Investors are already putting trillion-dollar numbers on orbital AI computing, while the engineers involved reckon useful space data centres remain decades away.
According to DigiTimes, SpaceX has made orbital computing part of its pitch to investors, with plans to fill low Earth orbit with satellites carrying AI processors and powered by solar energy. The idea is attractive because terrestrial AI data centres face increasingly awkward problems with electricity supplies, grid connections, cooling water and planning restrictions.
Space has plenty of sunlight and no neighbours complaining about a 500MW data centre appearing at the end of the road. Unfortunately, it comes with the minor inconvenience of launching several thousand tonnes of expensive electronics into orbit and the fact that no one has the ability to pull it off yet.
DigiTimes reports that capital markets are already placing enormous values on the idea of AI infrastructure in space following SpaceX’s public listing. However, companies actually developing orbital computing systems expect large-scale deployments to arrive during the next decade.
SpaceX has said it could launch its first experimental orbital computing systems as early as late 2027. The company has filed with US regulators for a constellation that could involve vast numbers of data-centre satellites. That timetable refers to experiments, not replacing Virginia with racks of GPUs floating above the atmosphere.
Launch capacity remains one of the largest problems. Orbital computing outfit Cowboy Space warned earlier this year that there simply are not enough rockets available to support the sort of deployments being discussed. Large projects are depending heavily on SpaceX’s Starship and Blue Origin’s New Glenn to cut launch costs dramatically. Even if those vehicles reach their planned prices, plenty of other customers will be competing for launch slots.
Cowboy Space believes many serious projects are aiming at the mid-2030s. Starcloud is taking a more modest approach, focusing initially on processing information already generated in orbit. Google has reached a similar conclusion while investigating its Project Suncatcher orbital computing system.
Its researchers calculated that today’s launch economics make supplying power to computing equipment in orbit brutally expensive. Orbital power works out at roughly $14,700 per kilowatt each year, compared with between $570 and $3,000 for a US terrestrial data centre.
Google believes the equation becomes more interesting if launch prices fall to around $200 per kilogram. At that point, which Google considers possible by the mid-2030s, orbital power costs could drop to about $810 per kilowatt annually. That would put the energy side of the equation within striking distance of terrestrial facilities.
Getting electricity cheaply is only one part of the problem as cooling processors in orbit is awkward because a vacuum cannot carry heat away through convection. Satellites instead have to radiate heat into space using large surfaces, adding mass and complexity to the system. Radiation presents another headache. GPUs and AI accelerators designed for terrestrial data centres were not built to spend years being bombarded by energetic particles.
Then there is the uncomfortable fact that AI hardware becomes obsolete astonishingly quickly. A GPU launched after months of satellite construction and rocket scheduling could become technologically obsolete before its mission has paid for itself.
Repairing failures presents a similarly traditional space-industry problem. A technician cannot wander down an aisle, pull out a faulty accelerator and stick another one in unless someone first invents an extremely expensive orbital server engineer. Latency and bandwidth mean orbital data centres are unlikely initially to replace conventional cloud infrastructure either.
The more realistic early applications involve processing information already produced in space, including Earth observation, scientific instruments, defence systems and satellite communications. Sending raw data down to Earth for processing, then transmitting the results back up, wastes both time and bandwidth.
Google is already moving towards real hardware tests. Its Project Suncatcher programme is preparing to send TPUs into orbit to examine how commercial AI silicon behaves under radiation, vibration and thermal conditions encountered in space. SpaceX has much grander ambitions and potentially the launch infrastructure needed to make them credible. Its ability to combine Starlink-style satellite manufacturing with Starship would give it an advantage if launch prices fall far enough.
The financial world, naturally, has priced much of this possibility long before engineers have determined exactly how to build it economically.
