New York — Elon Musk pledged this week to shake up another industry, just as he did with cars and rockets, and once again faces long-term odds.
The world’s richest man said he wants to put up to a million satellites into orbit to form massive solar-powered data centers in space – a move to allow expanded use of artificial intelligence and chatbots without causing power outages and higher utility bills.
To fund the effort, Musk merged SpaceX with his artificial intelligence company on Monday and is planning a large initial public offering for the combined company.
“Space-based AI is clearly the only way to scale,” Musk wrote on the SpaceX website on Monday, adding of his solar ambitions, “It’s always sunny in space!”
But even Musk — who has outmaneuvered Detroit to turn Tesla into the world’s most valuable automaker — faces formidable technical, financial and environmental hurdles, scientists and industry experts say.
Here’s a look:
Capturing the sun’s energy from space to power chatbots and other AI tools would ease pressure on power grids and reduce demand on sprawling computing warehouses that consume farms, forests and vast amounts of water for cooling.
But space poses its own set of problems.
Data centers generate tremendous heat. Space seems to offer a solution because it’s cold. But it’s also a vacuum, trapping heat inside things in the same way a thermos keeps coffee hot using double walls with no air in between.
“An uncooled computer chip in space will heat up and melt much faster than one on Earth,” said Josep Jornet, a professor of computer and electrical engineering at Northeastern University.
One solution is to build giant radiator panels that glow in infrared light to push heat “out into the dark void,” Jornet says, noting that the technology has worked on a small scale, including on the International Space Station. But for Musk’s data centers, he says, it will require a set of “huge, fragile structures that have never been built before.”
Then there’s space junk.
The failure of a single satellite or loss of orbit could result in a series of collisions, which could disrupt emergency communications, weather forecasting and other services.
Musk noted in a recent regulatory filing that he has only experienced “one low-speed debris generation event” in seven years of operating Starlink, his space communications network. Starlink has operated about 10,000 satellites, but that represents a small fraction of the million or so it now plans to put into space.
“We could reach a tipping point where the chance of collision will be very high,” said John Crassidis of the University of Buffalo, a former NASA engineer. “These objects are traveling at speeds – 17,500 miles per hour. There can be very violent collisions.”
Even without collisions, satellites fail, chips disintegrate, and parts break.
For example, special GPU graphics chips used by AI companies can become damaged and need to be replaced.
“On Earth, what you would do is send someone into the data center,” said Baiju Bhatt, CEO of Aetherflux, a space solar energy company. “You can replace the server, replace the GPU, do some surgery on the thing and then bring it back in.”
But there is no such repair crew in orbit, and GPUs in space could be damaged by exposure to high-energy particles from the sun.
One solution, Bhatt says, is to overstock the satellite with additional chips to replace those that fail. But that’s an expensive proposition, since they would likely cost tens of thousands of dollars each, and current Starlink satellites are only about five years old.
Musk is not alone trying to solve these problems.
A company in Redmond, Wash., called StarCloud, launched a satellite in November carrying a single AI computer chip made by Nvidia to test how well it fares in space. Google is exploring orbital data centers in a project called Project Suncatcher. Jeff Bezos’ Blue Origin company in January announced plans to begin launching a constellation of more than 5,000 satellites late next year, although its focus has been more on communications than artificial intelligence.
However, Musk still has an advantage: he has rockets.
Starcloud had to use one of its Falcon rockets to put its chip into space last year. Aetherflux plans to send a group of chips it calls Galactic Brain into space on a SpaceX rocket later this year. Google may also need to turn to Musk to launch the first two prototypes of its planned satellites on Earth by early next year.
Musk routinely charges competitors much higher than he charges himself — up to $20,000 per kilogram of payload versus $2,000 internally, says Pierre Lyonnet, research director at trade association Eurospace.
He said Musk’s announcements this week indicate he plans to use this advantage to win this new space race.
“When he says we’re going to put these data centers in space, it’s a way of telling others that we’re going to keep these low launch costs for myself,” Lyonette said. “It’s kind of a power play.”