Standing before a friendly crowd in March, Elon Musk laid out his plan for the future of his companies, and it was literally out of this world.

Musk announced that his space-launch company, SpaceX, which had recently merged with his artificial intelligence company, xAI, would put data centers into orbit around the Earth.

It all comes down to electricity, he explained. "You're power constrained on Earth," he said. "Space has the advantage that it's always sunny."

Musk envisions legions of data-crunching satellites spinning around the planet, powering the AI revolution from above. It's the perfect pitch for taking SpaceX public. This week, Bloomberg reported that the company had filed documents confidentially to the Securities and Exchange Commission with the goal of listing an initial public offering this summer.

Musk also claims it makes financial sense. "I actually think that the cost of deploying AI in space will drop below the cost of terrestrial AI much sooner than most people expect," he said. "I think it may be only two or three years."

Others are skeptical. Musk's timeline is "an optimistic interpretation," according to Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University who specializes in putting computers on satellites. The napkin math looks appealing, and power is free up there after all — but it turns out there are a lot of obstacles to building a data center among the stars.

A global power problem

Here on Earth, the problem is glaring: AI is gobbling up electricity around the globe. Global data-center power consumption is expected to roughly double to nearly 1,000 terawatt-hours by the end of the decade, according to an estimate by the International Energy Agency.

To fill the gap, some companies are building dedicated gas turbines, while others are investing in nuclear technology. It's not enough, according to Philip Johnston, CEO and co-founder of Starcloud, which is seeking to build orbital data centers.

"We're very quickly running up on constraints on where you can build new energy projects terrestrially," Johnston said. "Within six months, they'll just be leaving chips in warehouses because they don't have power for turning them on."

Starcloud launched its first spacecraft last fall with an Nvidia H100 chip on board. The company demonstrated the ability to run a version of Google's Gemini AI from space, and it plans to launch a second spacecraft in October. "That one has 100 times the power generation of the first one," Johnston said, though it's still expected to generate only around 8 kilowatts of power.

Google is also pursuing the idea of building data centers in space through a project known as Suncatcher. It envisions an 81-satellite cluster that it plans to build in partnership with the satellite-imagery company Planet. Two prototype satellites will launch in early 2027, according to the companies.

"Orbital data centers are an idea whose time has come," Will Marshall, Planet's CEO, wrote to NPR in an email. "When exactly it will be more cost efficient than terrestrial ones is debatable but now is the time to be working on this."

Everything must get bigger

To go from a handful of prototype satellites to something useful is not so easy. For one thing, the power requirements of the microchips used for artificial intelligence are enormous.

To get a sense of just how much power is needed, consider the largest power-producing facility in space right now: the International Space Station (ISS).

The solar panels of the ISS are around half the size of a football field and produce around 100 kilowatts of average power, according to Olivier de Weck, a professor of astronautics at the Massachusetts Institute of Technology. "It's basically the amount of power that a single big car engine produces."

To replicate a 100-megawatt data center in space would require a facility that's 500 to 1,000 times, depending on the orbit.

"Is that feasible? Yeah, I think it's feasible, but not next year and certainly not in three years," he said.

And power is not the only requirement; the satellites also have to provide cooling to the microchips. While it's true that space is cold, it's also a vacuum. This means that when a satellite gets hot, there's no easy way to get rid of that heat — it just builds up.

"All of that heat that the computer generates has to be dispelled," said Rebekah Reed, a former NASA official now at Harvard University's Belfer Center for Science and International Affairs.

The best solution is radiators, which move liquids out to giant panels where the heat can be dissipated. So in addition to solar panels, an AI satellite would need another set of large radiators.

"When you put those massive radiators together with massive solar arrays that are required in order to power and cool, you're actually talking about really large satellites, or very, very large satellite constellations," Reed said.

An alternative is to build smaller satellites and fly them in preset formations called constellations. Such constellations allow the heat and power problems to be distributed, but to work, the satellites would need to send huge amounts of data back and forth. That likely means using lasers to beam data between satellites. But even moving at the speed of light, the time it takes to get data from one satellite to another is long enough to slow down computing.

Google's Project Suncatcher proposes flying groupings of satellites in extremely tight clusters to reduce that latency. Musk, meanwhile, has proposed launching upward of a million satellites and placing them in orbit around Earth's poles. He recently unveiled the first generation "AI Sat Mini" spacecraft — with solar arrays spanning roughly 180 meters (about 600 feet) — during his presentation.

Launching all that into space would cost money — lots of money. At the moment, it can cost around $1,000 per kilogram to launch a satellite into orbit. Google believes that cost must drop by at least a factor of five to $200 per kilogram before data centers in space will begin to make sense.

Musk thinks he can do it with his new Starship rocket, which is still in development. Starcloud's Johnston says Starship is central to more than just SpaceX's vision. He told investors: "If you don't think Starship's going to work, don't invest in us — that's totally fine."

Upgrading the server

Even if a company could get a data center into space, running it would involve a lot more than just moving microchips into orbit.

Data centers on Earth are not just static buildings full of chips humming away, says Raul Martynek, the CEO of DataBank, a company that maintains 75 data centers, primarily located in the United States. They require constant maintenance and upgrades, all of which is done by workers.

Take DataBank's IAD1 data center in Ashburn, Virginia. The facility is 144,000 square feet filled with rows and rows of black computer cabinets, which are filled with microprocessors. It's fairly run-of-the-mill, as these facilities go, but it still consumes around 13 megawatts of power at any given moment (that's 130 times more than the International Space Station).

"We have vendors here every single day," says James Mathes, who manages IAD1.

Workers are constantly in and out of these data centers, installing new servers, upgrading microchips and fixing things. And to stay competitive, space data centers would need to do much of the same.

Some of that could be done through software, and Musk points out that chips can be rigorously tested on the ground before they're sent aloft. But the fact remains that the companies that rent data centers often want to access them physically for one reason or another.

Martynek, who has spent decades in telecom, says he's not worried about space data centers taking business from his company.

"It seems like there's a lot of ifs and a lot of advancements that would have to occur, and I find it kind of hard to believe that all that could happen in two or three years," he said. "No one in data center land is losing any sleep."

Contact Geoff Brumfiel on signal at gbrumfiel.13