Texas quietly imposed the most consequential brake on the U.S. artificial intelligence buildout so far. Governor Greg Abbott’s late-August directive tells the Public Utility Commission of Texas (PUCT) and ERCOT, the state’s grid operator, to freeze new data-center interconnections and audit every request already in the queue. The reason: the numbers coming in stopped looking like real projects and started looking like optionality.
Texas is not alone. Pennsylvania’s Governor Josh Shapiro signed a similar order in August, and Ohio quietly repriced study fees to $100,000, which caused reported data-center demand at AEP Ohio to fall by more than half. What was billed as an infinite AI power boom is meeting the same reality every prior capacity cycle has met: the queue is much bigger than the actual pipeline.
The 474 GW Number
ERCOT’s interconnection queue for large loads — the vast majority tied to AI data centers, crypto mines, and hyperscaler campuses — has ballooned from roughly 48 GW in 2023 to more than 474 GW today, according to Reuters’ reporting on the Abbott order. For reference, the entire U.S. data-center fleet consumes an estimated 60–70 GW at any given moment. Texas alone is being asked to plan for roughly seven times national existing usage.
The other big regional grids tell a similar story. Ten major utilities across the Midwest, Mid-Atlantic, and South have logged roughly 270 GW of new large-load requests. Combine the queues nationally and the number tops 700 GW — an order of magnitude beyond what the industry currently draws.
| Grid / Utility | Requested Load | Notes |
|---|---|---|
| ERCOT (Texas) | 474+ GW | Up from ~48 GW in 2023 |
| 10 Utilities (Midwest, Mid-Atlantic, South) | ~270 GW | Cumulative large-load requests |
| U.S. National Total (Approx.) | 700+ GW | Vs. ~60–70 GW current data-center use |
| Exelon (post-screening) | 11 GW | Down ~40% after credit and milestone checks |
| AEP Ohio (post-fee) | – | Down >50% after $100K study-fee rule |
Ghost Demand Isn’t a New Phenomenon
The industry term for it is “phantom” or “ghost” demand — developers filing interconnection requests at multiple utilities for the same eventual campus, hedging against permitting or land uncertainty. It’s rational for the developer and terrible for the grid planner, who has to size transmission, generation, and reserves against numbers that would double-count.
Screening tools cut through the noise fast. Exelon disclosed that after tightening its qualification criteria — requiring meaningful financial deposits and site-control milestones — its “high probability” data-center load estimate fell 40% to about 11 GW. AEP Ohio’s experience was similar: once the state approved a $100,000 study fee for large-load interconnections, aggregate demand requests fell by more than half. Neither utility changed its view on real AI adoption; they simply stopped counting projects that had no real capital behind them.
What the Abbott Order Actually Requires
The Texas directive, executed through the PUCT under chair Thomas Gleeson and coordinated with ERCOT, mandates that developers disclose:
- Ultimate ownership — not just the shell affiliate signing the interconnection paperwork.
- Any state or local taxpayer incentives claimed or expected.
- Water usage and on-site generation plans, addressing both consumption and grid-reliability offsets.
ERCOT has told regulators it aims to complete the audit by December 2026, in time to feed the results into the next long-term load forecast. Until then, no new interconnection agreements above the review threshold move forward.
Pennsylvania and Ohio Are Doing the Same Thing
On August 18, 2026, Governor Josh Shapiro signed an executive order tightening permitting for any data-center load exceeding 25 MW in Pennsylvania. Projects that bring their own generation or storage get preferential treatment; speculative campuses without secured power or committed customers do not. Of more than 100 announced Pennsylvania data-center projects, only about 20 have actually applied for permits under the new regime.
Ohio, through AEP’s tariff change, effectively priced out speculative interconnection filings by requiring the $100,000 study fee upfront. The result: reported load requests collapsed, but the projects that survived look real.
Why This Matters for Stocks
The market has been pricing hyperscaler capex, AI-chip demand, and utility load growth as if the queue were the pipeline. That was already a stretch — the queue and the pipeline are never the same thing — and the state-level cleanup is now formalizing the gap.
Hyperscalers (AMZN, GOOG, MSFT, META, ORCL)
Timing risk, not thesis risk. The largest AI campuses are backed by real customers and real deposits; those clear a Texas audit. Marginal builds — the ones being pitched to raise capital — are the ones that quietly disappear. Expect capex commentary next earnings to emphasize “power-secured” backlog rather than announced GW.
Data-Center REITs (EQIX, DLR)
Structurally advantaged in a “prove-you-can-power-it” regime because both already own power interconnection rights at scale. Speculative colocation entrants get squeezed hardest.
Utilities (EXC, AEP, VST, NRG, CEG)
Load-growth stories that leaned on the raw interconnection queue will need to reset. But utilities that discipline the queue — as Exelon and AEP have done — end up with cleaner earnings-power narratives, which is what regulators reward on rate-case decisions.
AI Infrastructure (NVDA, and the private-financing complex)
Nvidia’s near-term shipments are covered by contracts already booked. The risk is on the 2027–2028 revenue slope that assumes today’s queue converts. If half of the “ghost” GW never materializes, the compounding growth curve narrows without ever missing a quarter.
What to Watch Next
- ERCOT audit results — target December 2026. First hard read on how many of Texas’s 474 GW are backed by real capital.
- PJM’s next long-term load forecast. The Exelon and AEP screening results will flow into this. A downward revision is likely.
- Hyperscaler capex disclosures Q3 2026. Watch for a shift from announced GW to “power-committed” GW in earnings language.
- Data-center REIT leasing spreads. If the queue tightens, pricing power for EQIX and DLR should firm; if the AI demand narrative was overstated, spreads compress.
The “AI needs infinite power” story was always going to collide with the physical grid. What Texas has done is make the collision official, and it forces every serious investor in the AI-infrastructure stack to answer the same question: which of these gigawatts are real?
Sources
- Reuters via Yahoo Finance — Texas halts new data-center power (Sept. 1, 2026)
- Utility Dive — ERCOT aims to complete data-center audit by December (Aug. 20, 2026)
- Utility Dive — Pennsylvania permitting framework for 25+ MW loads (Aug. 2026)
- ERCOT historical load data
Disclosure: This article is for informational purposes only and is not investment advice.