Project Ludicrous

This past January, a motley crew of CEOs joined President Trump at the White House to announce the beginning of a new “Golden Age” in America: Larry Ellison, Masayoshi Son, and Sam Altman, “by far the leading expert” on artificial intelligence in Trump’s view. The catalyst would be Stargate LLC, a privately funded joint venture between Oracle, OpenAI, SoftBank, and the Emirati Sovereign Wealth Fund MGX.

With the blessing of the president, these companies have promised to invest half a trillion dollars into building new AI data centers on American soil over the next five years. Their goal is simple: win the AI race for the United States and achieve artificial super-intelligence, or “AGI”—short for “artificial general intelligence,” or “systems that are generally smarter than humans,” as defined by OpenAI. But this modern-day “New Deal,” to quote Altman, remains beset by two existential questions. Can it find the revenue to pay its investors? And where will it find the energy required to power five-hundred-billion-dollars’ worth of data centers?

The hope of Altman, Ellison, and Son—not to mention Mark Zuckerberg, Jensen Huang, and other Silicon Valley robber barons—is that answering one of these questions will help answer the other; that, in the same way much of the Intercontinental Railroad was brought into existence through unbridled speculation, government largesse, and the dramatic scaling up of oil production, the future of AI will be made possible by a mix of speculation, public subsidy, and an energy industry that has a vested interest in the discovery of AGI. Or, more accurately, a vested interest in the search for AGI.

The first element is already in the works: the foolhardiness of speculators is exactly what drove semiconductor manufacturer Nvidia’s market capitalization over $3 trillion last November—making it the most valuable company in the world—and, likewise, what explains its record $589 billion loss in valuation in late January, following the headline-generating release of R1, the new large language model from DeepSeek. That the Chinese startup’s latest model was cheaper and more efficient than ChatGPT hardly explains the record scale of Nvidia’s losses, not to mention the jaw-dropping trillion dollars lost across the broader AI industry over the span of a week. Rather, the magnitude of these numbers reveals that investment in AI, and AI infrastructure just as importantly, is less an investment in the technology itself than in AI companies’ future performance as an asset. It’s not news that we live in an asset economy, but the extent to which different industries across the economy have been swept up in the financialization of AI’s infrastructure may come as a surprise—especially considering many of its normal-person-facing applications come in the form of chatbots that poorly edit essays and churn out dubious therapeutic advice.

These vested interests—these speculators—take many forms. They include real estate investors hoping to make it big off of development in the towns surrounding new U.S.-based chip-manufacturing plants, or off of the data centers themselves (like private equity giants Blackstone and KKR); investors in massive semiconductor-manufacturers like Nvidia (such as asset management firms BlackRock and Vanguard); investors in privately traded model-building companies like OpenAI (including SoftBank and Microsoft); and speculators hoping to cash in on the seemingly endless power that the hunt for AGI will require, among them ExxonMobil and Chevron, who remain bullish in spite of DeepSeek’s supposed efficiencies. There is also the resurgent nuclear power industry. Though much smaller than oil and gas, it stands to grow exponentially from all the speculation surrounding AI. But whether the uptick in investment will actually produce functioning reactors is another question. 

At the center of the hubbub is Helion, a nuclear-fusion startup valued, as of now, at $5.4 billion. It’s backed by $375 million of Altman’s own money—making it his largest personal investment by far—as well as by billions from other famous Silicon Valley investors and investment firms, like SoftBank, Peter Thiel’s Mithril Capital, and Facebook cofounder Dustin Moskovitz’s Good Ventures Foundation. The twelve-year-old startup has not one lofty goal, but three: first, to produce commercially viable fusion energy, a feat that’s nearly been a century in the making; second, to build manufacturable, small-scale reactors; and third, to do both of these things by 2028. Helion claims to have a prototype “fusion machine” named Polaris up and running as of late 2024.

But across the nuclear energy industry and even in academia, there are concerns about the soundness of Helion’s science and the extraordinary expectations they’ve set for themselves. For one, Helion doesn’t engage in peer review—in fact, no one even knows if the fanatically secretive company’s Polaris “machine” is actually doing what they say it is. What we do know is that their reactor design is notably novel. Coupled with the fact that they’ve gotten in the habit of talking to YouTubers and pop scientists, rather than their peers or the press—they’ve even invited a vlogger to tour their secretive facilities—the company has a public face that’s sometimes less than serious. One of the fusion researchers I spoke to for this story compared Helion to Theranos because of its lack of peer review. Other scientists have called for more transparency in the industry, warning that “the public at large . . . needs to understand how the technology is progressing, which should not be entirely left to the promotional efforts of startup companies.”

On its face, Helion’s timeline appears ridiculous. A Bloomberg investigation over the summer reiterated that this was also felt internally: one employee wrote to colleagues that the promise to get Polaris functional by year’s end didn’t sound “feasible,” and another insinuated that “in private conversations, even Kirtley didn’t sound certain.” To meet that deadline, Helion offered all of its employees a $50,000 bonus as well as $50,000 in stock. The accelerated timeline has also worked to deflect criticism away from its lack of peer review. “Right now the focus is: How do we build, and ideally on how do we move as fast as possible to get Polaris up and running,” Kirtley told Bloomberg. The company’s timeline was even used by a press representative to turn down my request for an interview; they’re simply too busy.

Other fusion startups are operating on more realistic timelines. Firms like Commonwealth are taking it slower—and no major firm besides Helion claims that they’ll have commercially available fusion energy this decade. These spectacular and attention-grabbing objectives have worked so far—to reel in investment. In 2023, it entered a power purchase agreement with Microsoft, Altman’s former employer and OpenAI’s primary investor, to help the tech giant become carbon-negative by 2030. And as the joint-venture Stargate seeks to name an energy partner, the Son and Altman-backed startup, which has never turned a profit, looks to be a possible contender. Helion’s CEO, David Kirtley, told Forbes that his company “would be perfect for a project like [Stargate].”

AI speculation has even managed to resurrect Three Mile Island, the site of the most serious nuclear accident in U.S. history. Late last year, Constellation Energy announced they would be reopening the plant, after securing a twenty-year power purchase agreement with Microsoft. Crane Clean Energy Complex, as it was renamed, is slated to be operational in 2028 and will power the company’s data centers. Microsoft isn’t the only tech giant looking for nuclear energy. Amazon Web Services (AWS) is following suit, developing a data center campus up the Susquehanna River, next to a different power plant. And further south, in Tennessee, Google is funding the firm Kairos Power, which plans to build small modular reactors all to help support Google’s AI services. Meta has also been looking for nuclear-powered data centers, but, as the Financial Times reports, their most recent proposal, at an undisclosed location, was thwarted in part by the presence of a government-protected species of bee.

Meanwhile, Stargate LLC’s first data centers are being built at Abilene Christian University in West Texas and are due to be completed next year. What was once meant to be a renewable-powered Bitcoin-mining operation will now become at least ten, and up to twenty, data centers, each a half a million square feet. The first development is registered under the name “Project Ludicrous.” Earlier this year, it was reported that Project Ludicrous alone would require about 200 megawatts of power capacity and that the plan would be to eventually scale that up to more than a gigawatt.

Other speculators are hoping to cash in on some of the demand created by Stargate. In February, D.C.-based Last Energy, another nuclear startup, announced its plans to build thirty “microreactors” north of Abilene. The company said in a press release that this was to meet Texas’s data center deployment projections—never mind the fact that while they hope to deploy their first microreactor by 2029, only Russia and China have ever managed to actually build a small modular reactor let alone a microreactor, despite billions and billions of dollars in investment. The money continues to pour into Abilene: Rusty Towell, the lead developer of a separate nuclear reactor, on ACU’s campus itself, told the Texas Tribune that the rush around the tiny town was going to “bless the world.”

The maelstrom looks to consume more than just Abilene. According to the Department of Energy, data centers consumed 4.4 percent of total U.S. electricity in 2023—but will consume anywhere between 6.7 and 12 percent of total electricity by 2028. In order to meet the growing demand, the tech industry isn’t just relying on infrastructure that is yet to be built: it needs to leverage our existing electrical grid, or more precisely, the power plants within the grid.

Days before the presidential election, the Federal Energy Regulatory Commission (FERC), rejected a request from Houston-based Talen Energy to increase its nuclear power supply to the aforementioned AWS data center campus, near Talen’s Susquehanna plant in Pennsylvania. The request had been challenged by American Electric Power and Exelon, two massive public utility corporations, who argued that diverting any more power to AWS’s data centers would affect both the grid’s reliability and would increase the power bills of its customers. (The two companies also stand to lose a bunch of money if energy-hungry tech were to try to bypass the grid.) FERC agreed with the utility companies: its commissioner Mark Christie wrote in the decision that data centers being “co-located” next to power plants, including nuclear ones, “present an array of complicated, nuanced and multifaceted issues, which collectively could have huge ramifications for both grid reliability and consumer costs.” The chairman of the committee dissented, writing that blocking the deal put up “unnecessary roadblocks to an industry that is necessary for our national security.”

Talen has since sued. But according to Brandon Oyer, who leads carbon-free energy procurement for Amazon’s data centers, the company’s foray into “co-locating”—what it’s called when, in order to bypass the grid, a data center is built right on top of a power plant—was a “sort of experimentation,” and the company prefers to remain or return to being “grid-tied,” as they were before the AI boom days. That might be just good PR, of course, a way of pushing back against negative press, but it’s true that co-locating on top of existing plants hasn’t yet become the norm. Fossil fuel companies like Chevron have recently announced their plans to build new power plants on top of data centers, rather than the other way around, much in the same way that Trump was dreaming of during his press conference. But natural gas power plants like the ones Chevron is planning to build can be built a lot faster than conventional nuclear ones—which is one major reason why AI’s carbon footprint is expected to grow exponentially before less intensive nuclear energy is available at the necessary scale.

For now, Helion has yet to buck its reputation for blowing past its self-imposed deadlines. In 2014, Kirtley told the Wall Street Journal that, if all went according to plan, they’d have a working prototype that produced net energy output in the next three years; in 2016, they told Science News they’d have commercially ready energy in six years; and today they claim that Polaris is working—or, as a press representative told me tactfully: “We began initial operations in [sic] Polaris before the end of 2024 and we’re currently accelerating our testing program.” These sorts of timelines haven’t been completely ignored by the press: articles in Bloomberg, MIT Technology Review, and The Observer have raised serious questions about Helion. Investors, however, seem happy to look the other way: during their latest funding round, they raised a whopping $425 million.

But it’s not obvious that more capital is the answer.

Some have remained steadfast in their stance that “AI scaling laws will continue”—that diminishing returns on AI infrastructure are neither here nor right around the corner—yet, as made clear by Nvidia’s record drop in valuation in January, there are at least concerns about diminishing returns on investment. More institutional players in the field have started to hedge. In late February, Microsoft cancelled some leases with data center providers, lightly suggesting that perhaps the company had realized that it had overextended itself—maybe it wasn’t really going to need all that computing power. The slightest hint of there being an oversupply of data centers has deflated some of the hype around AI infrastructure: share prices for Constellation Energy immediately sank nearly 6 percent, and other nuclear companies’ stock soon followed. Last month, Microsoft followed up by walking away from 2 gigawatts worth of data centers, sending Constellation stock price down again.

Whether this trend will pick up steam remains uncertain, but for lack of a better word, the nuclear sector seems to be a ticking time bomb. Regardless of Helion’s bold timeline, scientists and engineers will need years to meaningfully scale up nuclear energy production—by which point investors may have soured on AI. OpenAI, for one, lost $5 billion last year, and that’s excluding stock-based compensation.

In the meantime, Trump plans to power AI the old-fashioned way. “Drill, baby, drill,” he cooed during the Stargate press conference, despite the fact that many American fossil fuel companies have made it clear they don’t want to drill more, and that production is already high enough to maximize shareholder value. Earlier this month, Trump also signed an executive order removing barriers to coal leasing and mining—with an eye toward using unreliable and expensive coal-fired electricity to help power new AI data centers.

The coming “Golden Age” will, it turns out, be powered by pretty much whatever is available—save for the offshore wind industry, which Trump has taken particular pains to go after. It appears that what matters is the fantasy that the source of energy conjures up, not its practical applications—be it hyper futuristic nuclear fusion technology or reindustrialized coal towns, both of which couldn’t be further from our collapsing present.