Do space-based AI data centers make economic sense?

Following the astronomical success of SpaceX’s IPO (which raised $85.7 billion, valued the newly public company in the trillions, and made Elon Musk the world’s first billionaire), building AI data centers in space, which many skeptics still view as a pie-in-the-sky idea, is on the horizon. There’s good reason for skepticism, but as a result of SpaceX’s windfall, the concept could at least move in a more plausible direction.

SpaceX has reliable and reusable Falcon rockets, and it also has more powerful rockets in its wings. Meanwhile, the company’s xAI has an insatiable need for computing power, and its space-based internet service, Starlink, is equipped with upgradable satellites. The interconnected entity’s engineering and technology now has billions of dollars in new capital needed to integrate these components in space, not only to power SpaceX’s massive internal AI operations, but also to provide commercial services to an array of paying customers, including Anthropic.

Some investors argue that if the company wants to justify its public market valuation over the long term, it has no choice but to let the idea become a reality. “This company is going to end up in a data center in space,” Duncan Davidson, a partner at Bullpen Capital, told CNBC’s “The Exchange” the week before the IPO. “It’s a big, long game.”

Engineering and technical issues are being resolved, Davidson said. His company is not an investor in SpaceX, but it does have an indirect stake in space startup StarCloud. However, he added, “Financially, we are at the limit at the moment.”

And given the ever-increasing practical, political, and public constraints on ground-based data centers, the prospect of launching data centers into low orbit, where the sun shines 24/7, is no longer science fiction.

As Mr. Musk said, if SpaceX’s large Starship rocket is operational next year — certainly an “if,” given its track record of falling short of previously promised schedules — launch costs, a critical barrier to affordability, will be significantly reduced. Meanwhile, while the cost of building data centers on Earth may rise, “data centers in space will start to get cheaper and cheaper,” Davidson said. “So I think (business) litigation is very strong on these things,” he said.

In January, SpaceX applied to the Federal Communications Commission for a constellation of up to 1 million satellites to form the basis of an in-orbit AI data center. Two months later, at an event in Austin, Texas, Musk reiterated his previous claims that space-based, solar-powered data centers would become more cost-effective than ground-based data centers within two to three years at the earliest. “Increasing power on Earth gets harder and more expensive over time, but in space it actually gets cheaper and easier over time,” he said.

The so-called AI1 satellites will be an upgrade to the ones used in the existing Starlink communications network, dramatically increasing the number of semiconductors needed. The scale required is so large that SpaceX, Tesla, and Intel partnered to form TerraFab. Terrafab is a 10 million square foot facility under construction in Austin, expected to open in 2029, and could cost up to $119 billion to build.

SpaceX declined to provide details about its plans, offering previously published information about the orbital data center concept and Terafab.

Jeff Bezos and Alphabet are also in the race

SpaceX isn’t the only company in the race to compute in space. Amazon CEO Jeff Bezos has expressed similar aspirations for his rocket and AI ventures Blue Origin and Prometheus, respectively. Last month, Bezos said in an interview with CNBC that building a data center in space is “very realistic,” but he questioned how long it would take. “Some of the timelines we’re hearing are very short. People will say something like two or three years,” he said, perhaps referring to Mr. Musk’s bold predictions. “That’s probably a little ambitious.”

In March, Blue Origin submitted plans to the FCC to launch 51,600 data center satellites into low Earth orbit as part of its Project Sunrise initiative. Deployment of the proposed satellite constellation, called TeraWave, is expected to begin in the fourth quarter of 2027, the company said.

Alphabet Inc.’s search giant Google entered the race with SpaceX (which owns 6.1% of the company) as a potential launch partner, partnering with Earth observation satellite maker Planet Labs on its in-orbit data center concept, Project Suncatcher. Google said the project will explore how an interconnected network of solar-powered satellites equipped with Tensor Processing Unit AI chips can make the most of the sun’s power.

The paper describing Suncatcher notes that historically high launch costs have hampered large-scale space-based systems, but suggests prices could fall to less than $200 per kilogram by the mid-2030s. At that price, an operational orbital data center could be on par with the reported energy costs of a comparable terrestrial data center on a per kilowatt/year basis.

Beyond this paper, a Google spokesperson responded to a request for comment and wrote, “We have nothing new to share.”

Satellite, rocket and robotics startups are testing

Beyond the multitrillion-dollar world of tech stocks, several startups are eyeing the skies.

Starcloud has already sent Nvidia H100 GPUs into space on a test satellite aboard a SpaceX Falcon 9 rocket. “It’s simply cheaper to launch into space,” CEO Will Marshall told CNBC in a recent interview. Another benefit, he said, is that there is no need for communities of people to compete for water or electricity. It’s a long-term project, “but it’s also an exciting project,” Marshall said.

StarCloud is also partnering with Rendezvous Robotics, a manufacturer of modular spacecraft systems that self-assemble in space to generate power for orbital data centers. The spacecraft is made up of hundreds of interconnected hexagonal tiles, each about 5 feet in diameter, stacked atop a launch vehicle.

“Our tiles have been tested three times, once on the Blue Origin New Shepard flight and twice on the International Space Station,” said Rendezvous President Joe Landon.

Another test on the ISS is scheduled for later this year. “We will be able to offer a full-scale system in 2028,” he said.

rocket lab‘s Electron rockets have launched nearly 90 unique satellites into space for NASA, the U.S. Space Force, and numerous global customers. Founded in 2006 by self-taught engineer Peter Beck, Rocket Lab is building a more powerful reusable rocket, Neutron. This gives the vertically integrated company the ability to compete with SpaceX in the orbital data center market, even if it is much smaller.

“If this becomes a large market, we will be in a great position to attack it either as a merchant supplier, or with our own applications, or a combination of both,” said CFO Adam Spice. But if pushed into a corner, he said, “we’d rather turn our customers into tenants of the infrastructure we own than help them build their own infrastructure.”

Cowboy Space, founded in 2024 under the original name Aetherflux by Robinhood co-founder Baiju Bhatt, also has an end-to-end strategy, but its rocket and data center infrastructure are still developed in-house. The novel approach “involves using the rocket’s second stage as a data center satellite itself,” Butt said. “We’ll have more to reveal and show about that in the not-too-distant future,” he said, adding that the company is targeting its first launch into space later this year. Meanwhile, the company applied to the FCC for a constellation of 20,000 satellites.

A new commercial space economy is rapidly taking shape, and despite all the continued investment in research, development, and orbital AI data centers in space, the fundamental economic problems remain the same, says Mark Weinziel, a Harvard economist and tax expert. Mark Weinziel became interested in space-based businesses about 10 years ago and now writes and teaches on the subject. “One of the biggest questions is, can’t we do it cheaper on Earth?” he said.

Analyzing the current business model in the space sector, he said, “I’ve never seen a company that can say it’s cost competitive at this point.” But Weinziel believes it’s reasonable to predict a future in which “the cost of doing it on Earth will rise over time, the cost of doing it in space will fall over time, and at some point those two curves will intersect.”

But that requires assumptions about future economics based on today’s existing trends.

“Technology can always change,” he said. “The next generation of chips will probably consume less energy than the current generation,” which could lead to lower costs for terrestrial data centers.

Earth has regulatory, environmental and political constraints that Weinziel says currently make the space-based case work better, at least in theory. In fact, the backlash against mass data center construction has grown, with more than 100 moratoriums proposed at the local, county, state, and national levels. The people are also protesting. A Heatmap News poll conducted in May found that 7 in 10 Americans oppose data centers being built near where they live, up from 4 in 10 in August of last year. Additionally, a new study from climate risk analysis firm First Street finds that 79% of data center capacity faces increased risk from severe climate change.

In Weinziel’s view, that’s a bit of a gamble rather than a sure thing, but not necessarily an outlandish one.

“If we’re optimistic that the cost declines we’ve seen with advances in launch and satellite technology and solar (power) will continue, those (curves) lines will cross sooner. That’s always a gamble, but that seems reasonable to me,” he said.