Texas Just Discovered It Has a Veto Over the AI Build
The Probability Desk — Monday, 17 August 2026
The most permissive large-load jurisdiction in the United States froze a 200-gigawatt data-centre queue on a governor's letter. The Desk puts 55% on a narrower gate reopening between November and April, and 35% on the gate staying shut past 30 April 2027 — past the outer bound ERCOT set for itself. The megawatts are not the story. The precedent is.
Desk view, in one paragraph
On 3 August the Governor of Texas wrote two letters and stopped the largest interconnection queue in the Western Hemisphere. No statute was passed, no rule was amended, no hearing was held. ERCOT suspended a scheduled classification step four days before it was due, and eight days after that the federal statistical agency responsible for forecasting American electricity demand cut its Texas number by more than half. The Probability Desk weights a narrower gate reopening between 1 November 2026 and 30 April 2027 at 55%, a fast reopen on or before 31 October at 10%, and the gate staying shut past 30 April 2027 at 35% — a band that includes ERCOT missing the nine-month outer bound it gave the Public Utility Commission on 14 August. Our disagreement with the observable consensus is not about how many megawatts slip. It is about what a universal owner should now assume about political risk in domestic infrastructure — an asset class this audience has been accumulating for a decade precisely because it was supposed to have none.
The Trigger
3 August 2026. Governor Greg Abbott wrote to Public Utility Commission of Texas Chairman Thomas Gleeson and ERCOT Chief Executive Pablo Vegas directing "a comprehensive verification and audit of all data centers advancing through ERCOT's interconnection process before any additional data centers are approved to move forward." The letter states that ERCOT "is currently considering approximately 474 gigawatts of requests… That is more than five times Texas' record peak electricity demand for ERCOT," and that "approximately 90 percent of the new power requests are data centers." Projects that fail to comply "must be denied connection to the Texas grid." The stated provocation was non-compliance with a PUC water-and-power survey conducted under the General Appropriations Act. The letter references an earlier directive of 10 June 2026. It sets no completion date — a point Jones Day flagged in its 7 August client note and one that does a great deal of work in the model below.
The same afternoon, at 15:22, ERCOT issued Market Notice M-A080326-01. The operator would not, as required, notify each interconnecting Distribution Service Provider and Transmission Service Provider how any Large Load had been classified in the Batch Zero Interconnection Study by 7 August 2026. It would instead seek "a good cause exception related to the timelines and process for Batch Zero set forth in ERCOT Planning Guide Sections 5 and 9," ahead of the Commission's open meeting of 20 August 2026.
Three further dates complete the trigger. On 10 August ERCOT filed that request (PUCT Control No. 59142, Item No. 42). On 11 August the U.S. Energy Information Administration published its August Short-Term Energy Outlook and wrote, in the Highlights: "On August 3, the Texas governor announced a pause on new data center development, and as a result, we have lowered our forecast for electricity demand in Texas. We expect electricity load in Texas will grow by 6% in 2027, in contrast to our forecast of 14% growth in the previous STEO." On Friday 14 August, at a special PUCT open meeting, ERCOT told the Commission the audit covers roughly 250 to 300 projects representing about 200 GW of future demand, and that verification would take "several months" but less than nine months.
That last figure is the spine of this report. It is a number a public institution said out loud, under questioning, about its own conduct. It can be graded.
Recency note. The freshest reporting on this file as of publication is same-day — Washington Examiner (17 August, 08:30 ET) and KERA/Texas Tribune (17 August, 08:34 CT) — but both are write-ups of Friday's meeting. No new substantive development has been reported since 14 August. We say so rather than dress three-day-old facts as this morning's.
The Forecast Question
On what date does ERCOT first issue Batch Zero Large Load classification notifications to interconnecting DSPs and TSPs — the step suspended on 3 August 2026, which had been due 7 August 2026?
Resolution bands, mutually exclusive and collectively exhaustive, clocked from t₀ = 14 August 2026 (the PUCT open meeting at which ERCOT gave its estimate):
| Band | Definition | Desk weight |
|---|---|---|
| FAST REOPEN | notification on or before 31 Oct 2026 | 10% |
| NARROWER GATE | notification 1 Nov 2026 – 30 Apr 2027 | 55% |
| GATE STAYS SHUT | notification after 30 Apr 2027, or the step is superseded by a durable statutory or PUCT-rule restriction | 35% |
Weights sum to 100 and are rounded to 5%.
The 30 April boundary is not arbitrary. The fortnight between our boundary and ERCOT's own bound is not free: the model puts 38.0% beyond 30 April 2027 and 35.9% beyond a literal nine months, so the band is about two points wider than a strict test of ERCOT's statement. We chose the calendar month-end for gradeability and disclose the cost. Nine months from 14 August is 14 May 2027. The tail band is therefore, near enough, ERCOT breaching the outer bound it set for itself — and it also brackets the 9 April 2027 date on which Batch Zero study results are currently due, which Gibson Dunn reported on 14 August is not yet affected. The word carrying the weight in that sentence is yet.
Graded from ERCOT market notices and PUCT Control No. 59142 filings. Resolution date: 30 April 2027.
Prior and Base Rate — and an honest problem
Methodology requires an outside view before a narrative. Here we hit a wall, and we are going to describe it rather than paper over it.
There is no adequate reference class for this event. A US independent system operator suspending a scheduled large-load classification step, at a governor's written direction, with no statutory completion date, on a queue equal to five times system peak, is close to unprecedented. ERCOT is the first ISO in the country to batch-study large loads at all — that process was approved by the PUCT only on 18 June 2026, eleven weeks before it was frozen. You cannot compute a historical frequency for the interruption of a process that is eleven weeks old.
What we have instead are three weak analogues, each stated with what it does and does not license:
New York, 14 July 2026. Governor Hochul's executive order imposed the first statewide moratorium on new hyperscale data centres, pausing discretionary state environmental permits for facilities of 50 MW or more for up to one year (Executive Order 62). Relevant because it is the nearest-in-time executive-action analogue and the only one with a stated duration. Misleading because it is a permitting moratorium under a governor's environmental authority, not the suspension of an ISO's own study process, and because "up to one year" is a ceiling rather than an expectation. Texas is the second domino here, not the first.
Abbott's own escalation ladder. The 10 June letter directing the PUC to ensure data centres pay their own interconnection costs did not stop the queue. The 3 August letter did. Relevant because it establishes that this governor escalates when the first instrument does not produce compliance, which is an argument for duration. Misleading because two data points are not a rate.
The Batch Zero design process itself. More than 80 stakeholder meetings, workshops averaging some 500 participants, roughly 200 survey responses and 290-plus written comments, from initiation to ERCOT Board adoption on 2 June 2026 and PUCT approval on 18 June 2026. Relevant as a measure of how long ERCOT's own consensus machinery takes on large-load questions — a year-plus, not a quarter. Misleading because designing a rule and verifying a project list are different tasks with different critical paths.
Consequence for the model. We set the base-rate leg near-uninformative at 12 / 48 / 40, tilted modestly long for one documented reason only: the directive sets no completion date, and administrative processes without a deadline do not, as a class, resolve early. And we down-weight this leg from the standard 0.40 to 0.30, logging the reason as the methodology requires: the reference class is too thin to carry standard weight. The freed 0.05 goes to the expert-prior leg, where the evidence actually is.
We would rather publish a wide, honestly-labelled prior than a precise, invented one.
Evidence-Update Table
No probability appears in this report without the evidence that moved it. Prior is the base-rate leg above.
| # | Evidence | Date | Source | Direction | Strength | Effect on the tail band |
|---|---|---|---|---|---|---|
| 1 | ERCOT: verification will take "several months, but less than nine months" | 14 Aug 2026 | ERCOT testimony, PUCT open meeting, via Gibson Dunn | ↓ tail | High | 40% → 34%. The operator has put a public ceiling on itself. Institutions rarely volunteer a bound they expect to breach. |
| 2 | The directive sets no completion date | 3 Aug 2026 | Abbott letter; flagged by Jones Day, 7 Aug | ↑ tail | High | 34% → 38%. The nine months is ERCOT's estimate, not the Governor's commitment. Only one of those two parties can restart the queue. |
| 3 | Batch Zero study results still due 9 Apr 2027, "no imminent impact (yet)" | 14 Aug 2026 | Gibson Dunn | ↓ tail | Med-High | 38% → 35%. A live downstream deadline is a forcing function; ERCOT has an operational reason to finish. |
| 4 | BNEF models a three-month delay as its working case; 49.8 GW exposed | 5 Aug 2026 | BNEF via POWER, 6 Aug | ↓ tail | Medium | 35% → 33%. The best-resourced outside estimate is materially shorter than ours. We take the hit. |
| 5 | Audit perimeter is 250–300 projects / ~200 GW, not the full 474 GW queue | 14 Aug 2026 | ERCOT testimony via Texas Tribune | ↓ tail | Medium | 33% → 31%. A bounded scope is a completable scope. This is the single most under-reported fact of the week. |
| 6 | Abbott's named 2027 legislative priorities: codify cost obligations, require new capacity not just demand, mandate water-efficient cooling, repeal the sales-tax exemption | reported 10 Aug 2026 | POWER; MultiState, 22 Jun | ↑ tail | High | 31% → 36%. The 90th Legislature convenes 12 Jan 2027. A hardening instrument inside the resolution window is the tail's main engine. |
| 7 | Three independent polls agree within one point: ~70% oppose a local data centre | Mar / May / Jul 2026 (field dates) | Gallup; Heatmap Pro–Embold; Fox–Beacon/Shaw | ↑ tail | High | 36% → 38%. Data centres now poll worse than a local nuclear plant (53% opposed, Gallup). Three houses, one answer, is not noise. |
| 8 | Hyperscaler capex guidance set in the eleven days before the pause — Alphabet to $195–205bn, Amazon to ~$220bn, Microsoft reiterated, Meta narrowed to $130–145bn — has not been withdrawn since, and operators moved to comply rather than exit (6–10 Aug) | 22–30 Jul 2026 | company releases and calls | ↓ tail | Medium | 38% → 36%. Every one of these figures predates 3 August, so none of it is a response to the pause; what is post-pause is that no guidance has been withdrawn and operators moved to comply. That is pressure to reopen — jobs, tax base, and PPAs already signed. The first genuine post-pause test is Q3 guidance in late October (Watch #11). |
| 9 | Nobody has cancelled a Texas project over the pause; Meta committed $14bn to El Paso with BlackRock on 28 Jul, and operators moved to comply rather than exit | 28 Jul – 10 Aug 2026 | company/state announcements via POWER | ↓ tail | Medium | 36% → 35%. Compliance is cheaper than relocation. Revealed preference matters more than survey sentiment. |
| 10 | ERCOT's own 2032 peak forecast of 367,790 MW is disowned by ERCOT — "higher than expected future load growth," Vegas | 16 Apr 2026 | Utility Dive, quoting ERCOT filings | ↑ tail | Medium | Held at 35%. The queue is known to be speculative, which makes the audit substantively harder to finish, not easier — you cannot verify projects that were never real. |
Posterior on the tail band: 35%. Each move is small because each piece of evidence is real but partial. That is what an honest Bayesian update looks like when the file is two weeks old.
The Scenarios
FAST REOPEN — 10%
Classifications issued on or before 31 October 2026.
The Commission grants ERCOT's good-cause exception at the 20 August open meeting with a compressed schedule attached. ERCOT discovers that a large fraction of the 250–300 audited projects either fail the document request outright or never respond — plausible, since the whole provocation was non-compliance with the PUC's survey — and the perimeter collapses to a manageable list within ten weeks. Operators, having already committed publicly to Abbott's standards, deliver the water, power, ownership and community-mitigation data quickly because delay costs them more than disclosure. The Governor, having demonstrated the veto exists, does not need to hold it.
Trigger that would put it in play: a dated schedule in ERCOT's supplementary filing before 20 August. Tripwire: any ERCOT market notice restoring a Planning Guide §5/§9 timeline with a Q4 2026 date. What would falsify it: the Commission taking the good-cause request under advisement without a schedule.
Why only 10%. Ten weeks to audit 250–300 projects across five information categories, in an agency that has just told the Commission it needs "several months," requires ERCOT to beat its own guidance by a factor of two. Institutions that lowball do so in the other direction.
NARROWER GATE — 55% (base case)
Classifications issued between 1 November 2026 and 30 April 2027.
The audit runs the course ERCOT described. The Commission grants the exception, preserves the 9 April 2027 study-results date as an anchor, and ERCOT works the perimeter down through the autumn. The queue that emerges is dramatically smaller and materially different in character: projects with signed firm-power arrangements, on-site generation, air-cooled or closed-loop thermal designs and disclosed ownership pass; speculative duplicate filings — the ones inflating a 474 GW number against a 91,089 MW record peak — do not. The 90th Legislature convenes on 12 January 2027 and files bills, but the gate reopens on ERCOT's clock before any of them reach the Governor's desk.
This is not a defeat for the AI build. It is a re-underwriting of it. The megawatts that survive are the ones that were always real, and they arrive with cost obligations attached.
Triggers: the exception granted 20 August with the April 2027 study date intact; ERCOT publishing an audit methodology or tranche schedule in September. Tripwires: classification notices restored in any month Nov–Apr; a published audit completion percentage; the November Quarterly Stability Assessment including the seventeen Large Loads ERCOT flagged as potentially eligible. What would falsify it: an audit-completion estimate revised past nine months, or a filed bill with a load cap moving out of committee before March.
GATE STAYS SHUT — 35% (the tail — high for a tail, and the branch we spend most of this report testing)
Classifications not issued by 30 April 2027, or superseded by a durable restriction.
Two distinct routes, and it is worth separating them because they resolve differently.
Route A — the audit simply overruns. ERCOT is verifying a queue that its own leadership has said in a regulatory filing is inflated. Chad Seely told the Commission that ERCOT "has concerns with using the preliminary load forecast values" behind the 367,790 MW figure; Pablo Vegas said flatly, "we believe this forecast to be higher than expected future load growth." Auditing a speculative queue is harder than auditing a real one: non-responses have to be chased, duplicate filings across service territories have to be reconciled, and ownership chains on special-purpose vehicles have to be unwound. The 474 GW in Abbott's August letter against the 438,000 MW ERCOT published on 18 June is itself a 36 GW discrepancy in six weeks that no source we found reconciles. Nine months is an estimate made two weeks into a task nobody has performed before.
Route B — politics overtakes process. The Legislature convenes 12 January 2027 with the Governor's priorities already public: codify cost obligations, require new data centres to add generating capacity rather than only demand, mandate water-efficient cooling, impose annual reporting, repeal the sales-tax exemption. Three independent polls put local opposition near 70%, and same-day reporting on 17 August has data centres as a live issue in the 2026 races. A legislature that convenes into that sentiment does not quietly ratify the status quo. If any instrument passes, the classification step does not resume on ERCOT's clock — it resumes on a new rule's clock, after rulemaking.
Triggers: an ERCOT filing revising the completion estimate; a pre-filed bill with a load cap or an interconnection condition; a second gubernatorial letter. Tripwires: the 9 April 2027 study date slipping; the November QSA excluding the seventeen flagged Large Loads; committee passage of any large-load bill before 1 March 2027. What would falsify it: classifications restored, or the Legislature adjourning sine die in May 2027 with no instrument enacted.
The Monte Carlo — Simulation Results
A real simulation was run. The code is published as mc.py alongside this report and reproduces bit-for-bit under seed 20260817, numpy 2.2.6, 50,000 paths.
Structure. Four latent factors are drawn under a Gaussian copula — audit complexity (z₁), opposition intensity (z₂), demand pull (z₃) and administrative slip (z₄) — with corr(z₁,z₂) = +0.20 and corr(z₂,z₃) = +0.15. The second correlation encodes a real mechanism: opposition and capex are both driven by build-out intensity, so more spending produces more backlash.
The reopen delay is D = LogNormal(ln 5.0, 0.42) · exp(+0.22·z₂) · exp(−0.14·z₃) · LogNormal(ln 1.15, 0.28) with a hardening branch: p = logistic(−1.55 + 0.75·z₂ − 0.25·z₃); where it fires, the reopen is pushed to at least sine die of the regular session (31 May 2027) plus Exponential(3.0) months of rulemaking.
Median audit length of five months is a plain reading of ERCOT's "several." The opposition coefficient is sized off the three-poll agreement; the demand coefficient is smaller because a governor's office is not obliged to answer to capex; the 15% median slip is a judgment about self-reported regulatory schedules.
Results.
| Band | Probability |
|---|---|
| FAST REOPEN (≤ 31 Oct 2026) | 7.6% |
| NARROWER GATE (1 Nov 2026 – 30 Apr 2027) | 54.3% |
| GATE STAYS SHUT (> 30 Apr 2027) | 38.0% |
| Percentile | Delay from 14 Aug 2026 | Implied date |
|---|---|---|
| P5 | 2.23 months | 20 Oct 2026 |
| P10 | 2.79 months | 6 Nov 2026 |
| P25 | 4.11 months | 17 Dec 2026 |
| P50 | 6.60 months | 2 Mar 2027 |
| P75 | 10.77 months | 7 Jul 2027 |
| P90 | 14.52 months | 29 Oct 2027 |
| P95 | 17.43 months | 26 Jan 2028 |
Diagnostics: mean probability of a hardening instrument 19.9%; 58.4% of paths reopen before the 9 April 2027 study date; 35.9% of paths run past ERCOT's own nine-month bound.
The uncertainty we publish is the specification envelope, not the sampling CI. The sampling confidence interval on the base leg is ±0.44pp, which is a statement about how many paths we drew and nothing whatsoever about whether the model is right. Across twelve alternative specifications and the central case — thirteen in all — audit median 4.0 to 6.5 months, sigma 0.30 to 0.55, no slip to 30% slip, opposition off to doubled, demand pull doubled, hardening base rate 8% to 30%, hardening adding six months rather than three — the bands run:
| Band | Envelope | Desk view |
|---|---|---|
| FAST REOPEN | 3.1% – 14.2% | 10% |
| NARROWER GATE | 46.4% – 60.6% | 55% |
| GATE STAYS SHUT | 30.3% – 50.5% | 35% |
All three published weights sit inside the envelope. No specification we could defend puts the tail below 30%. The single most influential parameter is the audit median: moving it from 5.0 to 6.5 months lifts the tail from 38% to 50.5%. That is where a reader who disagrees with us should push. (The envelope's own "central" row reads 38.2% rather than 38.0% because the sensitivity loop re-draws its Bernoulli and exponential variates from an advanced generator state; it is the same specification resampled, and the 0.2-point difference is sampling noise of exactly the size the ±0.44pp interval describes.)
Limitations, stated plainly. Every parameter in this model is judgmental. None is fitted. We could not locate a defensible fitted base rate for ISO administrative-process slippage inside this run's evidence window, and we did not invent one. The model prices a date, not a dollar — it says nothing directly about how much of the 200 GW ultimately energises. It is single-jurisdiction and does not price FERC pre-emption. The hardening branch is binary, whereas a partial instrument — cost allocation without a load cap — could let the gate reopen while still repricing everything behind it. And the nine-month bound on which the tail boundary rests is oral testimony reported by a law firm; we could not open the underlying filing or a transcript.
Market vs. Desk View
There is no traded instrument on this question, so we will not pretend to a market-implied number. What exists is a set of revealed positions, and they point one way.
| Observable | Reading | Implied stance |
|---|---|---|
| Alphabet 2026 capex guidance, 22 Jul | raised to $195–205bn from $180–190bn, itself raised in April from $175–185bn | no slowdown priced |
| Amazon, 30 Jul | raised to ~$220bn on the call, from ~$200bn — attributed to higher memory and component costs rather than added capacity; AWS +36.7%, fastest in 18 quarters | no slowdown priced |
| Microsoft, 29 Jul | quarterly capex and finance leases $41bn, +69%; guidance reiterated | no slowdown priced |
| Meta, 29 Jul | 2026 capex narrowed to $130–145bn; $14bn El Paso venture announced 28 Jul | no slowdown priced |
| BNEF, 5 Aug | central case is a three-month delay; 1.2 GW of ERCOT additions at stake to 1Q27 | ~90% reopen within two quarters |
Taken together these amount to an implied view somewhere near 30 / 60 / 10 — but that is our conversion of revealed positions into bands, on the same basis and with the same caveat as the expert-prior leg: none of these institutions published probabilities over these bands, and a reader is entitled to disagree with the conversion before disagreeing with the conclusion. On our conversion the Desk sits at 10 / 55 / 35, roughly 25 percentage points heavier on the tail. We note for honesty that our own expert-prior leg, built from an overlapping evidence set, puts the tail at 29% rather than 10% — the spread between those two readings is a measure of how soft any "market" number on this question has to be. BNEF's three-month central case, clocked from 3 August, implies a reopen in early November — inside our NARROWER GATE band, not FAST.
Why we think the consensus is wrong, and why it is also reasonable. It is reasonable because on any capex-relevant measure this event is immaterial. BNEF's own arithmetic is the cleanest statement of that: only 1.2 GW of ERCOT data-centre capacity was scheduled to arrive between 2Q26 and 1Q27, and the widely-quoted $15 billion is the top of a range that requires a 100% AI-compute mix and a full delay. BNEF's supermajority case is just over $8 billion; at a 10% AI mix it is about $2 billion. Against Alphabet's single-year capex guidance, the central case is a rounding error. The $15 billion figure is the ceiling of that range, not its centre, and it should not be read as a central estimate — a distinction BNEF makes and the trade coverage generally preserved.
We think the consensus is wrong because it is pricing the megawatts and the story is the mechanism. Three things happened in eleven days that had not happened before: an ISO suspended its own scheduled process on a letter rather than an order; a federal statistical agency halved a state demand forecast — 14% to 6% — and named a governor's press announcement as the cause; and the most permissive large-load jurisdiction in the country demonstrated that the permission was discretionary all along. The highest-conviction mispricing is not in power or AI equities. It is in the political-risk premium embedded in domestic infrastructure valuations — assets bought by this readership at multiples that assume regulatory permission is a formality.
What would prove the Desk wrong: classifications restored before 31 October on a schedule filed this week, the Legislature convening in January without a large-load bill of consequence, and the November Quarterly Stability Assessment including all seventeen flagged Large Loads. Any two of those three and our 35% is too high.
Universal-Owner Portfolio Heatmap
Strategic, not advice. Direction is the reprice under the tail scenario; magnitude bands are the Desk's judgment, not model output.
| Asset class | Base | Tail (gate stays shut) | Magnitude, 12m | Mechanism |
|---|---|---|---|---|
| US regulated utilities (ERCOT-exposed T&D) | neutral | negative | −3 to −8% | Deferred load growth defers rate-base growth |
| Independent power producers / merchant gas | mildly negative | negative | −5 to −12% | Signed and pending large-load PPAs slip; the scarcity trade cools |
| Hyperscaler equities | neutral | mildly negative | −1 to −3% | Immaterial to capex; material to siting optionality |
| Semis / AI hardware | neutral | neutral | ±1% | Demand is compute-constrained, not Texas-constrained |
| Unlisted digital infrastructure | mildly negative | strongly negative | −10 to −20% | Entry multiples assume permission. Repricing here is slow, unhedged and marked late |
| Infrastructure debt / project finance | neutral | negative | spread +25 to +75bp | Morningstar DBRS has already flagged stakeholder opposition as a potential material credit factor |
| Long-duration rates | neutral | mildly positive | −5 to −15bp | Lower load growth is lower capex is lower issuance |
| US inflation-linked | neutral | mildly negative | — | The retail-electricity pass-through is the contested channel; see red-team |
| Water rights and utilities | positive | positive | +3 to +8% | Cooling-water scrutiny is now a screening criterion, not a footnote |
| Gulf and Asian sovereign digital-infra programmes | neutral | positive | — | Jurisdictional arbitrage: capital routes to where permission is credible |
| Texas municipal credit | neutral | mildly negative | — | Deferred assessed valuation in host counties |
The line that matters for this readership. A universal owner is on both sides of this trade simultaneously — long the hyperscalers, long the utilities and IPPs that serve them, long unlisted digital infrastructure, long the water, and ultimately accountable to beneficiaries who pay the retail electricity bill. There is no security selection that resolves that. It is the definitional universal-owner problem, and the only instruments that address it are stewardship and policy engagement, not allocation.
Second- and Third-Order Effects
First order. Roughly 200 GW of Texas interconnection requests is halted pending verification; the EIA cuts Texas 2027 load growth from 14% to 6%.
Second order. Jurisdictional arbitrage. Texas and New York are now both gated; capital routes to states that have not yet discovered the veto — and, at the margin, offshore to Gulf and Asian programmes where sovereign sponsors control both permission and power. Cost allocation moves to the front of the queue. Oklahoma's HB 2992 passed both chambers unanimously on 5 May 2026, defining large-load customers at 75 MW and up and requiring landowner and Corporation Commission notice within 60 days of land acquisition. A unanimous red-state cost-allocation statute is a stronger contagion signal than any moratorium bill. Behavioural adaptation. Google's Meitner Energy Center is over 1 GW paired with its own wind, solar, storage and on-site gas, and is air-cooled rather than evaporative. Chevron and Engine No. 1's Project Kilby is 2.67 GW of co-located generation in West Texas under a 20-year PPA with Microsoft. The industry's answer to the veto is to stop being a load.
Third order. The speculative queue gets marked. If verification reveals that a large share of 474 GW was duplicate or non-serious filings, every load forecast built on interconnection-queue data — including capacity-market procurement — is impeached. PJM's Independent Market Monitor has already put a number on the analogous exposure: data centres accounted for $6.5bn, or 40%, of the $16.4bn in costs from the 2027/28 auction, of which roughly $6.2bn relates to data centres not yet built; across the last four base auctions, $29.4bn, or 46%, of $63.6bn. Joseph Bowring's objection is on the record: "PJM is continuing to act like it's business as usual… it is really a paradigm shift, and failing to do that imposes costs on other customers." Political risk re-enters domestic infrastructure underwriting. The category was accumulated on the premise that OECD permission risk is de minimis. And a governance question lands on this readership: when a portfolio company's growth plan depends on a discretionary permission that a single executive can withdraw by letter, that is a disclosure question for the stewardship team, not a footnote for the deal team.
Watch Dashboard
| # | Indicator | Reading now | Threshold that moves the model | Source |
|---|---|---|---|---|
| 1 | PUCT open meeting, 20 Aug 2026 | good-cause request pending | Granted with a dated schedule → FAST +10pp. Taken under advisement → tail +5pp | PUCT Control No. 59142 |
| 2 | ERCOT supplementary filing on timing | promised, not filed | Any completion estimate beyond nine months → tail +10pp | ERCOT / PUCT interchange |
| 3 | Batch Zero study-results date | 9 Apr 2027, intact | Any slip → tail +15pp | ERCOT market notices |
| 4 | Nov 2026 Quarterly Stability Assessment | 17 Large Loads flagged as potentially eligible | Fewer than 8 included → tail ↑ | ERCOT QSA |
| 5 | Classification notices to TSPs/DSPs | suspended since 3 Aug | This is the resolution variable | ERCOT market notices |
| 6 | Texas 90th Legislature, convenes 12 Jan 2027 | pre-filing open | A load cap or interconnection condition out of committee before 1 Mar → tail +15pp | Texas Legislature Online |
| 7 | Governor's 2027 priorities | five items named | A sales-tax-exemption repeal bill filed → tail ↑ | Governor's office |
| 8 | Audited perimeter | 250–300 projects, ~200 GW | Revised up → tail ↑; down → base ↑ | ERCOT testimony |
| 9 | Queue total | 474 GW (3 Aug) vs 438 GW (18 Jun) | A reconciled number below 400 GW → the write-down is real, base ↑ | ERCOT / Governor's office |
| 10 | ERCOT peak demand | 91,089 MW, 22 Jul 2026 | A new record before October raises the reliability argument for verification | ERCOT records page |
| 11 | Hyperscaler Q3 capex guidance | all raised or held | Any cut citing siting or permitting → the consensus has moved to us | company releases, late Oct |
| 12 | Texas project cancellations | none attributed to the pause | Two or more → demand-side capitulation, base ↑ | company announcements |
| 13 | Other-state actions | NY (Jul), TX (Aug) gated | A third state gating interconnection → contagion confirmed | state executive/legislative records |
| 14 | Local opposition polling | ~70% opposed, three houses | Sustained above 65% into the session → tail ↑ | Gallup / Heatmap / Fox |
| 15 | Infrastructure credit spreads on digital-infra project debt | DBRS has flagged the risk | Widening on a Texas-specific basis → the mispricing is closing | rating agency actions |
Red-Team: How This Could Be Wrong
1. We may be reading a procedural pause as a political turn. Gibson Dunn's assessment is explicit: "it is not a moratorium on data center development." ERCOT is moving verification to the front of a study process it always intended to run — Chad Seely's own framing was "moving that verification process now to the front of the line." Abbott declined to propose a moratorium in June when he had the opening. If this is queue hygiene wearing a governor's letterhead, our 35% tail is roughly double what it should be, and the base case should be closer to 70%. This is the strongest argument against the report and we have not defeated it. Our answer is narrow: queue hygiene does not normally require suspending a deadline four days out, and the EIA does not normally halve a state forecast for it.
2. The ratepayer-harm premise underneath the politics is contested — and we did not put a number on it for that reason. The intuition that data centres raise household bills is not settled. A Rutgers New Jersey State Policy Lab paper is titled, near enough, Are Data Centers Raising Your Electric Bill? Mostly Not. Yet.; EPRI has found downward price pressure through 2024; Marketplace reported in July 2026 that data centres have lowered bills in some jurisdictions. Widely-circulated figures — bills up 267% near data centres, Virginia +13%, Ohio +12% — trace to sources we could not open and would not cite. If the harm is smaller than voters believe, the political energy may dissipate faster than we assume.
3. Our tail may be the model's construction rather than the world's. The hardening branch contributes most of the mass beyond nine months, and its base rate — 17.5% at the median draw — is a judgment with no fitted counterpart. Set it to 8% and the tail falls to 31.4%. A reader who thinks Texas legislatures do not restrict Texas energy development has a real argument and can move our number by seven points on one parameter.
4. Two premises we started with turned out to be false, and we are recording it. Oklahoma SB 1488 — a 100 MW moratorium to November 2029 — is widely described as law. It is not. We verified only that it was filed on 22 January 2026; it never received a committee or floor vote and died at sine die on 14 May 2026, per two independent legislative trackers, though we could not confirm it from a primary legislative record. Separately, the local-moratorium counts circulating for 2026 (533 instruments across 42 states; Ohio 35, Michigan 34) come exclusively from aggregators we would not cite, and we have therefore used none of them. The defensible counts are MultiState's — 300-plus data-centre bills in 30-plus states in the first six weeks of 2026, at least 18 states with large-load rate-class bills — and Good Jobs First's, at least 12 in-session states with filed moratorium bills. Our contagion evidence is thinner than the headlines suggest, and that cuts against our own tail.
5. FERC is not in the model. Federal pre-emption of state large-load restrictions is a live route that would collapse the tail quickly. We could not price it and left it out, which biases the tail up.
6. The nine-month bound is second-hand. It is oral testimony at an open meeting, with no transcript in our evidence set — though it is now reported consistently by a law firm and by the wire coverage of the meeting, which is weaker than a transcript and stronger than a single source. The tail boundary is built on it. If ERCOT never said it, or said it with qualifications the report omitted, the band definitions need revisiting.
Methodology Box
Weights are the Probability Desk's, formed by a written aggregation rule across three inputs: base rate 0.30 (down-weighted from the standard 0.40 with the reason logged — the reference class is too thin to carry standard weight), expert priors 0.35 (raised from 0.30 by the 0.05 freed above), and a real 50,000-path Monte Carlo 0.35 (occupying the multi-agent-simulation slot; MiroFish was not run for this episode and is not implied anywhere in this report).
The three legs produced 12/48/40, 14/57/29 and 7.6/54.3/38.0 respectively, aggregating to 11.2/53.4/35.4 and rounding to 10/55/35.
One disclosure the expert-prior leg requires: the 14/57/29 figures are the Desk's readings of published positions held by ERCOT, BNEF, Gibson Dunn and Jones Day. None of those institutions published probabilities over these bands. We converted their stated positions into band weights ourselves, and a reader is entitled to disagree with the conversion before disagreeing with the conclusion.
Live market data via FRED (Brent DCOILBRENTEU $93.26, 11 Aug 2026; US 10-year DGS10 4.63%, 13 Aug 2026) is reported for context and does not enter this model. Source ledger: 50 entries, of which 38 were opened and read directly. Every figure in this report traces to a dated source; where a figure could not be verified it has been excluded and the exclusion recorded in the red-team.
Probabilities are the UAO Probability Desk's, weighted across base-rate, expert-prior and simulation inputs. Methodology available on request.
Editorial scenario analysis only. Not investment, actuarial, or geopolitical advice.
Source Ledger
Primary sources are marked P. Fifty entries were assembled; the load-bearing thirty are listed.
| # | Source | Date | Point used |
|---|---|---|---|
| 1 P | Abbott letter to PUCT Chairman Gleeson and ERCOT CEO Vegas | 2026-08-03 | 474 GW; "more than five times" record peak; ~90% data centres; audit before any further approvals; no completion date |
| 2 P | ERCOT Market Notice M-A080326-01 | 2026-08-03 | Batch Zero classification notifications suspended; good-cause exception; PUCT open meeting 20 Aug |
| 3 P | EIA Short-Term Energy Outlook, August 2026 | 2026-08-11 | Texas 2027 load growth cut from 14% to 6%, the pause named as cause |
| 4 P | EIA STEO full report | 2026-08-11 | 1H26 US generation +1.8%; solar +21%, wind +6%, hydro +9% |
| 5 P | ERCOT all-time peak demand records | updated 2026-08-06 | Record 91,089 MW, 22 Jul 2026; prior 85,508 MW, 10 Aug 2023 |
| 6 P | ERCOT, New Batch Connection Process for Large Electricity Users | 2026-06-18 | 438,000 MW queue; ~90% data centres; PGRR145/NPRR1325; Board 2 Jun, PUCT 18 Jun; 75 MW threshold |
| 7 P | PUCT approves ERCOT's Batch Zero process | 2026-06-18 | Approval date |
| 8 | Gibson Dunn — What Abbott's directive means for Batch Zero | 2026-08-14 | ERCOT filed 10 Aug (Control No. 59142, Item 42); "several months but less than nine months"; 9 Apr 2027 study date "not yet" affected; six Large Loads in the Aug QSA, seventeen for Nov; "not a moratorium" |
| 9 | Texas Tribune — Texas will audit up to 300 projects | 2026-08-14 | 250–300 projects, ~200 GW; 1,800+ projects overwhelmed the queue; Seely quote |
| 10 | KERA News (Texas Tribune syndication) | 2026-08-17 08:34 CT | Special PUCT open meeting confirmed for Friday 14 Aug |
| 11 | Washington Examiner — pause will take "several months" | 2026-08-17 08:30 ET | Freshest dated item; reconfirms scope and duration |
| 12 | Jones Day — Texas pauses data center approvals, stops short of a ban | 2026-08-07 | "The directive sets no completion date"; scope ambiguity as of 7 Aug |
| 13 | Greenberg Traurig — Texas imposes new audit requirement | 2026-08-08 | What Batch Zero replaced; repeated-study problem |
| 14 | Utility Dive — Facing an estimated 474 GW, Texas hits pause | 2026-08-05 | Independent reproduction of the letter's figures; New York precedent; Data Center Coalition response |
| 15 | POWER — Texas audit could delay 49.8 GW, BNEF warns | 2026-08-06 | BNEF (report 5 Aug): 49.8 GW exposed = ~20% of a 253 GW US pipeline; 1.2 GW of actual ERCOT additions 2Q26–1Q27; three-month delay modelled; $8bn at a 60% AI mix, ~$15bn only at 100%, ~$2bn at 10% |
| 16 | POWER — Operators commit to Abbott's standards | 2026-08-10 | Governor's 6 Aug release; Google Meitner Energy Center >1 GW, air-cooled; Chevron/Engine No. 1 Project Kilby 2.67 GW under a 20-year Microsoft PPA; Diode Ventures abandoned Henderson County 23 Jul; Abbott's five 2027 legislative priorities |
| 17 | Utility Dive — ERCOT says Texas demand could quadruple but cautions the forecast may be inflated | 2026-04-16 | 367,790 MW 2032 forecast; Seely and Vegas disowning it; summer 2026 projection 90,500–98,000 MW |
| 18 | Bracewell — Texas SB 6 overhaul of large-load interconnection | 2025 | SB 6 signed 20 Jun 2025; 75 MW threshold; mandatory curtailment capability for loads interconnected after 31 Dec 2025 |
| 19 P | Texas SB 6 (89R) bill text | 2025 | Statutory authority behind the queue data |
| 20 | Utility Dive — Data centers drove $6.3B in PJM capacity auction costs | 2026-07-20 | IMM: $6.3bn / 38% of $16.4bn (2028/29); $29.4bn / 46% of $63.6bn across four auctions; Bowring quote; DBRS credit warning |
| 21 | Utility Dive — Data centers were 40% of PJM capacity costs | 2026-01-07 | IMM: $6.5bn / 40% of $16.4bn (2027/28), of which ~$6.2bn for data centres not yet built |
| 22 P | PJM 2028/2029 Base Residual Auction report | 2026-07-14 | 138,318 MW at the $325.00/MW-day cap; ~6,821 MW short; 14.7% reserve margin |
| 23 P | PJM — 2026/27 auction procures 134,311 MW | 2025-07-22 | Record $329.17/MW-day, +22% |
| 24 | S&P Global — PJM clears at $269.92/MW-day | 2024-07-30 | The $28.92 → $269.92 → $329.17 ladder |
| 25 P | Gallup — Americans oppose AI data centers in their area | 2026-05-13 | Fielded 2–18 Mar, n=1,000 adults: 71% opposed (23% somewhat + 48% strongly); vs 53% opposed to a local nuclear plant |
| 26 P | Heatmap — Americans now overwhelmingly oppose new data centers near them | 2026-06-02 | Fielded 15–28 May 2026 (published 2 Jun), n=4,118 registered voters: 71% opposed (55% strongly), from 42% in Sep 2025; ≥20 projects cancelled on backlash in Q1 2026, >$41bn and ≥3.5 GW |
| 27 | Fox News poll — voters oppose data centers by 40 points | fielded 17–20 Jul 2026 | 70–30 opposition, n=1,003 |
| 28 P | Governor Hochul — first statewide hyperscale moratorium | 2026-07-14 | New York pauses state environmental permits for ≥50 MW facilities for up to one year |
| 29 P | Oklahoma House — ratepayer protection bill advances to the Governor | 2026-05-06 | HB 2992 passed both chambers unanimously 5 May; large load defined at ≥75 MW; 60-day notice duty |
| 30 | MultiState — state data center policy shifts as governors impose new restrictions | 2026-06-22 | Arizona sales-tax-exemption moratorium to Jun 2029; Illinois credits paused from 1 Jul; Ohio comparable; Abbott declined a moratorium in June |
| 31 | MultiState — state data center legislation in 2026 | 2026-02-20 | 300+ bills in 30+ states in six weeks; 18 states with large-load rate-class bills |
| 32 | Good Jobs First — data center moratorium bills are spreading | 2026-02-19, upd. 2026-03-09 | At least 12 in-session states with filed statewide moratorium bills |
| 33 P | Amazon Q2 2026 results | 2026-07-30 | AWS +36.7%, fastest in 18 quarters; TTM free cash flow an outflow of $7.6bn vs +$18.2bn, on a $66.1bn rise in property and equipment |
| 34 | CNBC — Amazon hikes 2026 capex to $220 billion | 2026-07-30 | ~$220bn from ~$200bn, guided on the call, not in the release; AWS to double power capacity by end-2027 |
| 35 | CNBC — Alphabet Q2, capex guidance raised | 2026-07-22 | 2026 capex to $195–205bn from $180–190bn; Q2 capex $44.9bn, +100%; Cloud +82% |
| 36 | CNBC — Microsoft boosts capital spending plans | 2026-07-29 | Capex and finance leases $41bn in the quarter, +69%; ~$175bn capex-and-finance-leases for the period; building useful life extended 15 → 25 years |
| 37 | CNBC — Meta stock drops on guidance | 2026-07-29 | 2026 capex narrowed to $130–145bn; free cash flow $784m vs $8.55bn; $14bn El Paso venture with BlackRock announced 28 Jul |
| 38 | Rigzone, reporting the EIA Hourly Electric Grid Monitor | 2026-08-10 | ERCOT hourly peak 91.1 GW on 22 Jul at 18:00 CT, 48% gas / 32% solar |
Live macro context from FRED via the Desk's data spine: Brent (DCOILBRENTEU) $93.26 at 11 Aug 2026; WTI $84.77; US 10-year (DGS10) 4.63% at 13 Aug; 30-year (DGS30) 5.21%; VIX 14.25 at 14 Aug; upper bound of the federal funds target range (DFEDTARU) 3.75%.
The Probability Desk publishes its calibration. Every scenario above is logged with its weights, its resolution window and its tripwires, and is graded when the horizon arrives. This one resolves on 30 April 2027.
Continue the briefing. Read the daily brief · watch the daily video briefing · listen to The Universal Owner · view the chart of the day.
Produced and edited by the UAO editorial desk. Not investment advice.