In the 1960s, a paper by the scientist Freeman Dyson envisaged a mega-construct that would surround the Sun with solar generators to harness infinite energy.
His theory was that a suitably advanced alien civilisation would inevitably develop such a vast power source. The concept has since been named a “Dyson sphere”.
Such an idea is great fodder for science fiction. But now, these fantastical concepts are being repeated by Sam Altman, the chief executive of OpenAI, the developer of ChatGPT, and SpaceX boss Elon Musk.
Altman has floated the idea of launching a vast fleet of orbital data centres, powered by solar energy, to fuel a futuristic superintelligence.
“I guess that a lot of the world gets covered in data centres over time,” the OpenAI boss told a podcast host. “Maybe we put them in space. Maybe we build a big Dyson sphere.”
Earlier in December, Musk also started to run with the idea. In a post on X, the world’s richest man said satellites would soon be the “lowest cost way to generate AI bitstreams”.
A clutch of start-ups is now racing to pursue the out-of-this-world concept, while Musk is said to be pinning his plans for a $1.5tn (£1.1tn) flotation of SpaceX, his rocket business, on building a network of data centres in orbit.
One business, Starcloud, has a head start. Last month, it launched the first prototype satellite, Starcloud-1, carrying a powerful Nvidia graphics card into space, which it has used for AI processing.
“Lots of people thought you couldn’t run these high-powered AI chips in space,” says Philip Johnston, the chief executive of Starcloud. “We’ve proved conclusively that you can.”
Starcloud published a white paper last year explaining its idea. It called for satellites with a vast 4km by 4km solar array to power a data centre of up to 5GW.
For comparison, a 5GW data centre planned by Facebook owner Meta in Louisiana is roughly the size of a small city.
For now, Starcloud is eyeing a constellation of smaller, server-carrying satellites.
Johnston says the start-up was “pilloried from all corners. People thought this was completely impossible.”
However, he says that in the “last month, people seem to have come around to the idea. I give Elon all credit for that”.
Starcloud is not the only space start-up getting into the data centre business.
Earlier this year, Lonestar Data Holdings, a Florida-based company, launched a module containing a tiny data centre about the size of a laptop aboard the Athena Lunar Lander to the surface of the Moon.
In the future, it plans to develop a data centre in lunar orbit that can serve as a kind of backup storage system for critical information.
Meanwhile, Axiom, a Houston-based space company, launched a shoebox-sized data centre module to the International Space Station this summer to run experiments and plans to launch a fully functioning data centre node by 2027.
As well as Musk, other “big tech” giants are also exploring the concept.
Google has said it is planning to develop its own orbital data centre project, called Project Suncatcher. The search giant argued that in orbit, a solar panel should be up to eight times more productive than on Earth, potentially making space the “best place to scale AI compute”.
It is working with a US-listed satellite business, Planet Labs, to develop its space-based data centres, with the aim of launching a prototype by 2027.
Jeff Bezos, the Amazon founder and a rival to Musk, is said to be developing technology through his Blue Origin rocket business, according to the Wall Street Journal.
Altman of OpenAI also reportedly held talks to acquire a rocket start-up in a bid to bring the ChatGPT developer into the space race.
Supporters argue that space-based data centres would have access to unlimited energy from sunlight and avoid the tedious environmental and planning hurdles that threaten to delay or halt projects on Earth.
However, other space experts argue there are major practical and technological barriers ahead for any business attempting to move data centre operations into orbit.
Andrew McCalip, an engineer at Varda Space Systems, writes in one widely shared blog post: “The physics doesn’t immediately kill it, but the economics are savage.”
He adds that a company with its own launch capability, such as SpaceX, could potentially recoup costs from such a project.
The challenges
Others are less generous, pointing out that vast fleets of data-beaming satellites threaten to add to a growing mass of space junk orbiting Earth.
Pierre Lionnet, the managing director of Eurospace, wrote in response to Musk’s post: “The insanity has begun.”
Talk of data centres in space and breathless mentions of “Dyson spheres” come as hype around AI reaches fever pitch, with OpenAI promising to spend $1.4tn on digital infrastructure.
Advocates argue that demand for AI computing power will be insatiable for the foreseeable future, forcing companies to seek alternative ways to meet it.
Critics of the proposals also point to technical challenges, such as developing solar arrays for space that are significantly bigger than the largest currently in orbit on the International Space Station.
Data centres also generate significant amounts of heat. In the vacuum of space, this heat needs to be dispersed with heavy-duty radiators, which must be blasted into orbit at high cost.
Johnston of Starcloud, a British entrepreneur who moved to the US to work in the space industry, says his start-up plans to address this challenge.
“About 70pc of our engineering team is working on building this very large, low-cost and mass-deployable radiator,” he says. “There’s no new physics that needs to be discovered. The innovation that we are doing is making this thing very cheap and light.”
Johnston also believes that SpaceX’s giant Starship rocket should bring the cost of space launches down to ultimately make it economical to fire data centres hundreds of miles above the planet.
“The break-even launch cost needs to be about $500 a kilo,” he says. “Today, it is around $5,000 a kilo. SpaceX is targeting $10 to $20 a kilo for the marginal launch cost with Starship. They’re saying customer costs are $50 a kilo.”
Still, to even get to that point, orbital start-ups will need Musk to perfect Starship. So far, the rocket remains in the testing phase, with several launches ending with unplanned explosions.
Lionnet, who worked on a space data centre project for the European Space Agency, says: “It is a long way before this thing is really affordable in terms of launch. The cost of launch is still way too high.”
Perhaps the bigger challenge for a minnow like Starcloud, which has so far raised around $34m, is going into battle with Musk himself.
In recent weeks, the multibillionaire has declared SpaceX will develop its own space-based data centres.
Lionnet, of Eurospace, warns that if SpaceX begins its own project, it will have “no incentive to lower the launch costs. I don’t see a space for other players.”
However, Johnston argues that some AI companies may prefer to rely on an alternative provider rather than work with a rival such as Musk.
“It’s great to be validated,” says Johnston. “The question I used to get from venture capital investors was, ‘This will never work. You guys are crazy.’
“Now [they say], ‘You guys are going to get crushed by Elon.’
“I think I prefer the second option to the first one.”