Shale gas, cheap renewables, fifteen years of flat demand, and then AI. How the U.S. power industry was rebuilt after 2005, and why electricity is suddenly scarce again.
Nakul Nandhu23 October 202615 min read
On September 20, 2024, Constellation Energy said it would bring Unit 1 at Three Mile Island in Pennsylvania back to life, rename it the Crane Clean Energy Center, and sell its power to Microsoft under a 20-year contract (Nuclear News Network, 2026). Unit 1 sits next to the reactor that partially melted down in 1979, the accident that, as Part 1 described, helped freeze new nuclear construction in America for decades.
The most famous name in American nuclear failure had become a symbol of something new. For the first time in a generation, electricity in the U.S. is scarce, and some of the largest companies in the world are lining up to pay for it. This article explains how the industry got here from 2005, and where the money is going now.
The short version
Shale gas made electricity cheap after 2008. Gas replaced coal as America's biggest power source, and generators that sold into markets saw their prices squeezed.
Wind and solar went from a rounding error to about a sixth of U.S. power, built by developers whose business depends on the cost of borrowing money.
Demand barely grew for fifteen years, so regulated utilities grew the only way the formula allows: by spending on their grids.
Then data centers arrived. Demand is rising again, but power plants, transformers and grid connections take years to add. Whoever already owns scarce capacity now has pricing power.
00Where Part 1 left off
By 2005 the U.S. power industry had split in two. In about half the country, mostly the Southeast and much of the West, one regulated utility still owned everything from the power plant to your meter and earned a set return on what it built. In the other half, generation had been broken off into independent power producers that sell into wholesale markets run by non-profit grid operators, while regulated "wires companies" kept the poles and lines. Both models were about to be tested by the same three forces.
01The shale shock
2008 to 2016
Drillers learned to combine horizontal wells with hydraulic fracturing and unlocked enormous amounts of natural gas trapped in shale rock. The price collapsed. Gas at Henry Hub, the main U.S. benchmark, peaked at $13.32 per million British thermal units in July 2008 (Congressional Research Service, 2009). By April 2012 it had fallen below $2, a level last seen in 2001 (U.S. Energy Information Administration, 2014).
Power plants noticed. In 2005, coal produced about half of America's electricity and gas less than a fifth. In 2016 gas overtook coal for the first time, and by 2025 gas produced about 41% of U.S. power while coal had fallen to about 17% (U.S. Energy Information Administration, 2026a).
Natural gasCoalNuclearWind and solar
Figure 1. How America makes electricity, 2005 to 2025. Share of total U.S. net generation by source. Gas passed coal in 2016. In 2025, wind and solar together produced slightly more electricity than coal.
Data: U.S. Energy Information Administration (2026), Monthly Energy Review, Table 7.2a, all sectors. Shares are the author's calculation.
To see why cheap gas mattered so much for money, picture how a competitive power market sets its price. Every hour, each plant offers its power at roughly what it costs to run. The grid operator accepts the cheapest offers first and keeps going up the list until supply meets demand. The last plant it needs sets the price, and every plant that was accepted gets paid that price.
The price-setting plant was usually burning gas.
In most U.S. markets, for most hours, the last plant needed is a gas plant. So when gas got cheap, the market price fell for everyone, including coal and nuclear plants that never burned any gas. Their costs stayed the same while their revenue dropped.
This is exactly the trap that caught TXU, the Texas company whose 2007 buyout became Vistra. The buyers had bet that gas, and therefore power, would stay expensive. Shale proved them wrong, and the company went bankrupt in 2014 (NPR, 2014). Regulated utilities were largely protected, because their revenue comes from the rate-base formula, not from market prices.
02The renewables wave
2009 to 2025
At the same time, wind and solar became cheap. Lazard, which publishes a widely followed annual comparison of power costs, estimates that the cost of electricity from a new utility-scale solar farm fell about 83% from 2009, when it was above $350 per megawatt-hour (Kennedy, 2024). Its 2025 estimate puts unsubsidized utility-scale solar at $38 to $78 per megawatt-hour (PV Tech, 2025). Output followed. Wind went from 18 terawatt-hours in 2005 to 464 in 2025, and solar from about 1 to 296 (U.S. Energy Information Administration, 2026a).
Renewables brought a new kind of business. A developer finds a site, secures a grid connection, signs a long contract to sell the power (often 15 to 25 years), builds the project, then holds it or sells it to investors. The key is that a solar farm has no fuel bill.
For a solar farm, the interest rate is the fuel price.
Almost all of a solar farm's cost is paid upfront, before it sells a single kilowatt-hour. So the price it needs over its life depends mostly on what it costs to borrow that money. Cheap money makes cheap solar. When interest rates rose sharply in 2022 and 2023, the same panels suddenly needed a higher price to make sense.
Federal tax credits added fuel of their own. The Inflation Reduction Act of 2022 extended and expanded them. That support is now being withdrawn: the One Big Beautiful Bill Act, signed on July 4, 2025, ends the main wind and solar credits for projects placed in service after 2027, unless construction began by July 4, 2026. Credits for storage and nuclear largely survived (Sidley Austin, 2025).
One company came to symbolize the era. NextEra Energy, which owns Florida's largest regulated utility and also built one of the world's biggest wind and solar businesses, briefly became the most valuable energy company in America. On October 7, 2020, its market value of $145.5 billion passed ExxonMobil's $141.6 billion (Crowley et al., 2020).
03Fifteen years of flat demand
2007 to 2021
For most of the 20th century, electricity demand grew every year. Then it stopped. U.S. retail electricity sales were 3,765 terawatt-hours in 2007 and 3,806 in 2021, a gain of about 1% in fourteen years (U.S. Energy Information Administration, 2026b). The EIA describes U.S. consumption as essentially flat for nearly two decades, as more efficient lighting and appliances and a shift away from heavy industry offset growth elsewhere (U.S. Energy Information Administration, 2025).
Figure 2. Demand went flat, then came back. Total U.S. retail electricity sales, in terawatt-hours. After 2021, sales rose about 7% in four years, led by the commercial sector, where data centers are counted.
Data: U.S. Energy Information Administration (2026), Monthly Energy Review, Table 7.6.
Flat demand created a puzzle for both business models.
Regulated utilities could not grow by selling more power, so they grew the way the formula from Part 1 rewards: by investing. Replacing aging equipment, hardening lines against storms and wildfires, and building renewables all add to the rate base, and the allowed return is earned on the rate base. Investment kept rising even while sales did not. Members of the Edison Electric Institute, the trade group for investor-owned utilities, planned to invest nearly $208 billion in 2025 alone (Edison Electric Institute, 2025). The catch is that customers pay for that investment through their bills, which is why rising rates have become a political issue.
Generators in competitive markets faced a different problem. With flat demand and cheap gas, energy prices were low, and plants that ran only at peak times struggled to cover their fixed costs. Several regions answered with capacity markets, which pay plants for being available when needed, not just for the energy they sell. Texas chose not to, keeping an energy-only market (POWER, n.d.). Its weakness showed in February 2021, when Winter Storm Uri knocked out a large share of the state's power plants and Texas later put the death toll at 246 (Spectrum News, 2022).
A new kind of owner also arrived. Infrastructure investors liked the steady, long-term cash flows of power plants, pipelines and even whole utilities. The sector grew big enough that BlackRock, the world's largest asset manager, bought Global Infrastructure Partners, completing the deal on October 1, 2024 and creating an infrastructure business with about $170 billion of assets under management (BlackRock, 2024).
04Demand comes back
2022 to today
Then the line in Figure 2 turned up. U.S. retail sales rose from 3,806 terawatt-hours in 2021 to 4,058 in 2025, a record, and commercial sales alone grew almost 10% between 2019 and 2025 (U.S. Energy Information Administration, 2026b). The EIA expects commercial demand, which includes data centers, to keep growing faster than any other sector (U.S. Energy Information Administration, 2025).
Data centers are the biggest single reason. A report by Lawrence Berkeley National Laboratory for the Department of Energy found that data centers used 58 terawatt-hours in 2014 and 176 in 2023, about 4.4% of U.S. electricity. It projects 325 to 580 terawatt-hours by 2028, or 6.7% to 12% of the national total (U.S. Department of Energy, 2024).
Figure 3. Data center electricity use. Actual use in 2014 and 2023, and the projected range for 2028, in terawatt-hours. The width of the 2028 range is itself the story: nobody knows exactly how fast this grows.
Data: Lawrence Berkeley National Laboratory, as reported by the U.S. Department of Energy (2024).
The technology companies building these data centers have become a new kind of electricity customer. Instead of simply buying from their local utility, they are signing long contracts directly with power plants, and they especially want nuclear, because it runs around the clock and produces no carbon:
Microsoft and Constellation (September 2024): a 20-year contract to restart the 835-megawatt Crane unit, targeted for 2027, helped by a $1 billion Department of Energy loan closed in November 2025 (Nuclear News Network, 2026).
Amazon and Talen Energy: after FERC blocked a plan in November 2024 to supply more power directly to a data center next to the Susquehanna nuclear plant, the companies restructured in June 2025 into a 1,920-megawatt deal through 2042, delivered through the grid (Vigliarolo, 2025).
Google and Kairos Power (October 2024): up to 500 megawatts from small modular reactors, with the first targeted for 2030 (NucNet, 2024).
Meta and Constellation (June 2025): 20 years of the roughly 1.1-gigawatt output of the Clinton nuclear plant in Illinois, starting in June 2027 (Stevens, 2025).
Prices responded. PJM, the grid operator for 13 states and Washington, D.C., runs an annual auction to buy capacity three years ahead. The clearing price for the 2024/25 delivery year was $28.92 per megawatt-day. For 2025/26 it was $269.92. The next three auctions all cleared at or near the maximum allowed, and the most recent, for 2028/29, cleared at the $325 cap while still falling 6,831 megawatts short of PJM's reliability target (PJM Interconnection, 2026). For the 2027/28 auction, PJM's demand forecast rose by 5,250 megawatts, almost entirely because of data centers (Howland, 2025a).
Figure 4. The price of being available. PJM capacity auction clearing prices by delivery year, in dollars per megawatt-day. In four years the price rose about elevenfold.
Data: PJM Interconnection (2026), 2028/2029 Base Residual Auction Report, Table 2.
Scarcity also brought consolidation. In January 2026 Constellation completed its purchase of Calpine, creating a company with about 55 gigawatts of capacity in a deal valued at roughly $26.6 billion including debt (Industrial Info Resources, 2026). It is the same Calpine that, as Part 1 described, went bankrupt in 2005 after the last merchant power boom.
05The bottlenecks
If demand is rising and prices are high, why not just build more? Because almost every piece of the system takes years to add.
Getting connected. Before a new power plant can sell electricity, the grid operator has to study whether the grid can handle it, and the line is long. About 2,061 gigawatts of generation and storage were waiting in U.S. interconnection queues at the end of 2025 (Rand et al., 2026), more than the roughly 1,250 gigawatts of capacity on the entire U.S. grid today (Edison Electric Institute, 2025). The median project built in 2025 took 61 months from its request to commercial operation, compared with 36 months in 2015. Most requests never get built: of projects that applied between 2000 and 2020, only about 19% had reached operation by the end of 2025 (Rand et al., 2026).
Figure 5. The waiting line. Total generation and storage capacity in U.S. interconnection queues at the end of each year, in gigawatts. The queue has shrunk since its 2023 peak, but it is still larger than the whole existing U.S. grid.
Data: Lawrence Berkeley National Laboratory, Queued Up reports (2022, 2023, 2024 and 2026 editions). Values for 2021 to 2023 are rounded as published.
Getting equipment. Power transformers, the machines from Part 1 that step voltage up and down, now take about 128 weeks to deliver on average, and generator step-up transformers about 144 weeks, according to a 2025 Wood Mackenzie survey. Power transformer prices are up about 77% since 2019 (Garrett, 2026). Gas turbines are just as tight. GE Vernova's backlog of gas turbine orders and reservations reached 116 gigawatts in mid-2026, and it was already taking reservations for delivery in 2031 (Martucci, 2026).
Getting the rules right. Regulators are still working out how data centers should connect, and who pays. In December 2025, FERC ordered PJM to create new transmission services for data centers built right next to power plants (Howland, 2025b). In June 2026, after a request from the Department of Energy, FERC ordered all six major grid operators it oversees to either justify their rules for large new loads or change them (Pomerleau, 2026).
When supply cannot respond quickly, the rewards go to whoever already holds the scarce thing: an existing power plant, a factory slot for a transformer or turbine, or a place near the front of the queue. That is the value chain at the end of the rebuilding era.
Power flowsLong-term contractsBottleneck
Diagram 3. The chain rebuilt, 2026. Compared with Diagram 2 in Part 1: coal is shrinking, wind, solar and gas developers sit alongside utilities and independent producers, financial investors own a growing share of assets, hyperscalers buy power directly from plants, and the slowest links are equipment and grid connections.
Based on the sources cited in Sections 1 to 5.
06Where the money is now
Put the pieces together and each link of the chain is in a different position than it was five years ago.
Link
What changed
Evidence
Equipment makers
Sold out years ahead
Turbine reservations for 2031; transformers take about 2.5 years to deliver.
Generators in tight markets
Paid more to exist
PJM capacity prices up about elevenfold since 2024/25.
Nuclear owners
New long-term buyers
20-year contracts with Microsoft and Meta; a reactor restart.
Regulated utilities
Growing rate base
Record investment, paid for through customer bills.
Wind and solar developers
Cheap power, fading support
Among the cheapest sources of new power, but key federal credits end for projects after 2027.
New projects of any kind
Stuck in line
A median 61 months from grid request to operation.
Table 1. Who gained and who is constrained. Compiled from the sources in Sections 1 to 5.
The series ends where it began, with a forecast. In the 1990s and early 2000s, companies built power plants on optimistic demand forecasts, and when the demand did not arrive, many of them went bankrupt. Today the industry is again building, buying and borrowing on a forecast, this time for data centers, and the official range for 2028 runs from 325 to 580 terawatt-hours. Which end of that range proves right will decide who gets paid in the next chapter.
I used AI (Claude) as my research and drafting partner. The questions, the structure and the cuts are mine, and every figure is checked against the sources listed below. If something is wrong, that is on me, and I would like to know.
This article is for educational purposes only and is not investment advice. Figures are as reported by the cited sources on the dates shown. All views are my own.
Rand, J., Cheyette, A., Talley, C., Zhang, S., Gorman, W., Wiser, R., Seel, J., Jeong, S., & Kahrl, F. (2026). Queued up: 2026 edition. Characteristics of power plants seeking transmission interconnection as of the end of 2025. Lawrence Berkeley National Laboratory. https://eta-publications.lbl.gov/sites/default/files/2026-06/queued_up_2026_edition.pdf
U.S. Energy Information Administration. (2025, May 13). After more than a decade of little change, U.S. electricity consumption is rising again. Today in Energy. https://www.eia.gov/todayinenergy/detail.php?id=65264