Access to electricity now limits how fast AI companies can grow, because their products run on computing resources that increase only as fast as new data centers can be connected to power sources. Anthropic recently spelled out that limit in its IPO filing, telling investors that demand for its services “has historically exceeded available capacity” even as its revenue grew more than tenfold over the first nine months of 2026. The filing also says that electricity “has emerged as a significant constraint on data center development and expansion in many geographies, including the United States.”
Anthropic’s filing is more evidence of a limit some already suspected the broader U.S. AI industry is facing, at the same time demand is expected to keep rising. The S-1 filing says the company’s AI agents, which carry out multistep tasks on their own, are “structurally more compute-intensive” than its other products, so every task that shifts to an agent adds to the load. Across the industry, Morgan Stanley estimates that U.S. data centers will come up about a third short of the power they need through 2028, even counting projects with their own on-site, or “behind-the-meter,” power plants. Even worse, these companies are already behind and paying for the shortfall. Anthropic, for example, contracts most of its compute 12 to 18 months in advance, and when demand outruns that, it buys far more expensive short-term capacity: $5 billion worth in the third quarter alone.
That sets up a critical frontier economy tension for the next two years: AI companies need power connected to compute faster, while many residents and the officials they elect want data center builds slowed down or stopped entirely. Six in ten Americans say they would be uncomfortable with a new data center in their area, and on Tuesday, data center opponents appear to have won nearly every one of about 40 local ballot measures in 10 states. In Ashville, Ohio, voters even moved to block a data center campus that’s already under construction. Voter resistance is at its very strongest in what we call the referendum belt, where three conditions overlap: a data center planned on farmland that must be rezoned, a town that has never hosted one, and a state law that lets residents force a public vote on the rezoning.
One way or another, these conditions will sort U.S. political geography into boomtowns and holdouts. Here, we use “boomtown” to describe a place that agrees to host data centers and connect them to power, and collects the construction jobs, tax revenue and grid investment that follow. The likeliest successful boomtown candidates have land already zoned for industry, experience hosting data centers or power plants, and state rules that settle where data centers can go before a local fight starts. Holdouts keep their farmland and local control, and investment capital flows elsewhere.
Tuesday’s election moved some of the biggest data center markets toward the holdout column. In Texas, which accounts for nearly 30 percent of the U.S. data center capacity due online in 2028, Democrat Gina Hinojosa defeated Gov. Greg Abbott after calling for a moratorium on new data centers until the legislature passes rules protecting communities and utility bills. Democrats who called for moratoriums of their own also won the governor’s races in Ohio and Georgia, and Bloomberg reports these wins could leave about half of proposed U.S. AI data center additions facing stricter scrutiny. Pennsylvania held steady by re-electing Gov. Josh Shapiro, who has courted data centers while criticizing speculative projects.
Financial markets are still sorting out what the results mean, and investors are split on who comes out ahead. AI data center stocks fell on Wednesday as investors weighed whether state restrictions would delay projects, and Libby Cantrill, head of public policy at the investment firm PIMCO, said she expects “a lot more sand in the gears of the AI buildout.” One reading, mentioned by Stephen Weiss of Short Hills Capital Partners in a CNBC appearance Wednesday, is that a forced slowdown blunts the advantage of the companies spending the most, since money cannot buy power that has not been connected; for example, Microsoft CEO Satya Nadella said last year that his company had “a bunch of chips sitting in inventory that I can’t plug in.” The opposite reading, from Morgan Stanley, is that new power is going first to “a handful of large, credit-worthy players,” which would widen their lead. Either way, every company will have to get more out of the power it has, which is why Nvidia now pitches its chips on how much AI output they deliver per watt.
Resolving this tension means making data centers worth hosting for the communities that decide, and three approaches are already in play in the U.S.:
- Make data centers pay their own way. Florida’s new law requires large data centers to cover the full cost of their electricity while leaving local governments the final say on where they go, and the biggest AI and cloud companies have pledged to pay for the power plants and grid upgrades they need.
- Let data centers bring their own power. Behind-the-meter power plants skip the wait for a grid connection, but Anthropic’s filing says such arrangements are “feasible only in some jurisdictions given the applicable regulatory frameworks,” which makes state rules decisive.
- Build on old industrial sites. The federal government is turning former uranium plants in Kentucky and Ohio into data center sites, and Pennsylvania’s retired Homer City coal plant is becoming a 3,200-acre data center campus.
Whatever part of the estimated one-third shortfall those approaches fail to close will push data centers toward experiments at home and toward other countries. After the election, Elon Musk posted that “data centers in space” were “looking pretty appealing right now,” but the reality is that data centers in orbit, at sea, and underground are still too small or experimental to deliver gigawatts of capacity in the next two years. Meanwhile, other countries are ready to host some of the data centers the U.S. can’t connect to power in time. Anthropic’s filing says it is negotiating for data center sites outside the U.S., and the company is hiring to lease capacity in Europe and Australia after signing a $10 billion deal for computing power in Norway in August.
Where that computing power lands will shape America’s current lead in AI. Morgan Stanley has already lowered its estimate of America’s share of the world’s computing capacity by five percentage points, to 55 percent, as more of it moves abroad. Computing power built overseas also falls under other governments’ rules, which Anthropic’s filing lists as “data sovereignty and localization laws, governmental access or assistance requirements.” More ominously, China has far more cheap electricity available to offer its homegrown AI companies. Europe, on the other hand, is falling further behind: connecting a new data center to the grid can take up to seven years there, and Roland Berger expects Europe’s share of global data center capacity to slip to about 10 percent by 2030. Even France, which has plenty of nuclear power, is still catching up. Mistral, its leading AI company, is bringing a 44-megawatt data center online near Paris, while a 1.4-gigawatt campus it is developing with Nvidia and other partners is not due until 2028.
For America, the best outcome is enough boomtowns at home to meet AI’s rising demand, built on clear rules set before any land is bought and ideally at the national level, so that AI companies, developers, utilities, investors, local officials and residents all know who pays for the power and what each community gets in return. Access to electricity sets how fast the AI industry can grow, and the officials elected this week, many of whom campaigned on slowing data centers, will help decide how much of that growth happens in the United States in the years to come.