The AI Power Bill — Digest
TL;DR
The regional exposure map this digest was built to find now exists: EIA state data confirms PJM residential rates rising 2–3x the national rate while ERCOT and Georgia stay below trend. Meanwhile, hyperscalers spent the first week of October writing the checks themselves — $4.3B from Google to Constellation's nuclear fleet, a $4.2B DOE loan to Vistra — and the market paid for contracts, not forecasts, with merchant nuclear names bouncing 15–21% while NRG and the regulated complex lagged.
The price map lands: PJM states are paying for everyone's data centers
The dispersion question is answered, and the pattern maps onto the grid, not onto data-center siting. EIA's Electric Power Monthly (July 2026 data) shows Maryland residential prices up +13.7% YoY, Michigan +12.2%, Ohio +11.9%, Pennsylvania +11.3%, and Virginia +11.2% — all 2–3x the national +4.9% state-electricity-price-dispersion
. The commercial sector — where EIA counts most data centers — is the sharpest fingerprint: Ohio commercial +22.3%, Virginia +14.4%.
"Commercial prices rose 20.3% in Ohio and 20.0% in Virginia in the past year. Only DC (20.8%) rose more."
— startbusinessbystate.com
The mechanism explains the geography: PJM's socialized capacity charges are shared across all 13 states plus DC, which is why Maryland and Pennsylvania lead despite hosting fewer data centers than Virginia. And the cross-state correlation runs negative — data centers locate where power is cheap (Texas business prices +0.4% YoY), with the damage showing up years later through regional capacity markets state-electricity-price-dispersion. This resolves last cycle's open question directly: the "cost lands here" map is PJM (MD, PA, OH, VA, IL, MI), while ERCOT and the Southeast stay below trend thanks to large-load tariffs.
What to watch: Whether ERCOT and Georgia's below-trend readings hold as their own large-load pipelines (1.5 GW in Montana alone, below) mature.
Hyperscalers are now funding the supply side — $4.3B at a time
The "bring your own capacity" principle went from PJM market rule to voluntary deal template in eight days. Google's 20-year PPA with Constellation, announced October 6, supports $4.3 billion of upgrades at 11 nuclear units across six PJM sites, adding 890 MW by end-2032, plus a 15-year agreement covering 2,700 MW of existing output hyperscaler-funded-nuclear-uprates
. Amazon followed on September 29 with a 20-year PPA covering 690 MW of Calvert Cliffs, and DOE conditionally committed up to $4.2 billion to Vistra for uprates at Beaver Valley, Davis-Besse and Perry — 433 MW of new capacity, all of it under contract to Meta.
"a direct response to PJM's 'Bring Your Own Power' proposal, providing a practical model for how major power consumers can expand new supply alongside growing operational requirements, without state or federal mandates."
— POWER Magazine
This extends the Georgia Power/Google template from last cycle into the merchant market: merchant nuclear owners are monetizing existing fleets twice — long-dated PPAs on existing output plus hyperscaler-funded uprates on top. The open question is whether the 890 MW registers as PJM BYONC, which determines whether it actually shields other ratepayers or merely re-labels bilateral contracting.
What to watch: Constellation and Google confirming BYONC registration for the 890 MW — the cleanest test of whether the template has real ratepayer-protection mechanics.
FERC hit pause on the backstop — and the clocks are decoupled
The reliability architecture and the physical shortfall are now on different schedules. FERC accepted PJM's Reliability Backstop Procurement on September 29 but suspended it for five months to February 28, 2027, killing the September 30 auction start with "actual timing to be determined" pjm-deploys-reliability-backstop-procurement-framework
.
"A contract that can deliver as late as 2032 covers the back half of a shortfall that opens in 2028."
— America's Energy News
The math is the problem: the 6,831.3-MW shortfall opens June 2028, the backstop can't take effect until February 2027 at the earliest, and procured resources have an in-service deadline of June 1, 2032. FERC sent "who pays and what collateral" to a paper hearing in Docket EL26-108. Meanwhile the underlying scarcity persists — Talen cleared 10,180 MW in the 2028/29 auction for roughly $1,208 million, consistent with the $325/MW-day cap pjm-deploys-reliability-backstop-procurement-framework. This supersedes last cycle's December 2 watch item: that auction never happened.
What to watch: The EL26-108 paper hearing and the February 28, 2027 suspension expiry.
Virginia's merger fight makes cost allocation a campaign issue
The $67B NextEra–Dominion deal has become the most visible referendum on who pays for AI load. Lt. Gov. Ghazala Hashmi formally opposed it October 6 — the first statewide official to do so — after a five-city listening tour nextera-dominion-merger-virginia-ratepayers
.
"A promise and a press release is not protection on a power bill. When the commitments expire, it will not be the shareholders who pay, it will be everyday working Virginians."
— State Sen. Mike Jones, via Virginia Mercury
The SCC's local hearings ran overwhelmingly against the deal, and opponents are reframing the four-year bill-credit package as expiring window-dressing against a structural stranded-asset question — what residential customers owe if forecast data-center demand doesn't materialize. Regulators have until January 11 to rule; Dominion trades at $61.64, down 13.1% over three months nextera-dominion-merger-virginia-ratepayers.
What to watch: The January 2027 decision — and whether ring-fencing conditions become the template for every future utility M&A in data-center territory.
The market pays for contracts, not forecasts
The September de-rate reversed selectively: Constellation bounced +18.6% over three months and Vistra +15.3% over five days on their contract catalysts, while NRG sits −40.9% from its high with two straight EPS misses and its demand story still mostly forecast ai-power-trade-derating. Talen is +21% over a month on cleared capacity revenues and a $3B buyback — despite Q2 EPS of $0.16 against a $3.20 estimate. The bifurcation from last cycle sharpened into a rule: signed PPAs and auction revenues re-rate instantly; uncontracted load forecasts don't price at all.
What to watch: Q3 2026 earnings calls for load-growth guidance revisions and the contracted-versus-forecast split.
The load map keeps broadening — this time to Montana
NorthWestern Energy jumped 8.5% on October 8 on three signed data-center development agreements expected to push large-load demand toward about 1.5 GW by 2030 northwestern-montana-data-center-load
. Against a Montana system peak of roughly 3.5 GW, that's transformational for a small regulated utility — and it fits the pattern exactly: Montana residential power at 13.84¢ is well below the 18.01¢ national average, and data centers go where power is cheap. The same cost-allocation questions now head to the Montana PSC that Ohio, Georgia and Pennsylvania have already answered with large-load tariffs. Caveat: the 1.5 GW figure is sell-side narrative, not a regulatory filing.
What to watch: NorthWestern's Q3 earnings materials to verify the 1.5 GW against actual filings.
What surprised us
- The negative correlation. Host states don't see the price damage — the cross-state correlation between commercial-use growth and price increases runs negative (−0.62 for business prices). The bill arrives years later, through the regional capacity market, in someone else's state state-electricity-price-dispersion.
- FERC suspended the fix for the problem it exists to fix. A five-month suspension of the backstop, with procured capacity allowed to arrive as late as 2032 for a 2028 shortfall, means the window where PJM prices sit pinned at administrative caps while reliability risk accumulates just got longer pjm-deploys-reliability-backstop-procurement-framework.
- Talen's market cap is disconnecting from its income statement. +21% in a month, RSI 78, on a $3B buyback — while earning $0.16 against a $3.20 estimate. The market is pricing capacity revenues and optionality, not earnings.
- New Jersey is the PJM outlier. Residential prices fell 0.5% — but only because its large June 2025 capacity-cost reset already rolled into base rates. The pain was prepaid state-electricity-price-dispersion.
Open threads worth a vote
- [watch] Constellation/Google confirm whether the 890 MW uprates register as PJM BYONC
- [watch] FERC suspension of PJM backstop procurement expires Feb 28, 2027 (EL26-108 paper hearing)
- [watch] Utility Q3 2026 earnings calls — load-growth guidance and capex to serve AI demand
The AI Power Bill — Digest (Δ from last cycle)
Since last time
Promoted
- State price dispersion — last cycle's open question ("which states are rising fastest?") is now the lead section, answered with EIA state data.
Escalated
- Hyperscaler-funded nuclear — the Georgia/Google template is now a wave: $4.3B Google–Constellation, Amazon at Calvert Cliffs, a $4.2B DOE loan to Vistra.
- NextEra–Dominion merger — was a passing mention in the Virginia risk stack; now its own section with the first statewide official opposition.
- PJM backstop — was a December 2 watch item; FERC has since suspended the whole framework, and the clocks decoupled.
- The IPP de-rate bifurcation — sharpened into a rule, and partially reversed for contracted names.
Demoted
- Georgia Power/Google NU-1 deal — now cited only as the template the new deals extend.
- National baseline (STEO) — the national +4.9% figure survives inline; the 18.20¢/kWh STEO framing and TVA demand records are gone.
Disappeared
- The governors' EO wave — Virginia/Maryland executive orders, the ten-state intervention count, and the January 13, 2027 General Assembly watch item are all absent.
- Bowring's official attribution number — the $29.4B / 46% figure and his separate-auction proposal are gone (the socialization mechanism he explained survives as the price map's geography lesson).
- Vistra's equity-kicker PPA and the ERCOT co-location story — no mention; Vistra reappears only via the DOE loan.
- New Jersey's Fair Share Act curtailment hierarchy — NJ appears again, but only for its prepaid price reset.
- Michigan's OpenAI "Stargate" arbitrage — Montana is the new arbitrage example; Michigan itself is unmentioned.
Unchanged
- Nothing qualifies — even the market section's conclusion flipped from last cycle.
Promoted — The price map lands: PJM states are paying for everyone's data centers
The dispersion work item is closed, and the pattern maps onto the grid, not onto data-center siting. EIA's Electric Power Monthly (July 2026 data) shows Maryland residential prices up +13.7% YoY, Michigan +12.2%, Ohio +11.9%, Pennsylvania +11.3%, and Virginia +11.2% — all 2–3x the national +4.9% state-electricity-price-dispersion. The commercial sector — where EIA counts most data centers — is the sharpest fingerprint: Ohio commercial +22.3%, Virginia +14.4%.
"Commercial prices rose 20.3% in Ohio and 20.0% in Virginia in the past year. Only DC (20.8%) rose more."
— startbusinessbystate.com
The mechanism explains the geography: PJM's socialized capacity charges are shared across all 13 states plus DC, which is why Maryland and Pennsylvania lead despite hosting fewer data centers than Virginia. And the cross-state correlation runs negative — data centers locate where power is cheap (Texas business prices +0.4% YoY), with the damage showing up years later through regional capacity markets state-electricity-price-dispersion. The "cost lands here" map is PJM (MD, PA, OH, VA, IL, MI); ERCOT and the Southeast stay below trend thanks to large-load tariffs.
What to watch: Whether ERCOT and Georgia's below-trend readings hold as their own large-load pipelines (1.5 GW in Montana alone, below) mature.
Escalated — Hyperscalers are now funding the supply side — $4.3B at a time
Last cycle's single Georgia tariff deal became a deal wave in eight days. Google's 20-year PPA with Constellation, announced October 6, supports $4.3 billion of upgrades at 11 nuclear units across six PJM sites, adding 890 MW by end-2032, plus a 15-year agreement covering 2,700 MW of existing output hyperscaler-funded-nuclear-uprates. Amazon followed on September 29 with a 20-year PPA covering 690 MW of Calvert Cliffs, and DOE conditionally committed up to $4.2 billion to Vistra for uprates at Beaver Valley, Davis-Besse and Perry — 433 MW of new capacity, all of it under contract to Meta.
"a direct response to PJM's 'Bring Your Own Power' proposal, providing a practical model for how major power consumers can expand new supply alongside growing operational requirements, without state or federal mandates."
— POWER Magazine
This extends the Georgia Power/Google template into the merchant market: merchant nuclear owners are monetizing existing fleets twice — long-dated PPAs on existing output plus hyperscaler-funded uprates on top. The open question is whether the 890 MW registers as PJM BYONC, which determines whether it actually shields other ratepayers or merely re-labels bilateral contracting.
What to watch: Constellation and Google confirming BYONC registration for the 890 MW — the cleanest test of whether the template has real ratepayer-protection mechanics.
Escalated — FERC hit pause on the backstop — and the clocks are decoupled
Last cycle's December 2 backstop auction watch item is dead: that auction never happened. FERC accepted PJM's Reliability Backstop Procurement on September 29 but suspended it for five months to February 28, 2027, killing the September 30 auction start with "actual timing to be determined" pjm-deploys-reliability-backstop-procurement-framework.
"A contract that can deliver as late as 2032 covers the back half of a shortfall that opens in 2028."
— America's Energy News
The math is the problem: the 6,831.3-MW shortfall opens June 2028, the backstop can't take effect until February 2027 at the earliest, and procured resources have an in-service deadline of June 1, 2032. FERC sent "who pays and what collateral" to a paper hearing in Docket EL26-108. Meanwhile the underlying scarcity persists — Talen cleared 10,180 MW in the 2028/29 auction for roughly $1,208 million, consistent with the $325/MW-day cap pjm-deploys-reliability-backstop-procurement-framework.
What to watch: The EL26-108 paper hearing and the February 28, 2027 suspension expiry.
Escalated — Virginia's merger fight makes cost allocation a campaign issue
The NextEra–Dominion deal — previously one line in Virginia's risk stack — is now the most visible referendum on who pays for AI load. Lt. Gov. Ghazala Hashmi formally opposed it October 6 — the first statewide official to do so — after a five-city listening tour nextera-dominion-merger-virginia-ratepayers.
"A promise and a press release is not protection on a power bill. When the commitments expire, it will not be the shareholders who pay, it will be everyday working Virginians."
— State Sen. Mike Jones, via Virginia Mercury
The SCC's local hearings ran overwhelmingly against the deal, and opponents are reframing the four-year bill-credit package as expiring window-dressing against a structural stranded-asset question — what residential customers owe if forecast data-center demand doesn't materialize. Regulators have until January 11 to rule; Dominion trades at $61.64, down 13.1% over three months nextera-dominion-merger-virginia-ratepayers.
What to watch: The January 2027 decision — and whether ring-fencing conditions become the template for every future utility M&A in data-center territory.
Escalated (and partly reversed) — The market pays for contracts, not forecasts
Last cycle's de-rate has selectively reversed. Constellation bounced +18.6% over three months and Vistra +15.3% over five days on their contract catalysts, while NRG sits −40.9% from its high with two straight EPS misses and its demand story still mostly forecast ai-power-trade-derating. Talen is +21% over a month on cleared capacity revenues and a $3B buyback — despite Q2 EPS of $0.16 against a $3.20 estimate. The bifurcation sharpened into a rule: signed PPAs and auction revenues re-rate instantly; uncontracted load forecasts don't price at all. Note this directly contradicts last cycle's surprise that "contracted growth isn't pricing" — now it is.
What to watch: Q3 2026 earnings calls for load-growth guidance revisions and the contracted-versus-forecast split.
New — The load map keeps broadening — this time to Montana
NorthWestern Energy jumped 8.5% on October 8 on three signed data-center development agreements expected to push large-load demand toward about 1.5 GW by 2030 northwestern-montana-data-center-load. Against a Montana system peak of roughly 3.5 GW, that's transformational for a small regulated utility — and it fits the pattern exactly: Montana residential power at 13.84¢ is well below the 18.01¢ national average, and data centers go where power is cheap. The same cost-allocation questions now head to the Montana PSC that Ohio, Georgia and Pennsylvania have already answered with large-load tariffs. Caveat: the 1.5 GW figure is sell-side narrative, not a regulatory filing.
What to watch: NorthWestern's Q3 earnings materials to verify the 1.5 GW against actual filings.
What surprised us
- The negative correlation. [NEW] Host states don't see the price damage — the cross-state correlation between commercial-use growth and price increases runs negative (−0.62 for business prices). The bill arrives years later, through the regional capacity market, in someone else's state state-electricity-price-dispersion.
- FERC suspended the fix for the problem it exists to fix. [NEW] A five-month suspension of the backstop, with procured capacity allowed to arrive as late as 2032 for a 2028 shortfall, means the window where PJM prices sit pinned at administrative caps while reliability risk accumulates just got longer pjm-deploys-reliability-backstop-procurement-framework.
- Talen's market cap is disconnecting from its income statement. [UPDATED — was Vistra's beat-not-pricing last cycle; now Talen] +21% in a month, RSI 78, on a $3B buyback — while earning $0.16 against a $3.20 estimate. The market is pricing capacity revenues and optionality, not earnings.
- New Jersey is the PJM outlier. [NEW] Residential prices fell 0.5% — but only because its large June 2025 capacity-cost reset already rolled into base rates. The pain was prepaid state-electricity-price-dispersion.
Open threads
Closed or absorbed from last cycle:
- Which states' residential prices are rising fastest? — Closed: answered by the price-map section (PJM states 2–3x national; ERCOT/Southeast below trend).
- December 2 Reliability Backstop results — Superseded: the auction never happened; replaced by the FERC suspension thread (Feb 28, 2027 expiry, EL26-108).
- EIA STEO monthly release — Absorbed: tracking moved to EIA state-level Electric Power Monthly data, now in the body.
- Georgia PSC ruling on NU-1 (Dockets 44280/56002) — Dropped without closure: the deal survives only as the template for the new merchant-market wave; the docket watch is no longer tracked.
- Virginia 2027 General Assembly by-right ban — Dropped without closure: the governors'-EO storyline is entirely absent this cycle.
New threads this cycle:
- [watch] Constellation/Google confirm whether the 890 MW uprates register as PJM BYONC
- [watch] FERC suspension of PJM backstop procurement expires Feb 28, 2027 (EL26-108 paper hearing)
- [watch] Utility Q3 2026 earnings calls — load-growth guidance and capex to serve AI demand