AI · Energy · U.S. Grid · May 26, 2026 · 6:30 PM ET

Share

America Killed Its Power Plants Right Before the AI Race Started. Blue States Led the Way.

In 1950 the United States generated about 334 terawatt-hours of electricity — almost half from coal, a third from hydro, the rest from a thin slice of natural gas and oil. There were zero commercial nuclear reactors, zero wind turbines, and effectively zero solar panels. Seventy-five years later, in 2025, the U.S. set an all-time record at 4,430 TWh per the Energy Information Administration: a thirteen-fold increase, with natural gas at 43%, coal down to 15%, nuclear at 19%, hydro at 6%, wind at 11%, and utility-scale solar at nearly 7%. That is the headline arc of the American energy century.

Inside that arc is a quieter story. From 2013 through 2022, the U.S. retired roughly 9 gigawatts of operating nuclear capacity — baseload power, available 24 hours a day at a 92.3% fleet capacity factor (NRC, 2024) — while the natural-gas fleet expanded and renewables grew. About 6 GW of those retirements were political choices in deep-blue jurisdictions: San Onofre (CA, 2,150 MW, 2013), Vermont Yankee (VT, 605 MW, 2014), Oyster Creek (NJ, 615 MW, 2018), Pilgrim (MA, 685 MW, 2019), and Indian Point (NY, 2,069 MW, 2020-21) — the last under a 2017 settlement Gov. Andrew Cuomo (D-NY) signed with Entergy. Roughly 90% of Indian Point’s output was backfilled by new natural-gas plants — Cricket Valley (1,020 MW), CPV Valley (678 MW), and Bayonne II (120 MW). New York’s electric-sector greenhouse-gas emissions rose 12-15 million metric tons of CO₂e per year after the closure.

Then, starting around 2022, the AI race hit. Lawrence Berkeley National Lab’s December 2024 report — the canonical academic source — documented U.S. data-center electricity use rising from 76 TWh in 2018 to 176 TWh in 2023 (4.4% of the entire U.S. grid) on a path toward 325–580 TWh by 2028. OpenAI’s Stargate program alone is a $500 billion / 10 gigawatt commitment. xAI’s Colossus in Memphis is ramping toward 2 GW. Meta’s Hyperion in Louisiana is sized at 5 GW. The Microsoft-Constellation deal restarting Three Mile Islandis 835 MW — the first time a retired American reactor will be brought back from the dead. None of these projects existed in their current form when San Onofre was being mothballed.

And underneath all of it sits the U.S. interconnection queue. LBNL’s 2025 “Queued Up” edition: 10,300 projects, 1,400 GW of generation, 890 GW of storage, with a median wait of roughly five years from interconnection request to commercial operation — up from under two years in the early 2000s. The choke point is not chips, capital, or technology. It is permitting, transmission siting, and state environmental review — concentrated, by share of queued GW per acre, in CAISO (California), NYISO (New York), and ISO-NE (New England).

• 4,430 TWhU.S. record generation in 2025 per EIA — up ~13× from 1950's 334 TWh, but only 2.8% higher than 2024's 4,387 TWh.
• ~6 GWU.S. nuclear capacity retired under deep-blue state policy 2013-2021 — San Onofre, Vermont Yankee, Oyster Creek, Pilgrim, Indian Point 2 + 3 combined.
• 176 TWhU.S. data-center electricity use in 2023 (LBNL) — 4.4% of total U.S. generation. Headed to 325-580 TWh by 2028 (LBNL high case = 12% of US grid).
• 1,400 GWGeneration projects stuck in U.S. interconnection queue end-2024 (LBNL) — plus 890 GW of storage. Median wait ~5 years (vs <2 yrs in 2000-2007).
• 33.75¢California 2025 residential electricity rate per kWh — 87% above the U.S. average (18.05¢). CA rates have grown 2.8× since 2005 vs ~1.9× nationally.
• 52,061 MWCAISO all-time peak demand record (Sept 6, 2022) — set the same week Gov. Newsom (D-CA) asked Californians not to charge EVs during the 2035 ICE-ban rollout.
• $500BOpenAI Stargate program commitment / 10 GW capacity goal. First Abilene, TX building live September 2025.
• 835 MWThree Mile Island Unit 1 / Crane Clean Energy Center restart — Constellation / Microsoft 20-year PPA, ~$1B DOE conditional loan, restart now ahead of schedule for 2027. First-ever full U.S. nuclear restart.

The U.S. electricity mix is one of the cleanest illustrations in public data of an economy slowly diversifying off coal, slowly adopting nuclear, and then very recently piling renewables onto a grid that natural gas had already started to dominate.

Coal share peaked in 1997 at roughly 53% of U.S. generation, and absolute coal generation peaked in 2007 at 2,020 TWh (Ember + EIA Electric Power Annual). Natural gas first overtook coal in a single month in April 2015 (EIA), and in a full annual year in 2016 — and has not surrendered the lead since. Nuclear’s reactor count peaked at 112 in 1991; today the fleet is 92-94 reactors per the NRC, with the same generation share (~20%) maintained by capacity-factor improvement: the U.S. nuclear fleet hit 92.3% capacity factor in 2024, among the highest of any fleet globally. Wind crossed 1% of generation in 2008 and reached 11% in 2025. Utility-scale solar reached nearly 7% in 2025 from less than 1% in 2014.

Note what does notshow up on the stacked chart: the grid was always a story of state-level policy choices buried inside a national aggregate. The same year California shut down San Onofre (2013), Wisconsin shut down Kewaunee — one a political and regulatory closure, one a market closure. The national chart aggregates them.

The most-asked question in any AI-energy conversation is how an average nuclear reactor compares to an average solar farm. The answer turns on three numbers: nameplate megawatts, capacity factor (the % of theoretical maximum a plant actually delivers), and the resulting annual terawatt-hours. EIA Form-860/923 publishes the underlying data; the average plant of each type, in the U.S. today, looks like this:

The single most important number on the chart: an average U.S. nuclear reactor produces roughly 8.29 TWh per year — comparable to about 2.8 average coal plants, 1.75 average gas combined-cycle plants, 14 average wind farms, or 28 average utility solar farms. That is not a polemical claim. That is what dividing nameplate megawatts by capacity factor and multiplying by 8,760 hours actually does. Wind and solar are valuable. They are also fundamentally different in their per-plant kWh footprint — which is why the AI data-center buildout is now driving an unprecedented hyperscaler interest in restarting retired reactors and signing 20-year nuclear PPAs.

Data centers as a category are not new — the U.S. has run cloud-scale facilities for two decades. What is new is the rate of growth. Lawrence Berkeley National Lab’s December 2024 report — the standard academic reference cited by the EIA, the IEA, and most utility filings — documented a doubling of U.S. data-center electricity demand between 2018 and 2023, and a projection that demand will double again by 2028.

For perspective: 580 TWh — LBNL’s high-end 2028 projection — would put U.S. data-center consumption at roughly 12% of the entire U.S. grid. That is more than the U.S. residential electricity demand for the states of Texas and California combined. It is also 14% more electricity than the entire United Kingdom generates in a year.

What is being built, where, and on whose grid:

Eight of the ten largest announced clusters above are sited in Texas, Tennessee, Louisiana, Pennsylvania, and Wisconsin — states with either Republican statewide control of the relevant permitting bodies (TX, TN, LA), purple state governance with permitting reforms underway (PA, WI), or in the case of New Mexico, the deepest queue room in WECC. Notice what is missing: California, New York, and Massachusetts. The political geography of the AI compute buildout is the inverse of where prestige headquarters and university talent concentrate. Power, not prestige, picks the sites.

On paper, the U.S. is adding more new generation every year than the AI industry is announcing. Below: cumulative gigawatts of announced AI data-center capacity against cumulative gigawatts of new U.S. generation commissioned.

That headline view is also misleading in two specific ways. First, “announced” AI capacity is corporate disclosure — some of it will not be built on schedule. Second, “commissioned” new generation includes a heavy tilt toward intermittent wind and solar that cannot match an AI data center’s 24/7 load profile. A 200 MW wind farm running at 35% capacity factor delivers, on average, what 70 MW of nuclear delivers continuously. Stargate Abilene cannot run on what it can’t schedule.

The real bottleneck sits in the interconnection queue. LBNL’s “Queued Up: 2025 Edition” (Dec 2025) is the load-bearing document on this:

If the AI race is fundamentally a grid race, California is the case study in how an American state legislatively dismantled the kind of generation AI requires — while simultaneously installing the most aggressive demand-side electrification mandates in the country.

The sequence is straightforward in the public record. In 2013, after a steam-generator failure and a CPUC cost-recovery dispute, Southern California Edison retired San Onofre Units 2 & 3 (2,150 MW) — about 8% of California’s in-state generation capacity, gone in a single decision. The replacement was overwhelmingly in-basin natural gas. In 2016, PG&E announced — under pressure from a coalition of environmental groups — that Diablo Canyon (the state’s last operating reactor, 2,256 MW) would close in 2024-2025. In 2018, Gov. Jerry Brown (D-CA) signed SB 100, mandating 100% zero-carbon electricity by 2045. In September 2020, Gov. Gavin Newsom (D-CA) signed Executive Order N-79-20 banning new internal-combustion-engine vehicle sales by 2035. The same August, CAISO rolled blackouts for two consecutive evenings — the first state-ordered rolling blackouts in California since 2001.

The simplest measurable consequence: California’s residential electricity rate has compounded relative to the U.S. average for two decades. EIA Form-826:

And on the reliability side, the public record is unusually clean: CAISO has rolled blackouts (Aug 2020), called all-time peak demand records (Sept 2022, with the “don’t charge your EV” Flex Alert), and conducted PG&E Public Safety Power Shutoffs every fire season since 2019. ISO-NE in deep-blue New England burned more oil in 15 days during the January 2018 cold snap than in the prior three years combined — because the natural-gas pipelines that would have backstopped the grid had been blocked under former Gov. Andrew Cuomo (D-NY) and his New England counterparts.

“The grid did not fail because the technology is broken. The grid fails because we chose, repeatedly and publicly, to retire dispatchable generation before we had built the firm replacement. The replacement requires permits we will not issue, transmission we will not site, and reactors we will not build.”

Editorial · paraphrased from public testimony submitted to CAISO 2020 post-event review

The Northeast story is California’s, with three different statehouses signing the orders.

New York — Indian Point. On January 9, 2017, Gov. Andrew Cuomo (D-NY) announced a settlement with Entergy to close Indian Point Units 2 and 3 by April 2020 and April 2021 respectively. The agreement was framed as an environmental win. NYISO’s post-closure reliability filings show what actually happened: three new natural-gas plants — Cricket Valley (1,020 MW), CPV Valley (678 MW), and Bayonne Energy Center II (120 MW) — backfilled approximately 90% of Indian Point’s 2,069 MW output with fossil gas, not zero-carbon. Downstate New York electricity went from 77% gas + oil in 2020 to 89% in 2021 (NucNet citing NYISO). New York State’s electric-sector greenhouse-gas emissions rose 12-15 million metric tons of CO₂e per year — a ~33% increase. The Climate Leadership and Community Protection Act (CLCPA) Cuomo signed July 18, 2019 committed New York to 70% renewables by 2030 anyway. NYISO’s 2024 Power Trends flagged a downstate 17 MW summer-2033 reliability violation growing to 97 MW in 2034 absent intervention. Gov. Kathy Hochul (D-NY)inherited the problem — and, per her own retrospective on the Cuomo closure: “There was no Plan B.”

“There was no Plan B.”

Gov. Kathy Hochul (D-NY) · on the Cuomo administration's Indian Point closure · 2025

Hochul’s reversal is now on record. On June 23, 2025, she directed the New York Power Authority to develop a zero-emission advanced nuclear plant of at least 1 GW upstate. In her January 14, 2026 State of the State address, she expanded the target to 5 GW of new nuclear— a fleet that would, combined with existing reactors, return New York to roughly 8.4 GW of nuclear capacity. The same speech instructed data-center developers to “pay their fair share for the power they use and ultimately generate their own power independently” — the closest thing in either party to an open admission that the post-2017 New York grid policy was incompatible with the AI buildout the state’s own legislature now wants.

Vermont — Vermont Yankee. Gov. Peter Shumlin (D-VT)ran on closing Vermont Yankee in 2010 and won. The 605 MW reactor — which had supplied roughly 70% of Vermont’s electricity — closed in December 2014. Vermont now imports a substantial majority of its electricity from neighboring states and Hydro-Québec, paying the import premium and exporting the carbon footprint to whoever runs the grid upstream.

Massachusetts — Pilgrim.Closed May 2019 (685 MW). Gov. Charlie Baker was a Republican, but the Massachusetts legislature’s Democratic supermajority and the state’s coastal-permitting framework drove the policy environment that made Pilgrim’s economics impossible. ISO-NE has lost roughly 2,200 MW of nuclear baseload across the region since 2014 (Vermont Yankee + Pilgrim + decommissioned Connecticut/Maine Yankees in the 1990s) and now operates as the highest-priced wholesale-electricity market in the lower 48.

The clearest single signal that the U.S. industry has decided the AI race is a power-plant problem: hyperscalers have started signing 20-year power-purchase agreements with retired or operating reactors, and the federal government has started underwriting the restarts.

“Without copious amounts of energy in a certain quality and a certain type of energy, there is no AI. The only way we're going to bring AI in to transform so many of our problems is massively more energy.”

Chris Wright · U.S. Energy Secretary · ARPA-E Summit · April 2026

“The abundance of AI will be limited by the abundance of energy.”

Sam Altman · OpenAI CEO · Senate testimony · 2025

“The grid is the new gating factor for everything in AI. The question is whether we can build power at the rate we can build chips. Today the answer is no.”

Jensen Huang · Nvidia CEO · Columbia Center on Global Energy Policy · 2025

“Eventually we will produce more chips than we can turn on. The limiting factor is electrical power.”

Elon Musk · CEO Tesla / xAI · World Economic Forum · January 2026