches every one of the ≥38 Texas data centers permitted through TCEQ's standard/minor route since 2024 [reported — Texas Tribune/Floodlight, 2026-07-09 and 2026-07-16]. By the xAI base rate — Earthjustice notice 2026-02-13 → filed complaint 2026-04-14, about two months — a filed complaint here lands around September 2026.

6. The regulator is moving the other way, in the same month. EPA proposed on 2026-07-01 to eliminate the minimum federal public-participation requirements for state minor NSR programs, handing TCEQ discretion over “whether, when, and for how long” to allow public comment; the hearing was held 2026-07-22 and comments close ~2026-08-21 [disclosure/reported — EPA release; Texas Tribune 2026-07-16] (2026-07-epa-minor-air-permit-rule). Litigants and the rule-writer are pushing on the same permitting route from opposite directions inside a single week.

7. Turbine scarcity got worse, not better — which locks the near-term builds onto engines. GE Vernova's gas equipment backlog went 100 → 116 GW in one quarter, guided to ≥125 GW at year-end, with >$5B of H1 2026 data-center orders and an output ramp of 20 GW/yr (Q3 2026) → 24 GW (2028) → 30 GW (2030) [disclosure — GEV Q2 2026 results and Form 8-K, 2026-07-22]. Siemens Energy is booked out to FY2028 on ~60 GW firm plus ~27 GW reserved, backlog €154B [disclosure/reported — S&P Global 2026-06-30, as of Q2 FY2026]. Mitsubishi still delivers no earlier than 2028 [reported]. Every 2026–2028 Texas BTM megawatt in this report therefore rides on reciprocating engines, aeroderivatives, or turbines ordered years ago.

8. There is still no defensible published $/MWh for behind-the-meter gas — and saying so is the finding. Lazard's 2026 LCOE+ figure of ~$90/MWh (up from ~$78) is grid-connected, utility-scale new-build CCGT [estimate — Lazard 2026, press-reported, not primary-verified] — a different cost object entirely. We publish the components instead, each scope-labelled, including the one new primary input: a disclosed ~6,800 Btu/kWh heat rate for the Wärtsilä 50SG [disclosure — Wärtsilä, 2026-04-23]. At $2.00/MMBtu that is ≈$13.6/MWh of fuel — our own arithmetic from two sourced inputs, not a figure taken from any source.

9. Announced megawatts remain a marketing unit. Roughly 100 GW of US data-center on-site gas is announced [estimate — BloombergNEF via Utility Dive, 2026-06-08]; roughly 2 GW operates, across four projects, rising to only ~2.8–3.2 GW by end-2026 [estimate — Cleanview census, cited by analyst name]. Sightline Climate puts grid bypass at ~3% of the 2026 delivery pipeline [estimate]. Texas is the largest single share of the paper: 32 of 74 proposed US data-center gas plants ≥100 MW, 43% of the national total (2026-07-eip-32-gas-plants; EIP, 2026-07-01) [estimate — advocacy analysis of permit filings].

1. The thesis, restated: what still drives the model, and what stopped

Rev. 2 of this report argued that BTM gas was a two-sided arbitrage: years of grid queue on one side, near-free stranded gas on the other. One side of that is now historical. The honest current thesis is narrower and, we think, more durable: BTM gas is a time-to-power trade with a fuel-basis kicker, and the kicker just shrank from “paid to take it” to “about a dollar off Henry Hub.”

1.1 The queue: unchanged, and still the whole argument

External reporting on ERCOT's June filings uses ~438 GW, ~89–90% data centers [disclosure/reported — Utility Dive, Jun 2026; Renewable Energy World, 2026-06-24]. The dataset carries both accountings in its scope note; neither is wrong, they count differently.

Against a queue the Midland County report characterises as a three-to-eight-year wait (midland-county-2026-06), the BTM cohort posts these intervals:

The grid-tied contrast from the same dataset: PowerHouse–Provident Grand Prairie's first-building energization slipped from May 2026 to Q1 2027 (powerhouse-grand-prairie); Skybox San Angelo's 1,200 MW has no construction date pending AEP/ERCOT load studies (skybox-san-angelo); Cipher's 1,000 MW Colchis direct-connect targets 2028 (cipher-colchis).

1.2 The fuel: the leg that broke

This is the single largest correction in this rebuild.

What happened, in order. Waha settled negative for 89 consecutive sessions beginning 2026-02-04 — the prior record was 27 days, set in 2024 — and set an all-time low of –$9.53/MMBtu on 2026-04-15 [reported — Hart Energy 2026-04-17; PGJ, Apr 2026; RBN Energy, Jun 2026]. Rev. 2's “record 25 consecutive negative days through 2026-03-12” and “–$3.70/MMBtu March monthly average” were accurate as of that date but described the middle of the streak, not its extent. (One aggregator reports a –$10.03 low on 2026-04-24; two better sources give –$9.53 on 2026-04-15. We print –$9.53.)

Takeaway then arrived early. Kinder Morgan's Gulf Coast Express expansion (+570 MMcf/d) entered service around 2026-05-31/06-10, and Waha printed positive in mid-June for the first time since 2026-02-03; the June cash average improved to –$0.24/MMBtu from about –$3.30/MMBtu in May [reported — BOE Report/Reuters 2026-06-16; Oil & Gas 360; RBN]. Energy Transfer's Hugh Brinson Phase I began intrastate flows on 2026-06-13 — not Q4 2026 as rev. 2 had it — moving 1.5 Bcf/d from Waha to Maypearl, with the full system scheduled complete 2027-03-01 and Phase II expanding to ~2.2 Bcf/d [disclosure/reported — Energy Transfer; Hugh Brinson Pipeline; OGJ; corroborated as “ahead of schedule” by AEGIS]. Blackcomb (2.5 Bcf/d, Waha→Agua Dulce) is still pending, tracking 2H 2026 — sources split between Q3 and Q4 [reported — OGJ].

Where the price is now. Waha spot cash averaged near $2.00/MMBtu over the week to 2026-07-24, with prompt (August) basis –$0.91/MMBtu and balance-of-summer basis ≈ –$1.28/MMBtu [estimate — AEGIS Hedging, Permian basis brief, as of 2026-07-24]. An independent print puts Waha cash at ~$1.65/MMBtu on 2026-07-20 [reported — NGI daily, via search index; the domain blocks automated fetch, and the figure is corroborated in direction by AEGIS]. AEGIS puts combined 2026 additions at ~4.5 Bcf/d (GCX +0.57, Hugh Brinson 1.5, Blackcomb 2.5) and expects the overbuild dynamic to extend into 2027 [estimate].

What this does and does not mean. It does not falsify Permian siting. A structural ~$1/MMBtu discount to Henry Hub, a dedicated 15-mile 1 Bcf/d lateral straight into the Waha Hub (gw-ranch-pecos), 5+ GW of dedicated source-priced supply inside one's own 3,200-acre tract (fo-permian-midland), a 20-year supply agreement (microsoft-reeves-pecos) and a 450,000 MMBtu/day pipeline contract (cloudburst-san-marcos) are all still real advantages, and long-tenor contracted supply is precisely what insulates a project from what just happened. What it does mean is that every merchant Permian BTM pro forma written against spot Waha in 2025 has a fuel line that is now materially wrong, and that the projects ramping into 2027 — GW Ranch, FO Permian, TCDC Phase 2 — are ramping into a normalised basis, not a subsidised one.

1.3 The rules: two hardenings in four weeks, both partly BTM-favouring

SB 6 (law 2025-06-20) gave ERCOT curtailment authority over large loads and requires on-site backup for 50% of load (2025-06-sb6-law). New grid-tied load takes firm curtailment risk; self-supplied load largely does not.

Abbott's June 10, 2026 directive made data centers bear their own interconnection and transmission-upgrade costs and recommended repealing the sales-tax exemption (2026-06-abbott-regulation).

Abbott, 2026-06-30: “we must prohibit them from building AI data centers in rural Texas neighborhoods” (2026-07-abbott-rural-prohibition) — a reversal of his “epicenter of AI development” framing [reported — Texas Tribune, 2026-06-30].

2026-07-23/24: the Governor's office converted that into a stated standard — pay for your own grid upgrades, add capacity rather than draw down headroom, run closed-loop cooling (2026-07-abbott-data-center-standards) — and released the PUCT/ERCOT response package: cost allocation to data centers, a new interconnection screening process, rules preventing large data centers from diverting existing power away from the ERCOT grid, mandatory rapid load reduction on ERCOT instruction, and a recommendation that data centers register with PUCT and ERCOT covering water and electricity operations [disclosure — Office of the Governor, 2026-07-24]. Read the anti-diversion rule precisely: it operationalises the SB6/PURA §39.169 restriction on moving existing grid-facing generation behind the meter. It does not restrict new, purpose-built on-site generation. The report does not conflate the two.

Proposed 16 TAC §25.194 would compress the large-load interconnection window from 180 days to 30 and require $50,000/MW of financial security plus site control before ERCOT studies a project (2026-06-puct-25194-window, proposed, not adopted).

Batch Zero's gate has closed and the roster is not public. The legacy Large Load Interconnection Study process ran through 2026-07-10; Batch Zero took effect 2026-07-11; the utility-side package and notarized attestation were due 2026-07-24; ERCOT issues classification notices 2026-08-07, deficiency cure 2026-08-31, binding MW allocations 2027-04-09, final transmission plan fall 2027, Batch 1 applications open summer 2027 (2026-06-batch-zero-deadlines; 2026-07-batch-zero-40pct-deferred) [disclosure — ERCOT Market Notice M-B062326-01; ERCOT trending-topic explainer, 2026-06-18].

The best available pre-notification screen is Enverus (2026-06-16), built from a 197-project database on a 12-signal readiness matrix: 55 projects / 21.7 GW positioned to qualify, 62 projects / ~37 GW screened into Batch 1+ out of roughly 59 GW screened, with ~23 projects / ~12 GW pursuing behind-the-meter generation [estimate] (2026-06-enverus-batch-zero-estimate).

A figure this report will not print, and why. A “~40% of a ~98 GW pipeline deferred” claim circulated in July and briefly entered this tracker. It could not be sourced to any analyst, and the “~98 GW” base appears to collide with ERCOT's ~98,086 MW summer 2026 peak-load scenario — a demand scenario presented to the Legislature in April 2026, not a queue and not a deferral base (2026-07-batch-zero-40pct-deferred, corrected 2026-07-25) [disclosure — ERCOT Large Load updates to Senate Business & Commerce, 2026-04-01, and House State Affairs, 2026-04-09]. Cite the base, not a percentage of an unsourced pipeline. ERCOT publishes the real roster on 2026-08-07.

2. The ledger: who is building behind-the-meter gas, at what scale

Projects whose primary power strategy is behind-the-meter or off-grid and whose on-site generation is gas-fired (grid.strategy = btm-gas or off-grid, gas powerSource):

Total: 10 projects, 35,500 MW.

Two arithmetic notes the reader is owed. The 655 MW that isn't gas: the dataset's off-grid/BTM bucket contains 12 projects and 36,155 MW. last-energy-haskell is a 600 MW nuclear PWR-20 microreactor park (still pre-application at NRC; no named data-center customer), and sellers-street-lufkin is 55 MW of captive biomass operated by Sim Agro (acquisition closing extended to Oct 31, 2026, cash portion unfinanced per SEC filings). Both are behind-the-meter. Neither is gas. Any BTM-gas total must exclude them.

The figures mix cost objects, because the records do. Microsoft Pecos's 2,000 MW is campus load; GW Ranch's 7,650 MW is permitted generation. Kilby is a third number again: Microsoft's own Jun 22 blog says ~2 GW of campus, Chevron's release says 2.67 GW of plant, and the Jun 10 Reeves County report — which predates both — carries 2,500–2,595 MW (microsoft-reeves-pecos). All three are retained with attribution; none is silently reconciled.

2.1 The at-risk block: 13,000 MW, 37% of the total

This is the number rev. 2 left as an open editorial question. It is now large enough that it belongs in the headline.

Fermi America HyperGrid — 11,000 MW. Executing on equipment while tenant-less. On 2026-07-21 three Siemens SGT6-5000F turbines (up to 260 MW each simple-cycle, 780 MW combined, 40 MW/min ramp) arrived at the Port of Houston; TSK was named for phase-two early works and Primoris for balance of plant on the six phase-one SGT-800s; the company disclosed 7,570 acres secured or under long-term lease and over $1.4B invested in site buildout to date (2026-07-fermi-siemens-turbines-houston). In the same release the ramp target became “approximately 1.5 gigawatts through 2027,” subject to entering into binding tenant agreements, with capital deployment matched to commercial progress — quietly replacing “first 1 GW by late 2026.” The release does not say the 2026 target was abandoned; we report a restatement, not an abandonment. The February construction pause (2026-02-fermi-pause) has not been described as lifted by any source — turbine delivery is not a resumption announcement. Behind it: the unnamed “First Tenant” terminated a $150M Advance in Aid of Construction Agreement on 2025-12-12, shares fell ~34% that day, and the resulting securities class action is pending in SDNY with a class period of Oct 1 – Dec 11, 2025 (2025-12-fermi-first-tenant-terminates). An EIP analysis flags the campus at an estimated ~40M tons GHG/yr (2026-07-matador-super-emitter) [estimate — advocacy]. Amarillo's executed water agreement is dated 2025-10-30 for up to 2.5M gal/day at twice the standard municipal rate, with Fermi paying all connection infrastructure (fermi-matador).

Poolside Project Horizon — 2,000 MW. CoreWeave's exit and the ~$2B raise collapse are logged at confirmed (2026-04-coreweave-poolside-exit), yet the June 10 and July 10, 2026 Pecos County reports both still present CoreWeave as the active anchor with 40,000+ GB300s and no mention of the exit — an unreconciled conflict the tracker holds open rather than resolving (poolside-horizon). Meanwhile the company keeps spending on municipal process: Ordinance No. 26-113 appeared in Fort Stockton's published 2026 ordinance directory on 2026-07-14, the day after the Jul 13 council rezoning hearing (2026-07-poolside-ordinance-26-113-published) — published, not confirmed approved; the scan has no text layer and the Jul 13 minutes are still unposted — and a second, separate Board of Adjustment variance hearing is noticed for 2026-07-27 (2026-07-poolside-variance-hearing). Both concern in-town corporate-housing parcels, not campus acreage. Pecos County's $5.5B abatement, approved 2026-01-12, predates the reported exit and stands (2026-01-poolside-abatement). The developer's own newsroom carries no 2026 items at all; its most recent post is 2025-07-28.

2.2 One project left the ledger this month

CyrusOne Freestone (760 MW) was reclassified out of behind-the-meter gas into hybrid. The tracker had counted it as BTM on a single trade-press groundbreaking report. The primary source — Constellation and CyrusOne's joint release of 2026-02-09 — states that Calpine will provide “power, grid connectivity, and site infrastructure,” and CEO Joe Dominguez sells the deal on “leveraging existing infrastructure and grid connections.” The record moved to hybrid and its 760 MW left this report's total; the release also gave us the phasing our note had carried as undisclosed (380 MW phase one plus an exclusive 380 MW phase two) and moved the record to confirmed (2026-02-cyrusone-freestone-agreement; groundbreaking at 2026-06-cyrusone-freestone-groundbreaking). Stated plainly because the direction matters: our BTM gas total went down on evidence, not up on enthusiasm.

2.3 The hybrid tier adds a second, uncounted layer of gas

13 projects / 16,140 MW carry hybrid strategy, and much of it is gas-inflected.

stargate-abilene (1,200 MW) runs an on-site gas microgrid alongside AEP grid power. The permitting detail is now on the record: 10 gas turbines and 62 backup diesel generators — roughly a 360 MW on-site plant — authorized in 2024 through TCEQ standard permits and permits-by-rule with no public-input process (permitted for >1.6M tons/yr GHG and >1,000 tons/yr combined criteria pollutants), followed a year later by a 2025 major-permit application for 41 additional turbines and 18 generators (2026-07-stargate-permit-layers) [confirmed — Texas Tribune/Floodlight, 2026-07-09]. This also resolves rev. 2's open turbine-count conflict: 10 units / ~360 MW is what is authorized; the reported 29 GE LM2500XPRESS + 12 DT35 fleet (41 units) is what is applied for.

tcdc-ector (1,000 MW) sequences 200 MW of grid Phase 1 into 450 MW of BTM gas in Phases 2–3 (Thunderhead 250 MW and EYP 400 MW power-island designs), with an Energy Dome CO2 Battery Plus MoU claiming to lift peak output up to 100% without additional gas (2026-04-tcdc-energy-dome). sailfish-comanche-circle (5,000 MW) leads with BTM gas, first 600 MW, with grid, solar and SMRs behind it.

cyrusone-bosque-whitney (400 MW) runs a Calpine on-site power program; core-scientific-pecos (1,500 MW) pairs contracted grid supply with planned BTM gas; nrg-delaware-basin (1,100 MW) is a dedicated NRG plant with ERCOT injection, contingent on FID against a long-term PPA; ampz-angelina (2,100 MW) plans 1 GW of self-generation by 2029.

meta-el-paso (1,000 MW) is its own category and worth naming as one: 366 MW of modular gas (813 units at ~450 kW each, ~$473M), built by El Paso Electric on an adjacent 31-acre site, operated by Enchanted Rock, supplying 225 MW at 95% load, with Meta funding the full cost during an initial five-year “bridge period” during which the data center stays off the grid [reported — DCD; El Paso Matters, 2026-01-18 and 2026-03-29]. Utility-owned, developer-funded, temporarily off-grid — neither pure BTM nor a grid PPA. (Note the MW conflict: 366 MW nameplate vs 225 MW contracted at 95% load. They measure different things and are not interchangeable.)

2.4 Where Texas sits nationally

EIP, 2026-07-01: 32 gas plants ≥100 MW proposed in Texas strictly to power AI data centers — 43% of the national total — projected at 287M tons GHG/yr; nationally 74 plants (71 new + 3 expansions), ~143 GW, 662M tons/yr (2026-07-eip-32-gas-plants) [estimate — advocacy analysis of permit filings; unrevised as of 2026-07-25]. This is the narrowest and best-evidenced base available because it counts permit filings rather than announcements.

Announcement-based censuses are roughly double and disagree by vintage: BloombergNEF puts US planned on-site data-center gas at ~100 GW — about 18% of the total existing capacity of all US gas power plants [estimate — via Utility Dive / Energy Innovation, 2026-06-08]; Cleanview's earlier, 2025-weighted census was ~56 GW [estimate — cited by analyst name]. These are not in conflict; they are different vintages and inclusion rules, and the report names both rather than picking one.

The operating base rate did not move: ~2 GW of BTM generation actually online across four projects as of mid-2026, projected to reach only ~2.8–3.2 GW by end-2026 — against ~100 GW announced, roughly 2–3% [estimate — Cleanview census], directionally consistent with Sightline Climate's ~3% grid-bypass share of the 2026 delivery pipeline (16 GW announced 2026 US pipeline vs ~5 GW actually under construction; 30–50% delay/cancel risk) [estimate — Sightline Climate, mid-May 2026].

The forward view: Enverus (2026-07-21) forecasts ~62 GW of new Lower-48 data-center capacity through 2030, with BTM ≈40% of installed capacity and ~1.3 Bcf/d of incremental gas demand by 2030, about half concentrated in ERCOT and PJM, and 96% of 2028 additions already under construction or in late-stage development [estimate].

3. The economics, redone — the component table

Rev. 2 published a cost stack in which fuel collapsed toward zero. That section is replaced, not patched.

No single defensible $/MWh exists for BTM gas at AI data centers in the open literature as of 2026-07-25. Rather than invent one, here are the components, each individually sourced and scope-labelled. The scope labels are load-bearing: two of these rows must never appear in the same sentence.

Correction carried from the refresh brief: Wood Mackenzie's $600/kW is by end-2027, equipment only. Secondary coverage restating it as “end-2026” is wrong, and placing it beside Bloomberg's installed-plant series without labels produces a number that means nothing.

3.2 Why the gap is a finding, not a limitation

As of 2026-07-25 there is still no defensible, independently sourced single $/MWh LCOE for behind-the-meter gas at AI data centers in the open literature.

Four reasons it stays open. (1) Scope mismatch: Lazard's ~$90/MWh is grid-connected utility-scale CCGT; BTM data-center gas is predominantly recips and aeroderivatives at higher heat rates, higher $/kW, higher capacity factors and no transmission cost. (2) The one circulating BTM-specific figure — “an industrial combustion turbine at high capacity factor carries an LCOE above $120/MWh, and LCOE analysis strongly favours front-of-the-meter grid connection over the long run” — could not be traced to an originating analysis or its assumptions [unverified — not printed as a sourced figure]. (3) SemiAnalysis's full TCO table remains paywalled. (4) No OEM, developer or hyperscaler has disclosed a BTM $/MWh.

The direction of travel is nonetheless legible: equipment prices up 195% since 2019, installed quotes up ~50% in six months, grid-side CCGT LCOE at a series high, and the fuel term no longer zero. BTM gas is not automatically the cheapest fast option in 2026; it is the fastest option whose speed is still worth paying for.

3.3 The externality argument now has a published counter-case

The strongest published rebuttal to the “off-grid protects ratepayers” line — which gw-ranch-pecos, microsoft-reeves-pecos and fo-permian-midland all lean on in their summaries — is Energy Innovation's (Jeffrey Rissman and Eric Gimon, 2026-06-08): natural gas is a traded commodity, so large new off-grid demand competes with other buyers and lifts gas prices, which lift electricity prices, because gas generators set the clearing price in most market hours. The analysis contrasts grid-connected data centers, which face regulatory oversight and cost-allocation mechanisms, with off-grid facilities that sign direct supply contracts outside utility jurisdiction — and it cites the Energy Transfer–CloudBurst agreement in Texas by name [reported — Utility Dive, 2026-06-08].

4.1 The enabling stack — turbines: scarcer than three months ago

GE Vernova, Q2 2026 (released 2026-07-22): gas power equipment backlog 116 GW, up from 100 GW at Q1 — +16 GW in one quarter — with guidance to “at least 125 GW by year-end 2026.” Q2 orders $24.2B, +88% organically; total company backlog $176B; data-center orders >$5B in H1 2026, more than double all of 2025. Production ramp: 20 GW/yr in Q3 2026, 24 GW in 2028, actions toward 30 GW in 2030 [disclosure — GEV Q2 2026 results and Form 8-K, 2026-07-22]. Rev. 2's “~100 GW backlog (Q1 2026)” is stale; the figure has moved every quarter and every mention here is date-stamped. The exit rate is rising, so the tail may compress slightly at the back end — but nothing about 2026–28 deliveries softened.

Siemens Energy: ~60 GW of firm gas turbine orders plus ~27 GW reserved, backlog €154B (record), “booked out” until fiscal 2028, >25% of the gas-services backlog now data-center driven, guiding to 90–100 GW of total commitments by end-FY2026 — as of Q2 FY2026 (quarter ended 2026-03-31) [disclosure/reported — S&P Global, 2026-06-30; Siemens Energy press]. Rev. 2 cited €146B; both are correct for their dates. Q3 FY2026 reports 2026-08-05.

Mitsubishi Power: no Q2-equivalent disclosure surfaced. Standing position unchanged — orders placed now deliver no earlier than 2028 [reported — Utility Dive; DCD]. Industry-wide OEM capacity is projected to rise from ~70 GW (2024) to >90 GW (2029) [estimate — directional].

Six Siemens SGT-800s are already delivered to Fermi and three SGT6-5000Fs landed Jul 21 (fermi-matador); GE Vernova H-class turbines supply Project Kilby, with additional capacity from Solar Turbines (Caterpillar) (microsoft-reeves-pecos); Poolside specified aero-derivative turbines with SCR (poolside-horizon); GW Ranch pairs small- and large-scale gas turbines with solar and storage (gw-ranch-pecos).

4.2 Reciprocating engines: the only 2026–28 path, and Texas is the order flow

VoltaGrid–INNIO is now 3.8 GW cumulative, not 2.3 GW. VoltaGrid ordered an additional 1.5 GW of Jenbacher engines (300 × J624/J620) in Feb 2026, delivery by 2028, on top of the record 2.3 GW order; VoltaGrid has raised $5B in project financing with Halliburton, ABB and Oracle partnerships [disclosure — INNIO and VoltaGrid releases, Feb 2026]. In the dataset, VoltaGrid's 700 MW / 210 Jenbacher engine microgrid is what takes vantage-frontier-shackelford fully off ERCOT.

Wärtsilä has passed 2.4 GW of US data-center capacity across five orders, including its first Texas order: 790 MW of off-grid primary power, 42 × Wärtsilä 50SG engines, equipment delivery 2028, fully operational late 2029 [disclosure — Wärtsilä, 2026-04-23] — updating rev. 2's “1.6 GW across four projects.” A further €292M Liberty Energy order (34SG engines, delivery 2029–2030) was announced 2026-06-29; MW not disclosed [disclosure].

Caterpillar: last hard data is Q1 2026 — record $63B backlog (+79% y/y), Power Generation revenue $2.817B (+41%), large-recip backlog up >3.5× since Jan 2024, capacity expansion raised from 2× to nearly 3× 2024 levels [disclosure/reported — Manufacturing Dive, 2026-04-30]. Rev. 2's “sold out through 2027” rests on 2025/early-2026 trade reporting and was not re-confirmed in the July window; it is dated and attributed here, not presented as a current-quarter fact. Q2 reports 2026-08-05. Cat gensets are the power island at fo-permian-midland.

Cummins–Circe Energy: ~2 GW of gas gensets for a West Texas behind-the-meter AI campus platform, deliveries 2026–2030 (2026-06-cummins-circe-2gw, reported) — now an event in the Intel feed, still not a project record.

4.3 Batteries, water, and the rest of the island

GPU training loads swing violently, and every serious BTM design pairs gas with storage: the MV BESS at crusoe-microsoft-abilene; Cat PGS1260 units “to stabilize transient AI loads” at fo-permian-midland; 1.8 GW of BESS in the GW Ranch stack; BESS in Fermi's. The most unusual entry is TCDC's Energy Dome CO2 Battery Plus MoU (2026-04-tcdc-energy-dome). On water, the pattern is closed-loop and air: Vantage's near-zero water-utilization closed-loop chilled-water system, Crusoe–Microsoft's closed-loop non-evaporative liquid cooling, Poolside's sealed loop on non-potable groundwater. Nexus is attempting something else entirely — AirJoule metal-organic-framework units producing high-purity distilled water from data-hall waste heat under an on-site Water Purchase Agreement, first systems targeted 2H 2026 (2025-12-nexus-airjoule-wpa, reported; commercial terms undisclosed).

5.1 The constraint set — permitting is the schedule, and the method is now the target

The air permit, not the interconnection, is the schedule-critical document. The dataset shows both ends of that. GW Ranch's 7.65 GW TCEQ permit — the largest power-generation air permit ever granted in the US — is the ceiling (2026-02-gw-ranch-permit), authorizing 12,000+ tons/yr of criteria pollutants and up to 33M tons/yr of GHG. Fermi stopped construction specifically to wait for its Clean Air Permit, granted 2026-02-25 (2026-02-fermi-pause, 2026-02-fermi-permit) — the floor.

A correction we made to our own record. Rev. 2 and the tracker asserted a separate “final” GW Ranch permit on 2026-06-19, clearing full-scale construction. A dedicated sweep on 2026-07-25 found no corroboration anywhere: Pacifico's own newsroom carries no June or July 2026 GW Ranch item, its most recent being the 2026-01-26 release; every primary and credible secondary account dates the authorization to late Jan / early Feb 2026; and the June item reads as a re-report. The event was retained but downgraded from reported to rumor and reframed as a sourcing artifact (2026-06-gw-ranch-final-permit). The Jun 19 claim does not appear in this report. What does not depend on it and stands: the 7.65 GW gas + 1.8 GW BESS + 750 MW solar mix, the dedicated 15-mile 1 Bcf/d Waha lateral, first 1 GW in H1 2027 and 5+ GW by 2031, and filing PSDTX1672 recorded 2025-08-04 — consistent with one permit track, not two.

The mechanism most projects actually use is the one now under attack. Since 2024, ≥38 Texas data centers have obtained on-site-power air authorizations through TCEQ's standard-permit route — historically meant for small generating units, avoiding contested-case hearings and major new-source review — alongside authorizations for 2,100+ backup diesel generators statewide [reported — Texas Tribune/Floodlight, 2026-07-09; re-confirmed verbatim 2026-07-16]. Stargate Abilene is the worked example (§2.3).

On 2026-07-22, that method was served with a Clean Air Act notice of intent to sue. The Environmental Integrity Project, the Sierra Club and Public Citizen named Vantage Data Centers and VoltaGrid LLC over two Vantage projects on the outskirts of San Antonio, alleging a “piecemeal approach to Clean Air Act permitting to fast-track these projects” — permitting the generation and the data center separately although they share a site, so each registers as a minor source rather than aggregating into a major source triggering PSD/BACT review and public participation (2026-07-minor-air-permit-lawsuit, confirmed) [reported — two independent outlets: Inside Climate News 2026-07-22; E&E News/POLITICO 2026-07-23]. EIP enforcement director Mary Greene: “Minor source permits are generally meant for small projects, like mom-and-pop dry cleaners. They are not meant for massive projects that can power an entire mid-sized city.” The on-site generation at issue is reported as two gas plants totaling ~423 MW plus 151 diesel backup generators totaling ~454 MW [reported — single-source, DCD; the notice's existence, parties, date and theory are two-source]. Former regulators quoted by Floodlight/Texas Tribune say the piecemeal approach likely violates EPA policy [estimate/reported].

Why this is the highest-consequence new fact in the report: it is a permitting-methodology challenge, not a site-specific one. It reaches every Texas campus built through the standard/minor route. The tracker has flagged the corporate exposure on vantage-frontier-shackelford precisely — the Shackelford campus is NOT named in the notice, which concerns two San Antonio sites; the risk is that the same developer and the same on-site power provider are the named recipients and the theory targets the route itself. Base rate for what comes next: at xAI's Southaven plant, the Earthjustice notice of 2026-02-13 became a filed complaint on 2026-04-14 — about two months — putting a filed complaint here around September 2026.

And the regulator is pushing the other way. EPA proposed on 2026-07-01 to eliminate the minimum federal public-participation requirements for state and local minor NSR programs, handing states — TCEQ included — discretion over “whether, when, and for how long to provide opportunities for public participation,” with EPA framing it as not altering emission standards. The public hearing was held 2026-07-22 and comments close ~2026-08-21 (2026-07-epa-minor-air-permit-rule, confirmed) [disclosure — EPA release; reported — Texas Tribune 2026-07-16; KJZZ 2026-07-16]. The Aug 21 date is press-reported; EPA's own release says 45 days after Federal Register publication without giving a date — verify against the docket. Texas-specific stakes on the record: Jennifer Hadayia (Air Alliance Houston) expects TCEQ, “already reluctant to provide rigorous oversight,” to adopt its own rules eliminating public-participation requirements; Cyrus Reed (Sierra Club): “Texas generally doesn't go above federal requirements.” TCEQ did not respond to requests for comment [reported].

Federal backdrop, carried forward unchanged: NSPS for stationary combustion turbines finalized 2026-01-15; EPA's January 2026 ruling that “mobile” turbines are stationary sources requiring permits; the proposed repeal of power-plant GHG standards; the 2026-02-12 Endangerment Finding rescission [disclosure/reported]. The NAACP / Mississippi NAACP complaint against xAI (filed 2026-04-14, N.D. Miss.) remains pending; DOJ moved to dismiss in June 2026 on national-security and citizen-suit-structure grounds, and no ruling had issued as of 2026-07-25 [disclosure/reported — CNBC 2026-06-16; The Hill; Turbomachinery].

5.2 Permit risk inside the ledger

Two of the ten projects have an unresolved permit sitting under a funded build.

nexus-hubbard. A $400M, 491,378 sq ft data hall with a state construction filing and an Oct 2027 completion date, financed by a Transition Equity Partners strategic equity stake (2025-12-nexus-transition-equity) and two nine-figure Eagle Point senior secured facilities, the second closed 2026-03-12 (2026-03-nexus-eagle-point-financing) — sitting on a power island whose TCEQ air permit has not issued. Permit identifiers referenced in coverage as 183046 / PSDTX1700 / GHGPSDTX269 for applicant NEXUS HUBBARD POWER, L.L.C. are stated as referenced, not asserted: the indexed TCEQ document is a scanned PDF that could not be machine-read, and the coverage carrying the numbers is on a site that does not resolve. The buildings are outrunning the power.

fermi-matador and gw-ranch-pecos are the counter-case: permits in hand, tenants absent.

Neither sailfish-comanche-circle (5,000 MW, hybrid, BTM-gas-first) nor alpha-digital-reeves (2,000 MW, of which 58.3 MW is filed with ERCOT) has an issued permit event for full-scale generation in the log.

5.3 Water and local politics: the constraint that actually kills projects

Texas's first clean cancellation of the new era was water and siting, not gas. Diode Ventures withdrew a proposed data center near Cedar Creek Reservoir, Henderson County, on 2026-07-23, citing community opposition and Abbott's rural-ban push (2026-07-diode-cedar-creek-cancelled, confirmed). Reported site profile: a ~12-year buildout drawing 5 million gallons/day of lake water by year five; the community group Save Cedar Creek Lake gathered thousands of petition signatures [disclosure/reported — Office of the Governor 2026-07-23; Texas Tribune; KERA; Tyler Morning Telegraph]. Two caveats this report keeps: it is not a BTM-gas cancellation — no on-site-generation dispute is reported — and the withdrawal was announced by the Governor's office as a policy success, which is a framing incentive both parties benefit from. One event is not a Texas pattern. The Governor then converted it into a template (2026-07-abbott-data-center-standards).

Nationally the pattern is real and large. Data Center Watch counts at least 75 projects worth ~$130B blocked or delayed in Q1 2026 alone — described as the largest single-quarter concentration on record, roughly matching all of 2025 in three months — with active opposition groups rising from 396 (end-2025) to 833 (March 2026) across 49 states, >300 state data-center bills filed in the first six weeks of 2026, statewide moratorium proposals in 14 states, and Maine within one House vote of the first statewide ban [estimate — advocacy/industry tracker; corroborated by NBC News, Jun 2026, and Newsweek]. No Texas-specific breakdown is published; none is inferred here. This retires rev. 2's unverified “~25 cancelled” figure.

In-state: San Marcos passed an outright municipal data-center ban 4–3 on 2026-06-16 (2026-06-san-marcos-data-center-ban) — and reporting states explicitly that it does not reach the CloudBurst campus, already under construction on the city's border, with no extraterritorial provision described (2026-06-san-marcos-ban-cloudburst-unaffected, reported). Hill County passed the state's first county moratorium, was sued for $100M, and rescinded it 24 days later in favour of a developer checklist (2026-06-hill-county-moratorium-rescinded) — the checklist now applies to Nexus. Hood County approved Comanche Circle without conditions after a lawsuit threat (2026-06-comanche-circle-forced-approval). Tom Green County suspended new industrial water-tapping permits for computing facilities (2026-07-tom-green-water-permit-suspension); San Angelo mandated closed-loop cooling with hard caps on city water draw (2026-06-san-angelo-water-ordinance). A UT/Texas Politics Project poll puts opposition at 56% of Texans and 62% of rural voters (2026-06-ut-poll-opposition). A Tribune census counts 335 existing and 248 planned Texas data centers, with half the planned ones sited in unincorporated county land (2026-06-texas-dc-census) — which is where nine of this report's ten BTM projects sit, and where legal resistance is weakest.

Midland codified BTM gas as the price of entry: no municipal water for cooling, sites outside residential limits, self-generated natural-gas power (2026-06-midland-dc-conditions; midland-county-2026-06). fo-permian-midland exists in that shape because of it.

Panhandle groundwater is tightening in the same quarter Fermi is buying Amarillo municipal water. On 2026-07-09 the Panhandle Groundwater Conservation District converted transport registrations into transport permits, with stricter notice and hearing requirements, penalties for unpermitted transport, and board authority to require hydrology studies [confirmed — Amarillo Tribune, 2026-07-15]. The article does not mention data centers or Fermi, and no causal link is asserted; the adjacency is the analysis, not the source.

5.4 Commercial structure remains the dominant failure mode

The market has effectively split BTM gas into anchored and merchant, and the split maps almost perfectly onto execution.

Every project that is actually building has a counterparty before steel: a hyperscaler owner (microsoft-reeves-pecos, crusoe-microsoft-abilene on a 900 MW Microsoft lease), a Stargate lease (vantage-frontier-shackelford, Oracle → OpenAI), or a contracted power/gas counterparty (cloudburst-san-marcos's Energy Transfer agreement — though note that agreement was contingent on CloudBurst reaching FID with its customer, and no FID or tenant has been announced since Feb 2025).

Every project that is stalled lacks one. Fermi has none and says so in its own guidance. Poolside's is disputed. alpha-digital-reeves, gw-ranch-pecos, nexus-hubbard and fo-permian-midland all carry “TBD” in the tenant field.

Developer silence is itself a signal, and the July sweep quantified it: poolside.ai/newsroom has zero 2026 items (newest 2025-07-28); cloudburstdc.com/press-releases carries exactly one release, Feb 2025, for a 1.2 GW campus targeting Q4 2026 first power; fopermian.com/news carries one item, Sep 2025, against a late-2026 Phase 1 target; Pacifico's newsroom has no GW Ranch item since Jan 26, 2026. Against that, vantage-frontier-shackelford is the outlier in the right direction — Vantage's own Shackelford County page still states, as of this sweep, that “the first data center is scheduled for delivery in the second half of 2026.” It is the only large off-grid gas campus in the tracker with a 2026 delivery still affirmed by the developer.

No confirmed case was found of a Texas BTM gas project reverting to grid or switching to nuclear/solar+BESS. Re-checked 2026-07-25; the observable Texas pattern is hybridization rather than reversal — GW Ranch folding solar and storage alongside 7.65 GW of gas, Fermi combining gas, nuclear, wind and solar. The closest non-Texas contrast remains Stargate Michigan's 1.4 GW DTE grid PPA. [flagged as not found, not disproven] Likewise, no AWS behind-the-meter gas announcement was located as of 2026-07-25 — which is not the same as saying AWS is absent.

5.5 The fiscal question arrives in two days

The state's subsidy for all of this is now under formal fiscal review. Senate Finance (Chair Joan Huffman) holds a public hearing on 2026-07-27 at 10:30 a.m. in E1.036 on interim charge “Data Center Investment and State Fiscal Effects,” studying Tax Code §§151.359 and 151.3595. The notice's own figure: state cost has grown from an estimated $14.6M for the 2014–15 biennium to a projected $3.3B for the 2028–29 biennium (2026-07-senate-finance-hearing-notice, confirmed). That is a different horizon from the Comptroller's $3.1–3.2B two-year estimate (2026-04-comptroller-exemption-cost) and the two are not interchangeable; Texas Tribune independently pegs the current cost near $1B/year [reported — 2026-04-08], and Good Jobs First puts local abatement losses at $1.3B in FY26 rising to $1.6B in FY27 (2026-06-goodjobsfirst-abatements) [estimate — advocacy]. A leaked Senate Finance draft proposes full repeal with no grandfather clause (2026-07-tax-repeal-draft-bill, reported).

The hearing has not occurred as of this report. Its outcome is not characterised here.

6. Watchlist: dates that will date this report

Methodology — dataset tier

Scope. All dataset figures from Texas AI Data Center Intelligence Hub dataset v2.25 (as-of 2026-07-25): 72 projects, 222 events, plus the report and news archives. Confidence labels (confirmed / reported / rumor) and disclosed-vs-estimated attributions are reproduced from the records. Where a figure is a county-report inference rather than a company disclosure — FO Permian's $1.05–1.5B capex and 300–500k sq ft; the McKinzie 20–40 MW estimates — it is labelled as such per the source record.

Aggregation method. Strategy cohorts follow each record's grid.strategy; capacity sums use each record's headline powerMW (disclosed figures only; nulls excluded, so google-armstrong, google-ranger-haskell, tpl-bolt-permian, aws-round-rock and others carry no MW and are absent from all totals).

BTM/off-grid, all fuels: 12 projects / 36,155 MW. Non-gas exclusions: 655 MW — last-energy-haskell (600 MW nuclear), sellers-street-lufkin (55 MW biomass). Behind-the-meter GAS: 10 projects / 35,500 MW — fermi-matador 11,000, gw-ranch-pecos 7,650, nexus-hubbard 7,200, poolside-horizon 2,000, microsoft-reeves-pecos 2,000, alpha-digital-reeves 2,000, vantage-frontier-shackelford 1,400, cloudburst-san-marcos 1,200, crusoe-microsoft-abilene 900, fo-permian-midland 150. By status inside that: under-construction 10,700 MW (4 projects); at-risk 13,000 MW (2); early-works 7,800 MW (2); pre-construction 4,000 MW (2). Hybrid: 13 projects / 16,140 MW. ERCOT-grid: 33 projects / 15,199 MW. Dataset total: 72 projects / 78,119 MW. Reclassified out this cycle: cyrusone-freestone (760 MW), btm-gas → hybrid, on the 2026-02-09 Constellation/CyrusOne primary release (2026-02-cyrusone-freestone-agreement).

Totals mix operational, under-construction and planned capacity, and mix campus load with permitted generation where that is what the record discloses. Both caveats are stated in the body.

Undisclosed means undisclosed. GW Ranch, CloudBurst and Alpha Digital capex; Nexus's campus total (the $400M is one building); Microsoft Pecos investment (industry-estimated ~$7B for the initial phase, not a company disclosure); Crusoe–Microsoft cost; Fermi total financing; Wärtsilä's Liberty Energy order MW — all carried as undisclosed because the records carry them that way.

Report records relied on: midland-county-2026-06 (queue trajectory, three-to-eight-year characterization, Midland conditions framework, re-skilling thesis), ector-county-2026-06, pecos-county-2026-06. The principal event records are listed on this page under Extracted events.

Methodology — external tier

Sources. Market-analyst web-research refresh brief, sweep of 2026-07-25, which supersedes and extends the 2026-07-12 sweep; and the project-scout BTM-gas/off-grid project sweep of 2026-07-25. Every external claim carries the briefs' [disclosure] / [reported] / [estimate] label with source and date, reproduced unchanged.

Principal external sources: GE Vernova Q2 2026 results and Form 8-K (2026-07-22); S&P Global on Siemens Energy (2026-06-30); Manufacturing Dive on Caterpillar Q1 2026 (2026-04-30); INNIO and VoltaGrid (Feb 2026); Wärtsilä (2026-04-23, 2026-06-29); Wood Mackenzie (2026-04-01); Bloomberg (2026-04-01); SemiAnalysis (2025-12-30) and Grid Capacity Intelligence (2026-03-05); Lazard LCOE+ 2026 (press-reported); Hart Energy (2026-04-17), PGJ, RBN Energy, BOE Report/Reuters (2026-06-16), Oil & Gas 360, AEGIS Hedging (as of 2026-07-24); Energy Transfer / Hugh Brinson Pipeline / OGJ; ERCOT market notices and legislative-hearing decks; Enverus (2026-06-16, 2026-07-21); Environmental Integrity Project (2026-07-01); Inside Climate News (2026-07-22); E&E News/POLITICO (2026-07-23); Texas Tribune / Floodlight (2026-07-09, 2026-07-16, 2026-06-30, 2026-07-23, 2026-04-08); KJZZ (2026-07-16); EPA news release and NSR regulatory actions (2026-07-01); Office of the Texas Governor (2026-07-23, 2026-07-24); Data Center Watch Q1 2026 (with NBC News and Newsweek corroboration); Sightline Climate (mid-May 2026); BloombergNEF via Utility Dive / Energy Innovation (2026-06-08); RMI; Texas Senate Finance hearing notice; CNBC (2026-06-16); DCD and El Paso Matters on Meta El Paso; Amarillo Tribune (2026-07-15).

Cleanview is cited by analyst name only; no Cleanview links appear anywhere in this report.

Known conflicts, handled rather than resolved

Turbine cost series scope: WoodMac's ~$600/kW is equipment only, by end-2027; Bloomberg's $2,000→$3,000/kW is installed plant; the EIA-derived +66% is installed CCGT. Presented separately; never combined in a sentence.

GE Vernova backlog: 80 / 83 / 100 / 116 GW reflect different as-of dates and slot-reservation treatment. Every mention is date-stamped; 116 GW as of Q2 2026 is current.

Waha record low: –$9.53/MMBtu on 2026-04-15 (two sources) vs –$10.03 on 2026-04-24 (one aggregator). We print –$9.53.

Nexus Hubbard capacity: 7,200 MW (Cleanview, Apr 2026, the basis of the second-largest-in-Texas ranking) vs 600–612 MW in trade coverage. The only reconciliation found — a claim that the TCEQ application covers ~7,230 MW of continuous generation plus ~19.5 MW of emergency standby — was on a non-resolving site and was not ingested. The gap is carried, not closed.

Project Kilby capacity: ~2 GW campus (Microsoft blog), 2.67 GW plant (Chevron/CNBC), 2,500–2,595 MW (Jun 2026 Reeves County report). All three printed with attribution.

Poolside anchor: Apr 2026 CoreWeave-exit reporting vs Jun/Jul 2026 county reports still naming CoreWeave. Both carried; the at-risk posture holds until directly resolved.

Meta El Paso: 366 MW nameplate vs 225 MW contracted at 95% load. Different measurements; not interchangeable.

Blackcomb in-service: “Q3 2026” vs “Q4 2026” vs “2H 2026.” Not in service as of 2026-07-25; we use “2H 2026, not yet online.”

Vantage/VoltaGrid MW split (423 MW gas + 454 MW diesel across 151 gensets): single-source to DCD; the notice's existence, parties, date and theory are two-source.

Caterpillar “sold out through 2027”: 2025/early-2026 trade reporting, not re-confirmed in the July window. Dated and attributed, not asserted. Rechecks 2026-08-05.

EPA comment deadline 2026-08-21: press-reported; EPA's own release says “45 days after FR publication” without a date. Verify against the docket.

Abilene turbine inventory — RESOLVED this cycle. Rev. 2 printed neither of the conflicting counts. 2026-07-stargate-permit-layers reconciles them: 10 units / ~360 MW authorized; 41 units (29 GE LM2500XPRESS + 12 DT35) applied for.

Deliberately not printed

Any “~40% of a ~98 GW pipeline deferred” figure — unverifiable as stated; likely a collision with ERCOT's ~98,086 MW summer peak-load scenario.

Any single-figure BTM-gas LCOE — none exists in open literature; components published instead. The uncorroborated “>$120/MWh industrial CT” claim. The ~101 GW / 57 GW ordered / 7 GW under construction / Texas 38 GW cluster — single-origin, originating analyst page not retrievable; BNEF's ~100 GW is used instead, with attribution. The “–$10.03/MMBtu” Waha low.

Any assertion that AWS is absent from BTM gas — none was found as of 2026-07-25, which is not the same thing. The trade-blog claim that Google backs Nexus Hubbard and it is “set to serve Anthropic” — no corroboration in any release, filing or wire story; recorded as seen and rejected. The “final GW Ranch permit, Jun 19, 2026” — withdrawn by our own correction.

Dataset follow-ups this rebuild surfaces

PowerPlay AI, greater Abilene — 400 MW behind-the-meter gas, announced 2026-07-20, JV with an unnamed Nasdaq-listed neocloud, IPP selection via RFP in progress, initial 400 MW targeted 2028. Not yet a tracker record.

Cummins–Circe ~2 GW West Texas portfolio is now an event (2026-06-cummins-circe-2gw) but still not a project record.

Energy Innovation's ratepayer analysis (2026-06-08) and the Panhandle Groundwater rule change (2026-07-09) are scout-brief proposals not yet in the Intel feed; both are cited here from the external tier only.

Nexus Hubbard TCEQ application (183046 / PSDTX1700 / GHGPSDTX269) needs a direct read — it is the key to the 612 MW vs 7.2 GW question and to this report's largest single uncertainty.

Editorial decisions of record

1. The headline capacity keeps the at-risk block in. 35,500 MW stands, with the 13,000 MW / 37% at-risk share stated in the same breath (§2.1, insight #3). It keeps the figure comparable with every other capacity number this tracker publishes, and discloses the caveat rather than netting it away.

2. Nexus Hubbard's 7,200 MW stays in the total, with the unreconciled 7,200-vs-612 MW gap and the pending TCEQ air permit stated in the ledger and §5.2. We publish the tracked figure and show the gap; we do not substitute our own estimate for it.

3. Retained as drafted: the three-way “98 GW” collision sidebar; the ~$13.6/MWh fuel figure, printed only inside the component table and labelled as our own arithmetic; Lazard's ~$90/MWh with its wrong-cost-object hedge; and the Vantage/VoltaGrid framing, which states twice that the Shackelford campus is not named in the notice.

Known revision points — this report is dated by design. The Senate Finance hearing on the sales-tax exemption falls on 2026-07-27; ERCOT publishes the actual Batch Zero roster on 2026-08-07; Siemens Energy and Caterpillar report on 2026-08-05; EPA comments on the minor-permit rule close around 2026-08-21 (press-reported — verify against the docket). Three of the figures here will move within a fortnight.

Not yet tracked, routed for a follow-up pass: PowerPlay AI (400 MW, greater Abilene), a Cummins–Circe campus record pending a disclosed county, and the Energy Innovation ratepayer analysis and Panhandle Groundwater Conservation District rule change as Intel-feed events. Each is cited here from the external tier only and labelled as such.