Researchers see potential benefits, astronomical challenges for orbital data centers

Back in 2003, an obscure candidate in a Belgian election received far more votes than predicted — 4,096 more, to be precise. The country was already using electronic voting, and the value in that computer error proved to be significant as it’s equal to 2 to the 13th power. An examination of the voting machines found that, during the time of the election, a cosmic event had occurred, the burst of radiation flipping a single bit value and changing the entire result of the election data.

Such events are rare on Earth, but they’re more common in space, where cosmic rays and solar radiation from the sun are not filtered through the atmosphere. When it comes to the hot topic of orbital data centers — or data centers in space — bit flipping is just one of the many potential hurdles that experts say need to be overcome to make the potential benefits worthwhile.

First, the pros: If positioned along the horizon, satellites could have continuous access to full sunlight, maximizing power uptake to solar cells. They also avoid the community and environmental issues of disrupting neighborhoods on Earth. And if they’re used to provide edge computing on-orbit, they could reduce latency. 

One of the major public concerns about data centers on Earth is the large amounts of water required for cooling. In the cold vacuum of space, the environment can take care of that temperature regulation, though the balance is more complex than simply offsetting the heat of the graphics processing units (GPU). Unlike on Earth, there’s no weather to deal with, so solar panels are always exposed fully to the sun. That’s good for power generation, but not for thermoregulation.

Without the filter of the atmosphere, the sun is also quite hot — the Webb Space Telescope, for instance, required a particular thermal engineering design and is always oriented in a very particular way so as not to overheat. And that’s only the first of the hurdles.

“There’s a lot of challenges that I see that are not trivial that makes having data centers in space a very challenging engineering problem,” said Virginia Tech aerospace engineer Samantha Kenyon.

Perhaps the largest is the cost of sending items to space, especially when you consider the size and scale of what’s being discussed, specifically SpaceX’s announced plan of sending up to 1 million data center satellites into low-Earth orbit.

Assistant Professor and Marty and Anna Irvine AOE Faculty Fellow Brad Denby, who works at the intersection of computer systems and space systems, said each of these satellites would house a GPU rack about the size of a bookshelf, equivalent to the NVIDIA GB300 computing platform. Each satellite would consume roughly 150 kilowatts of power — equivalent to the entire International Space Station — and would require unfurling a solar array at least half a football field long, much larger than solar arrays on current Starlink satellites. A data center on Earth contains 12 to 24,000 of these bookshelf-sized GPU racks, meaning that achieving the same amount of computing power in space would require more satellites than the number in Starlink’s entire current fleet — currently over 9,000.

“The only way to get to those numbers would be for the larger starship to become operational, and hundreds of starships would need to be launching practically around the clock,” said Denby.

The current computing capacity of xAI across its three terrestrial data centers is still a full order of magnitude smaller than the million satellites it has discussed launching. That’s all before getting anything off of the planet.

“Never has it been cheaper to do something in space than to do something on the ground,” said Kenyon.

The cost-per-pound of launching anything into space is already staggering before you get into the cost of fixing anything that goes wrong. Unlike data centers on earth, one can’t simply deploy a technician over to deal with a faulty GPU.

“Right now, you can’t service satellites unless you send an astronaut,” said Kenyon, an endeavor she said has been done for the Hubble Space Telescope a few times but is “very rare and very risky. A servicing satellite could be sent to repair a satellite, but the technology required to do so is still under development."

Smaller concerns include the added atmospheric drag that such a large object experiences as well as its physical footprint, which makes it at least marginally more likely to get hit with space debris. There’s also the issue of latency, as it takes much longer to transmit information from space than from the ground. If Starlink was only going to be using these for its own network in low earth orbit, that proximity could alleviate the downlink bottleneck, but it would not do so for a network back on Earth.

And while it may seem like the smallest of all the problems, the radiation that causes bit flipping can be a hugely disruptive issue.

“There’s all these highly energetic particles that come from the sun, from other distant galaxies, from big, catastrophic events in astrophysics,” said Kenyon. “All those particles come to meet objects in space and can disrupt electronic activities.”

Even with advancements in error-correction software, the total ionizing dose — or the slow corruption of the semiconductor from radiation exposure — may well drastically lower the lifespan of these systems. Denby said these generations of GPUs haven’t been built with the kind of radiation hardening required and will need to be tested for radiation resilience. Even then, there may be some very quick loss of some percent of the units.

“Just because you have a computer that works on the ground doesn’t mean it’s going to work well in space,” said Kenyon.