 "Even the sky may not be the limit as AI data centres prepare for orbit")
By Eli Tan & Ryan Mac
If the architects of the artificial intelligence (AI) boom are right, it is only a matter of time before data centres — the giant computing facilities that power AI — will float in orbit and be visible in the night sky like planets.
The science-fiction-like dream is being driven by AI and space industry leaders who are growing increasingly worried that data centers will eventually require more energy and land than are available on Earth. So one solution — perhaps the only solution, they say — is to start building them in space.
Google announced in November that it was working on Project Suncatcher, a space data center project that would begin test launches in 2027. Elon Musk said at a recent conference that space data centers would be the cheapest way to train AI “not more than five years from now.” Others pledging support for the idea include Jeff Bezos, the founder of Amazon and Blue Origin; Sam Altman, the chief executive of OpenAI; and Jensen Huang, the chief executive of Nvidia.
“It is not a debate — it is going to happen,” said Philip Johnston, the chief executive of Starcloud, a space data center startup. “The question is when.” The notion has gained traction as the AI race hits a fever pitch, fuelling fears of a potential bubble. Meta, OpenAI, Microsoft, Amazon, and other big tech companies are investing hundreds of billions in data centres worldwide, with OpenAI alone committing $1.4 trillion to such projects. Saudi Arabia and other nations are also pouring money into these efforts, while smaller companies pile up debt and take on financial risks to join the frenzy. Yet earthbound data centres are increasingly running into limits. In many places, the projects lack enough available power for the computing needs. Local opposition has also flared over whether data centers are driving up utility bills and exacerbating water shortages.
Technologists and scientists have researched the idea and concluded that some version of these projects may be possible in the next few decades. But skeptics said the proposals flew in the face of physics and would be astronomically expensive.
Data centres in space would look different from the football-stadium-size facilities on Earth. Most mockups from companies like Starcloud look like large satellites with a cluster of servers housing AI chips at the centre of miles of solar panels to power them.
The data centres would need to be rebuilt every five years, which is when the computer chips are typically replaced, Johnston of Starcloud said. They would be visible at dawn and dusk from Earth, he said, appearing in the sky as about a quarter the width of the moon. But it is too expensive to create space data centres today. A kilogram of material costs around $8,000 to launch into space, said Pierre Lionnet, a space economist and director at Eurospace. The cheapest rate — around $2,000 per kilogram — is offered by the rocket maker SpaceX, he added.
Tom Mueller, a former SpaceX executive who believes humans will hit the limits of terrestrial energy sources by 2040, said part of the reason Musk and other AI leaders were talking about space data centers was the financial opportunity. “The hottest thing to invest in right now is AI, and the second-hottest thing is space,” Mueller said. “Now they’re converging.”