← Briefing history

The cost-allocation fight has moved from tariff dockets to the ballot box and the governor's mansion: Georgia's regulators faced a hostile…

Read-only snapshot of The AI Power Bill

Sep 21, 2026 · 6 findings · ran 6m 5s

The AI Power Bill — Digest

TL;DR

The cost-allocation fight has moved from tariff dockets to the ballot box and the governor's mansion: Georgia's regulators faced a hostile crowd over the OpenAI deal, Virginia's merger fight turned into a $2.25B bidding war for ratepayer goodwill, and Texas froze new data center interconnections outright to audit its own load forecast. Meanwhile, the equity market has already voted — the merchant AI-power trade is down 35-43% from its highs as regulators cap the upside.

The price collar is holding — and failing — at the same time

PJM's capacity market is now pinned at its regulatory ceiling while the underlying scarcity gets worse, which is an unstable equilibrium by construction. The 2028/29 Base Residual Auction cleared at $325/MW-day — the FERC-approved cap — for the third straight time, with a 6,831 MW shortfall against the reliability requirement and only 525 MW of new generation clearing pjm-2027-2028-capacity-auction-price-caplandgate.comutilitydive.com.

"The outcome demonstrates that the current system doesn't work to bring online new capacity or stimulate demand response, the two things we need the most." — Julia Hoos, Aurora Energy Research, via Utility Dive

Without the collar, PJM says the auction would have hit ~$555/MW-day RTO-wide — nearly $30 billion instead of $16.4 billion — and the shield is visibly cracking at the retail level: Delaware and D.C. posted the largest non-Hawaii retail price increases in the country at +13.5% year over year pjm-2027-2028-capacity-auction-price-caplandgate.comutilitydive.com. The next intervention is already queued: PJM's Reliability Backstop Procurement, accelerated to September 2026, carries a proposed cap 71% above the BRA cap, and Pennsylvania's governor has ordered the PUC to develop backstop cost-allocation rules on a fast timeline.

What to watch: Whether Jefferies' call for structural reform of the BRA toward "materially lower prices" gains traction — that's the direct threat to the merchant generators currently booking record capacity revenue.

Ratepayer backlash goes retail: Savannah, Richmond, and the ballot

The cost-allocation debate has escaped the docket and become electoral politics. Georgia's PSC held its first out-of-Atlanta administrative meeting in roughly two decades and got an earful over the OpenAI contract — a $20 billion, 3.2 GW, 25-year Effingham County campus that requires up-front payments, minimum take, and up to 1,000 MW of flexible demand response georgia-power-large-load-contracting-frameworkgeorgiarecorder.comans.orgprnewswire.com.

"I feel like I am the one that has to put on my back all of the health and issues that could come up because of this data center, because of decisions y'all are making to save me ($15) a month. That don't equate." — Raymond Carver, Effingham County resident, via Savannah Morning News

The $15/month "downward pressure" pledge doesn't start until 2029 — after two PSC seats hit the November ballot — and Southern Company is already running negative Q2 free cash flow while Morgan Stanley cut the stock to Underweight on affordability concerns. Virginia's version of the same fight is bigger: NextEra and Dominion doubled residential bill credits to four years of $10/month ahead of the SCC's merger review, funded mostly by shifting away credits that would have gone to large data centers nextera-dominion-merger-virginia-ratepayersvirginiamercury.compilotonline.comutilitydive.com. Opponents aren't buying it — Del. JJ Singh: "The purported benefits are temporary, while the merger is permanent."

What to watch: The November 2026 Georgia PSC elections, where two seats are on the ballot and the OpenAI contract is the flashpoint.

Texas hits the pause button on the load forecast itself

The demand-side story took its sharpest turn yet: Texas imposed a moratorium on new data center interconnections pending a state audit of whether the forecast is real. Gov. Abbott's August 3 directive requires a "comprehensive verification and audit" of every project in ERCOT's queue before any further approvals, with completion targeted by December texas-ercot-data-center-interconnection-auditgov.texas.govercot.comgibsondunn.comutilitydive.com+1.

"The audit, which ERCOT anticipates will take several months to complete, must occur prior to any additional data center being approved to move forward in the interconnection process." — Gibson Dunn analysis

The stakes are the gap between forecast and contract: ERCOT's preliminary long-term forecast projected 278 GW peak in 2029 and 367.8 GW in 2032 — numbers its own Independent Market Monitor said warrant a resource-adequacy constraint on interconnection. This is the demand-side mirror of Ohio's financial requirements, which cut AEP Ohio's data center queue roughly 80% texas-ercot-data-center-interconnection-auditgov.texas.govercot.comgibsondunn.comutilitydive.com+1. If the audit finds phantom load, every Texas-exposed generator's buildout thesis compresses with it.

What to watch: The audit findings and revised Long-Term Load Forecast, both due around December 2026.

Louisiana: the biggest ask, the least transparency

Louisiana remains the extreme case of speed over scrutiny. Entergy's application for 2.26 GW of new gas at Richland Parish to serve Meta's Hyperion — the largest single data-center-driven generation ask in the country — heads to a December LPSC decision under the fast-track "Lightning Initiative," and the Commission voted 3-1 in August to let Meta keep key project information private, overturning a subpoena louisiana-ai-data-center-power-boomfortune.comall4energy.orgfacebook.comthecentersquare.com.

"The data centers that are in our plan are already signed." — Entergy CEO Drew Marsh, Q1 2026 earnings call

That "already signed" claim is the entire ratepayer defense — Entergy argues current customers' bills will be lower because data centers pay their share — but the Cleco fast-track for the Applied Digital load drew a commissioner objection that competitive bidding shouldn't be waived for a "system resource primarily paid for by ratepayers." Whether the LPSC imposes hard minimum-payment and ring-fencing guarantees in December is the test of whether Louisiana's model is a template or a warning.

What to watch: The December LPSC decision and whether hard ratepayer guarantees attach to the Meta Phase 2 gas expansion.

The equity trade de-rates before the demand even lands

Here's the punchline of the cycle: load growth is booming and the AI-power equities are in a bear market anyway. Constellation is 36.6% off its 52-week high, NRG is 43.3% off and at its low, Vistra and Talen are down ~35%, and even regulated beneficiaries (SO, AEP, D) are 10-13% below highs ai-power-trade-deratingpowermag.com. The fundamentals are real — Talen's Q2 capacity revenue jumped from $88M to $237M — but three headwinds are eating the capture: regulators extending the PJM price collar, debt-financed capex (negative Q2 free cash flow at SO, AEP, ETR, and D alike), and single-name execution misses like Talen's $0.16 versus a $3.20 estimate ai-power-trade-deratingpowermag.com. The investable question has shifted from "who has load growth" to "who has contracted, ring-fenced revenue."

What to watch: Whether Jefferies' expected PJM reform to a cost-of-service capacity model lands — it would formalize the regulatory cap on merchant upside.

What surprised us

  • Texas blinked first. The state with the most aggressive data center growth became the first to freeze the queue entirely — an audit of the load forecast is an implicit admission that the 367 GW 2032 number may be fiction texas-ercot-data-center-interconnection-auditgov.texas.govercot.comgibsondunn.comutilitydive.com+1.
  • The market priced the backlash before the regulators acted. A 35-43% de-rate across the merchant complex happened with record capacity revenue in hand — investors are pre-paying for the cost-allocation crackdown ai-power-trade-deratingpowermag.com.
  • Ratepayer benefits are now being funded by taking money from data centers. NextEra-Dominion's expanded credits explicitly shift away data-center-allocated credits — a reversal of the usual "hyperscalers subsidize residential" pitch nextera-dominion-merger-virginia-ratepayersvirginiamercury.compilotonline.comutilitydive.com.
  • The $15/month problem. Georgia's headline ratepayer benefit doesn't start until 2029 — three years of backlash risk priced against a promise that hasn't begun georgia-power-large-load-contracting-frameworkgeorgiarecorder.comans.orgprnewswire.com.

Open threads worth a vote

  • [watch] Virginia SCC decision on NextEra–Dominion merger — hearings begin November 2026, decision expected January 2027
  • [watch] Georgia PSC elections — two seats on ballot amid data center backlash
  • [watch] ERCOT data center audit findings and revised Long-Term Load Forecast

Findings from this cycle

Current topic brief

Shown for context; the brief may have changed since this cycle ran.

Track how the AI data-center buildout flows through to electricity markets, utilities, and ordinary ratepayers — and who profits or pays. That it's happening is well covered; the gap is the regional, investable exposure map: which utilities, grids, and regions absorb the load and where consumer bills rise as a result. Core entities: regulated and merchant utilities signing data-center load (Constellation, Vistra, Talen, NRG, Southern, Dominion, AEP); grid operators and interconnection queues (PJM, ERCOT, MISO); independent power and nuclear/gas supply; and the hyperscaler buyers (Microsoft, Amazon, Google, Meta) signing PPAs. I want to track new load commitments and PPAs, rate-case filings and commentary about cost allocation between data centers and residential ratepayers, interconnection and capacity-auction outcomes, and EIA/FRED data on electricity prices and demand by region. Follow utility earnings calls for load-growth guidance and the capex to serve it. Flag where a region's residential rates are rising to fund AI load, and any divergence between utility load forecasts and what's actually contracted. The thesis: AI's power demand is quietly repricing electricity by region — surface where the cost lands and who captures the upside.