
The Free Press

This week’s launch by OpenAI of its o3-pro artificial-intelligence reasoning model, its mightiest yet, is such a big step forward that Sam Altman hailed the arrival of what he called “the gentle singularity.” “Scientific progress is the biggest driver of overall progress,” OpenAI’s chief executive wrote in a blog post on Tuesday. “It’s hugely exciting to think about how much more we could have.”
But do we have enough energy to fuel it?
The AI boom is also a growing challenge to America’s strained power grid. As of six months ago, ChatGPT received more than a billion queries a day. The number is surely much higher now. And imagine that figure multiplied by the accelerating productivity of AI reasoning models, which will only increase their popularity and the power needed to run AI data centers.
Altman seems sanguine about that, writing in his blog post that the average ChatGPT query uses only as much electricity as a high-efficiency light bulb does “in a couple of minutes.” Many other people are worried.
In Abilene, Texas, construction is underway on the first site of the Stargate Project, a new company formed by OpenAI, Oracle, and SoftBank that plans to spend $500 billion on the largest data-center project in history.
The project could eventually consume five gigawatts of power a year, equivalent to Miami’s annual electricity consumption. A typical ChatGPT query takes more than 10 times as much power as a Google search. The microchips that ChatGPT runs on, called “graphics processing units” (GPUs), can consume as much power as a single-family home. Stargate is expected to use 64,000 Nvidia-made GPUs by 2026.
Abilene was so hungry for the project that the city and county agreed to giant tax breaks. “It will impact the rest of the economy—our restaurants, our homebuilders—with that many new people coming in and taking these jobs,” Mayor Weldon Hurt told Bloomberg.
The problem is that even before the AI surge, America’s energy grid was already slipping into disrepair. Successive reports during the last few years from the North American Electric Reliability Corporation (NERC), a grid watchdog, have warned that more and more of the U.S. is in danger of capacity shortfalls. That may sound like jargon, but what it means is a higher potential for blackouts like the ones we saw in Europe last month.
In 2023, NERC cited our energy policy as the No. 1 threat to grid reliability. We have been losing reliable power plants while replacing them with less-reliable wind and solar assets. NERC reports that while we are expected to add more than 400 gigawatts of power supply during the next decade, only 60 of those will be from power plants that can quickly ramp up the amount of electricity that they supply to the grid.
“We have to face some hard realities,” John Bear, the president and chief executive of Midcontinent Independent System Operator, a power market that spans from Minnesota to Louisiana, wrote last year. The transition “to get to a decarbonized end state is posing material, adverse challenges to electric reliability.”
In other words, America is struggling to keep the lights on, let alone absorb more annual growth in power demand than our country has seen since the Carter administration. Last month, Energy Secretary Chris Wright issued an emergency order to keep a coal-fired power plant in Michigan in operation to help the Midwest get through this summer.
Wright and President Trump are doing what they can to help more reliable power plants come online. Last month, Trump issued four executive orders aimed at hastening a nuclear renaissance. Nuclear power is our cleanest, safest, and longest-lasting power asset.
Large-scale data centers known as hyperscalers have been snatching up access to nuclear power wherever they can find it. Last year, Microsoft made waves when it announced that it would restart the remaining nuclear reactor at Three Mile Island, site of the accident that blackened the nuclear industry’s name in the late 1970s. This month, Facebook owner Meta signed a 20-year agreement with the utility Constellation to buy power from a nuclear plant in Illinois.
Data centers could gobble up so much power that they put some states in a bind. Illinois plans to shut off several Hoover Dams’ worth of power capacity in order to reach its climate goals. Unfortunately, the state can’t replace any of that power with reactors like the one at Clinton because of a partial moratorium on new nuclear construction.
All that aside, new nuclear power is still years away from deployment. President Trump’s executive order in April aims to provide incentives for reinvestment in coal plants, but executive orders aren’t enough. Stargate will be powered by natural gas. Elon Musk’s xAI’s mega data center near Memphis, Tennessee, has 35 gas turbines. About 20 percent of the gas we recover comes from oil drilling, but low oil prices reduce the incentives to drill even more.
Even if we could deploy gas turbines at will, hyperscalers are now running into obstacles about their “social license to operate,” or the level of public tolerance for data centers. For example, the xAI project in Memphis, where Grok churns out responses to queries on X, has been accused by some local residents of installing more gas turbines than it has permits for, and for increasing air pollution while harming air-quality standards.
Louisiana suffered blackouts in May when a nuclear plant tripped offline. Meta has a data center project near New Orleans that uses 2.5 times as much power as the city itself. Nearby residents are up in arms about the potential amount of power that could be siphoned off the grid by Meta, especially if that worsens their already-poor power service or raises their bills. California is the worst of both worlds. Despite the state’s reputation as an environmental bastion, it has added a gigawatt of diesel generation for data centers in Silicon Valley, much of it located near homes and schools. Residents aren’t happy about it.
Add these social concerns to the fact that not all data centers are Stargate. They don’t necessarily add many jobs, or even much tax revenue. Abilene will gain about 350 jobs when construction of the Stargate site there is complete. Last month, a major data-center project in Indiana was dropped for the same reasons. Even Loudoun County, Virginia, home to about a third of the world’s hyperscale data centers, has tightened zoning requirements that now make it more difficult to build data centers.
What if, as we build more data centers, we find that people want them less and less? This happened to nuclear, wind, and solar power. Why should data centers be any different, especially if the new world of AI makes people feel ill at ease rather than brimming with hope?
Then there is the nagging uncertainty that we don’t know exactly what this new power demand will do to our existing system. Late last year, BloombergNEF energy analyst Michael Liebreich predicted that AI “will also help mitigate energy demand in ways we can only begin to imagine.”
Joseph Dominguez, the president and CEO of Constellation, which struck the deal with Meta for all that nuclear power, said that “some out there” have “overstated the amount of new demand for their own reasons.” Even more modest growth would still be good for Constellation, though, he said.
There are those who think that the power grid can handle all this AI if hyperscalers try to avoid peak power-usage hours. “Our findings suggest that a small amount of flexibility from large loads like data centers can unlock significant grid capacity,” Tyler Norris, co-author of a recent study from the Nicholas School of the Environment at Duke University, told me. “The U.S. can’t replicate China’s top-down build-out pace, but we can harness technology to maximize utilization of our existing grid.”
Whether you think AI is a fundamental shift in our way of life or just a more sophisticated TI-83 graphing calculator, our future might depend on our ability to manage the power grid. If it fails, we fail.