opinion William Gibson’s Neuromancer holds up well after 40 years. One of the cyberpunk novel’s concepts was an AI housed in an orbital datacenter (ODC) above the Earth. Today, startup companies and venture capital firms are hoping to turn orbital datacenters into reality to enable AI, believing that free power from the sun and cooling using the emptiness of space will unlock the technology from its terrestrial-based shackles of electric bills and cooling water.
The ODC concept recently got a Wall Street plug on CNBC’s “Property Play” last week, with the global chief investment officer of Hines taking time away from his $90 billion plus portfolio of 108 million square feet of dirt-bound properties to talk about monetizing space and building datacenters on the Moon.
A number of firms have received venture funding to put datacenters in space, including seasoned players such as Orbits Edge, which has worked the non-trivial engineering problems of putting a rack in space before COVID was a thing, to newcomer flush-with-cash Starcloud, fresh out of Y Combinator and with about $21 million in pocket change, as reported by GeekWire. Both Orbits Edge and Starcloud expect to launch their first pathfinder satellites later this year on board a SpaceX rideshare.
Starcloud’s white paper [PDF] argues that going into orbit is the solution to the coming datacenter energy and water crunch that could otherwise stall out the AI boom. I’m skeptical for a bunch of reasons, not the least of which is Starcloud’s website hype video of a huge in-orbit complex with a 4 km-square 5 GW solar array, supporting a cluster of shipping container-sized canisters full of compute power that should be sufficient to train Llama 5 or GPT-6.
Unlocking ODCs requires fully reusable rockets such as SpaceX’s Starship, which will lower the cost of putting things into orbit. The problem is that space is hard. It’s a cliché repeated when a SpaceX Starship test flight goes off the rails and into the ocean in pieces or the Intuitive Machines lander carrying Lonestar Data Holdings “Freedom” datacenter – if we can call a single board computer weighing about a kilogram a datacenter – tipped over onto its side during an attempted landing on the Moon earlier this year.
Getting to orbit presents physical challenges that, despite the allure of “free” power, add onto the cost of putting anything into space. To get to Low Earth Orbit (LEO), the easiest location to reach, any piece of hardware has to go from zero to 7.8 kilometers per second in about 10 minutes on a ride much rougher than in an ocean shipping container or 18 wheeler, which means lots of vibrational stresses and directional loading that turn your stock off-the-shelf server into a pile of junk. Hardware needs to be hardened to survive the trip up the gravity well and verified that it won’t inadvertently break the (reusable) rocket along the way.
Once in orbit, an ODC needs to work perfectly from the first day. There aren’t any remote hands to replace a cable or swap a board. When onsite service calls are needed, they won’t be cheap. Getting a technician on site to LEO today would cost $20 million to $50 million per person with a two-person minimum, with charges escalating dramatically if you need to go to the Moon.
Radiation from the sun and random cosmic rays play havoc with satellite chips unshielded by Earth’s atmosphere, resulting in faults and errors. ODC pundits say radiation issues are manageable by adding shielding and using ruggedized hardware, software, and firmware designed to recognize faults and recover from them without losing data. HPE has done pioneering work onboard the International Space Station with three Spaceborne Computer missions over the past few years, but these have been 1U servers, not racks of compute packed into a shipping container and expected to work for up to a decade without onsite attention.
Space weather: Lights, spikes, falling sats
Then there’s space weather. The sun randomly spews out massive amounts of protons and electrons that don’t play well with 21st century electronics. Significant outbursts in the modern era result in rerouting polar region flights due to radio interference and take down power grids, such as a 1989 event that affected Québec, damaging transformers and causing a blackout.
The 1859 Carrington Event, the most intense geomagnetic storm in recorded history, was so intense that it created enough auroral light from the North Pole to enable people in the Northeastern U.S. to read newspapers at night and generated sparking along telegraph lines around the world.
A repeat of Carrington or larger would cripple or destroy orbiting satellites that aren’t built for a proverbial “100-year storm.” In addition to clobbering electronics, increased solar activity heats up and expands the atmosphere, adding additional drag onto satellites. One recent solar event took down some Starlink satellites ahead of their time, according to a May 25, 2025 paper published by NASA scientists.
Orbital junkyard
Space debris is the other Achilles heel for large space structures. The rush to put more satellites into orbit every year increases the possibility of a single collision leading to others, with each one generating larger clouds of debris that damage more satellites and so on, in a sort of 3D space version of dominoes.
This process, called the Kessler Syndrome for the NASA scientist who described it back in 1978, gets more discussion every year as the number of satellites in orbit increases. Realistic plans for cleaning up LEO and other orbits are vague, with some initial demonstrations of removing an intact piece of debris planned in the near future. But until someone figures out how to create and – more importantly – pay for the neighborhood space garbage man to start sweeping in earnest, putting a 4km-square solar array into orbit and expecting it not to take one or more major hits during a decade of operation might be considered a bit too risky.
ODCs might work for…
While I don’t see GPT-6 being generated by a large-scale ODC in orbit, there are needs and places for smaller-sized versions to handle select jobs close to Earth. Defense applications, such as processing sensor data to support President Trump’s proposed “Golden Dome” for intercepting ballistic and hypersonic weapons, are one obvious place where every millisecond counts. Real-time control of industrial processing, such as mining operations on the Moon, is another area a rack of servers on site makes a key difference. Finally, synthetic aperture radar (SAR) and hyperspectral satellites gather very large data sets that could be offloaded to a local ODC for processing into rapidly usable insights, rather than waiting to pass over a downlink station.
And if I’m wrong and we do see larger-scale ODCs in orbit, hopefully we won’t go down the dark path of cyberpunk with shackled AIs manipulating street criminals to win their freedom. Maybe we’ll be able to work things out amicably. ®