The Energy Wall

On a humid morning in late June 2024, the Greater Memphis Chamber called a press conference and announced, with the cadence of a city luring an NFL franchise, that Elon Musk’s xAI had chosen Memphis as the site of the world’s largest AI training cluster. The chamber’s chief executive, Ted Townsend, told reporters that the project would bring billions in capital investment and put Memphis on the map of the artificial intelligence age. He did not mention how the building, an old Electrolux appliance factory in the South Memphis industrial belt, would be powered. Nobody asked. The question would come later.

Within weeks, contractors were trucking in row after row of mobile gas turbines, the kind oil companies use at remote drilling sites, and lining them up along the property’s perimeter. By August, residents in Boxtown, the historically Black neighborhood across the rail yard, were reporting a steady metallic hum and a smell they had not had to live with for a generation. The Shelby County Health Department, asked why xAI was running roughly thirty-five generators with no air permits, replied that the equipment was technically portable, and that portable generators in one location for fewer than 365 days fell outside its enforcement remit. The Southern Environmental Law Center filed a notice of intent to sue. The NAACP joined. Justin Pearson, a state legislator who had grown up nearby and made his name fighting an oil pipeline in those same blocks, told a rally that clean air was a human right. By the time the county issued a partial air permit for fifteen turbines in the summer of 2025, xAI had quietly doubled the count and moved a second array across the state line into Mississippi.

Colossus, as Musk had nicknamed it, came online in early September 2024 with a hundred thousand H100 GPUs racked in a building that, six months earlier, had been empty. By the spring of 2025 it held two hundred thousand accelerators of various generations, all of them hungry. The first phase pulled roughly 150 megawatts. Memphis Light, Gas and Water completed a dedicated substation in early 2025, and the Tennessee Valley Authority agreed to deliver up to 150 megawatts of grid power, which xAI then supplemented with onsite gas and a wall of Tesla Megapacks. The math was inescapable: a single training cluster, dropped into a single zip code, was now consuming as much electricity as a small city, and it had gotten there in a year by running natural gas turbines that the local health authority lacked the staffing or the will to regulate.

Memphis was the loud version of a story playing out, more quietly and with better lawyers, in dozens of American counties. The chip war had been a war over wafers, lithography tools, and export licenses. By 2024, it had a new front, and the front was the wall socket. The constraint had moved.

Training one of the GPT-4-class systems available in 2024 took roughly twenty to twenty-five megawatts of continuous power for about three months, on Epoch AI’s estimates. That alone was the steady-state load of a small American town. Training was the smaller half of the bill. Inference scaled with usage, and usage was rising on a slope nobody in the industry quite trusted to be linear. By the time Sam Altman was telling reporters in late 2024 that ChatGPT had three hundred million weekly users, the cost of serving them was best measured in gigawatts. Frontier training compute had grown roughly 4 to 5 times per year for a decade. Power demand for individual training runs was doubling roughly every nine months. Epoch projected that, on trend, the largest single training run in 2030 would draw on the order of a gigawatt all by itself, the size of a single nuclear unit running flat out for a quarter.

The hyperscalers, who had spent a decade telling Wall Street their capital intensity would moderate, abandoned the line. Microsoft, Alphabet, Amazon, and Meta together had spent about $162 billion on capex in 2022. By 2025 that figure was on track for roughly $440 billion, three quarters of it pointed at AI infrastructure. Goldman Sachs analysts started circulating client notes that put aggregate hyperscaler capex above six hundred billion dollars by 2026. The grid had been built for a country whose electricity demand grew at half a percent a year, when it grew at all.

Power planners began to understand the problem in the language of their own auctions. PJM Interconnection, the grid operator running the wholesale market across thirteen states from Illinois to North Carolina, holds a yearly capacity auction to ensure enough generation to meet peak demand three years out. For the better part of a decade, the auction had cleared at quiet prices, between $30 and $50 per megawatt-day. In the auction held in mid-2024 for the 2025-2026 delivery year, the price cleared at $269.92 per megawatt-day in most zones, a roughly tenfold jump that PJM’s own market monitor traced overwhelmingly to data center load growth in northern Virginia and central Ohio. The Citizens Utility Board in Illinois calculated that the spike would translate into roughly $9 billion in higher electric bills, paid not by the data centers but by the grandmothers in apartment buildings up and down the corridor. A year later the auction cleared near the cap PJM had hastily put in place after the first shock. Capacity, in the most-modeled grid in the world, had become scarce in a way it had not been since the 1970s.

In Texas, on the other grid, the planners watched the same wave coming. ERCOT, which oversees most of Texas, normally projects load growth in single-digit percents. In late 2024 it was tracking sixty-three gigawatts of large-load interconnection requests, mostly from data centers and hyperscaler co-locations. A year later that figure had risen to two hundred and twenty-six gigawatts, more than three times the size of the entire ERCOT peak that summer. ERCOT publicly acknowledged that not all of it was real, that developers were filing speculative applications in multiple substations to lock in queue position, and that historical experience suggested only about half the requested megawatts would actually materialize. Even the deflated number was a tripling of Texas’s grid in eight years. Northern Virginia, the densest concentration of data centers on Earth, was already drawing more than four gigawatts continuously by 2024 and was on track for thirteen gigawatts of contracted demand by 2030 in Dominion Energy’s planning forecasts. Loudoun County, the suburban exurb that called itself “Data Center Alley,” was now the largest single load center in any American utility’s service territory.

The first reflex of the hyperscaler chief financial officers, when they understood the wall they were running into, was to reach for nuclear. Nuclear plants ran at capacity factors above 90 percent, produced no carbon, and, crucially, came in firm 800-megawatt lumps that could be plumbed straight into a campus. The problem was that very few of them had been built in the past thirty years, and the existing ones were already selling everything they made.

On Friday, September 20, 2024, Constellation Energy, the largest nuclear operator in the United States, announced it would restart a reactor that had been shut down five years earlier. The reactor sat on Three Mile Island in the Susquehanna River, ten miles south of Harrisburg. Unit 2 had famously suffered a partial meltdown in March 1979 and never run again. Unit 1, its undamaged twin, had operated cleanly until 2019, when Exelon retired it on cost grounds, citing low natural gas prices and a bad power market. Constellation now intended to bring Unit 1 back. Microsoft would buy every electron it produced under a twenty-year power purchase agreement, the largest in Microsoft’s history. Constellation would rebuild the turbine generators, refurbish the cooling tower, requalify the operators, and rename the plant the Crane Clean Energy Center after the late Exelon chief executive Chris Crane. The plan called for the reactor to come back in 2028. Constellation’s chief executive Joe Dominguez framed the deal in the language of someone who had spent a decade defending nuclear’s relevance. “In this rebirth,” he said, “we see the most powerful sign that America will turn to the enduring promise of nuclear energy.”

The market read the announcement instantly. Constellation’s stock jumped roughly nine percent at the open, adding more than five billion dollars to its market value before lunch. Analysts at BMO and Morgan Stanley estimated that Microsoft was paying somewhere between $98 and $115 per megawatt-hour for the contracted power, a premium of perhaps fifty percent over the prevailing wholesale rate in the PJM market. The premium told the story. Microsoft was not buying clean energy attributes. Microsoft was buying physical electrons attached to a real piece of generation that nobody else could outbid them for.

The same logic was running through every other hyperscaler procurement office. Amazon Web Services, six months earlier, had bought the Cumulus data center campus next to Talen Energy’s Susquehanna nuclear station in Pennsylvania for $650 million, with a co-location arrangement that called for AWS to draw up to 960 megawatts directly from the plant, behind the meter, bypassing the regional transmission system entirely. The structure was elegant from AWS’s perspective and infuriating to everyone else on the grid. PJM filed an amended interconnection agreement that would have raised the co-located load from 300 to 480 megawatts. Exelon and American Electric Power objected, arguing that the AWS load was effectively using PJM’s transmission network without paying for it, and would impose roughly $140 million a year in cost shifts onto ordinary ratepayers. On November 1, 2024, the Federal Energy Regulatory Commission voted 2 to 1 to reject the amendment. Talen’s stock dropped sharply. Constellation’s dropped with it. The signal was unmistakable: the federal regulator was not going to let hyperscalers quietly siphon nuclear plants off the grid without first answering who paid for the wires.

A month later, in October 2024, Google signed a master agreement with Kairos Power for up to 500 megawatts of small modular reactor capacity, with the first 50-megawatt unit slated for service by 2030. Days later, Amazon announced a $500 million investment in X-energy, the Maryland startup developing a high-temperature gas-cooled SMR, with a stated goal of bringing more than five gigawatts of new nuclear capacity online by 2039. Equinix paid Sam Altman’s reactor startup Oklo $25 million up front against a future order of as much as 500 megawatts. Meta, late to nuclear and unwilling to be left behind, would by early 2026 sign a sweeping set of agreements with Constellation, Vistra, TerraPower, and Oklo for as much as 6.6 gigawatts. None of these reactors existed yet. None would deliver a watt before the end of the decade. The deals were calls on a future supply of firm power, made by companies that had concluded they could not afford to wait and see whether anyone else would build it.

Sam Altman, who had personally invested $375 million in the fusion startup Helion, watched Microsoft sign a 2023 power purchase agreement with that company for a fifty-megawatt fusion plant by 2028. Fusion in 2028, on a commercial schedule, was a wager that no scientist outside Helion’s payroll would have made. But the calculation in Redmond was not really about whether fusion would work. It was about optionality, about being on the call list when something firm and clean came online. The hyperscalers had become venture investors in the electrical sector because the electrical sector had not, on its own, planned for them.

Nuclear, slow as it was, was still the cleaner side of the new market. Behind it, in plain sight, the gas turbine industry was having the best year of its life. GE Vernova’s order book stretched out past 2028, and so did Siemens Energy’s. OpenAI and Oracle, building the Abilene flagship of the Stargate program in West Texas, placed what SemiAnalysis described as the largest single onsite gas generation order in history, a 2.3-gigawatt plant alongside the data center campus. Meta announced in late 2024 a 2-gigawatt data center in Richland Parish, Louisiana, called Hyperion. By the spring of 2026 the project had grown into ten new gas-fired plants commissioned through Entergy Louisiana, totaling more than seven gigawatts and representing roughly a thirty percent expansion of the entire Louisiana grid. The optics were difficult for companies that had spent a decade assuring shareholders they would be net zero by 2030. The math was simple. Gas could be permitted, financed, and synchronized in two to three years. Nuclear could not.

The shift in framing inside the industry was sharper than the public statements admitted. At a Stripe conference in early 2025, Microsoft’s CTO Kevin Scott told the audience that the binding constraint on AI was no longer GPUs but power. Jensen Huang, asked the same question on stage at his company’s developer conference, gave a number that was hard to verify but easy to remember: a gigawatt of AI infrastructure could throw off ten to twelve billion dollars of revenue a year. At those returns, hyperscalers paid almost any price for firm capacity. SemiAnalysis tracked the global AI data center buildout through satellite imagery and concluded AI-specific load had crossed ten gigawatts by early 2025 and was on a track the consultancy itself called “absurd.” The International Energy Agency, in an April 2025 special report, projected global data center electricity demand would double from 415 terawatt-hours in 2024 to 945 terawatt-hours by 2030, with the United States and China accounting for roughly eighty percent of the growth. By the end of the decade, the IEA wrote, American data centers would consume more electricity than aluminum, steel, cement, and chemicals combined. Goldman Sachs put the increase at 165 percent by 2030, requiring roughly $720 billion of grid investment globally.

The numbers were dizzying enough that the people writing them down began second-guessing them in print. Some of the load requests in ERCOT’s queue were obvious double-counting. Some of the gigawatt training clusters announced in press releases would be silently scaled back when the reality of substation lead times set in. Transformers, the ordinary steel boxes on telephone poles, were on eighteen-to-thirty-six-month back order, and the heavier ones for utility substations on three-year-plus waits. The grid had absorbed the rise of personal computing, of the internet, of cloud computing, with a steady hum that nobody in finance had to think about. AI was different. AI scaled with the size of the cluster, not the size of the user base. A single building could consume a small grid.

For the first time since the early 1970s, an American technology industry was being constrained by physical infrastructure the country had stopped knowing how to build. Skilled trades were short. Heavy electrical equipment was on allocation. Permitting timelines for new transmission lines stretched past a decade. Nuclear engineers were being recruited out of retirement. Boomers in the natural gas industry were getting calls from headhunters about jobs they had aged out of years earlier. The fab-capacity story that had defined the chip industry from the 1980s through the 2020s had a sibling now. The new sibling did not have a glamorous photo. It looked like a substation in a soybean field outside Columbus, or a transformer yard in DeSoto County, or a polar crane being recertified at Three Mile Island.

The chip war’s central question, since 2022, had been who could fabricate the wafer. By 2026, an equally hard question sat next to it: who could deliver the gigawatt. Power had become its own choke point, and unlike chips, it could not be shipped overnight from Taiwan in a 747. It had to be built, line by line and stack by stack, somewhere downwind of the people who would have to live with it.