Someday, but not soon. Chip execs don’t need to start designing for space just yet.
The merger of xAI and SpaceX, the efforts by some startups, and research by Google have raised the question of when there will be data centers in space.
The big advantage of space is the plentiful availability of solar energy, in a sun-synchronous orbit.
Fig. 1: A possible orbital data center of the future. Source: Wikipedia, Creative Commons.
SpaceX’s impressive Starlink uses space to deliver high-bandwidth wireless networking globally with thousands of satellites. Now SpaceX has filed an application with the Federal Communications Commission (FCC) to develop orbital data centers with up to a million satellites.
But there are a lot of challenges. The big ones include:
Let’s look at these four challenges. If you want more detail read Google’s paper, “Towards a future space-based, highly scalable AI infrastructure system design” and Starlink’s technology description.
Electricity is a major concern in data center ramps since demand is outstripping supply — in the USA at least. Electricity is 20% to 30% of total data center annual operating cost today.
Sun-synchronous orbits are near-polar orbits where satellites are in sunlight most of the time. Solar radiation in orbit is more intense than on Earth since it doesn’t have to go through the Earth’s atmosphere. So orbital solar energy is free, except for the cost of the solar panels.
To generate 1 kilowatt in space takes 5 to 7 m2 of high-efficiency (20% to 30%) solar panels.
An Nvidia GB200 NVL72 rack consumes 120 to 130 kW of power. That requires ~7200 m2 of solar panels. An NVL72 rack consists of about 45 1U “pizza boxes,” each about one-third of a square meter = 15 m2 if arranged flat instead of stacked. So there is a form factor challenge. It would be great if each “pizza box” of electronics had a solar panel on one side that generated as much power as the pizza box consumed. Instead, the solar panel area required is about 500x.
Another problem is heat dissipation. We think space is cold, so heat dissipation is not an issue. Space is very cold, but its vacuum is a near-perfect insulator, so heat generated by an object is not conducted into space. It can only be dissipated by radiation in the infrared part of the spectrum. This requires cooling systems that carry heat to radiators that are out of the sunlight, which is on the opposite side of the solar panel. With AI accelerators operating at 75° C maximum, 1 kW of power can be dissipated by a radiator of about 2.5 square meters.
Summarizing, 1 kW:
So, like Starlink, the electronics payload must be surrounded by a much larger area for power generation and heat dissipation.
This is why there won’t be a single gigawatt data center satellite, but instead hundreds to millions of smaller satellites that will interoperate.
A single Nvidia GB200 NVL72 rack (36 compute trays and 9 switch trays) weighs about 1500 kg and costs about $3 million. Google’s paper talks about launch costs reaching <$200/kg to LEO (low Earth orbit) by the mid-2030s. So in 10 years it will cost $300,000 to launch a single Nvidia GB200 NVL72 rack into space — just one-tenth of today’s cost. But now to launch this rack costs at least $3 million, about the same as the electronics.
Existing SpaceX Falcon9 rockets can carry about 17,500 kg into LEO, or about 12 Nvidia GB200 NVL72 racks worth of electronics. A gigawatt data center has about 7000+ of these racks requiring 500+ launches. SpaceX’s Starship will be able to carry about 5x more per launch.
Fig. 2: How is this going to fit or work in space? Source: Google
Data centers as configured now won’t work in space. They will need to be modularized, like Starlink, into “pizza boxes” that fit in the payload canister and can either unfold into single units or self-assemble into a single unit. Rendezvous Robotics, a new startup, is working on self-assembling dinner-plate size electronic modules. The problem for AI is that the AI power consumption is, as we’ve seen, 10x to 100x the power generation of the similar area of solar cells in space. So there will need to be a central “core” of compute and a 10x to 100x larger area for power generation and heat dissipation. Heat dissipation requires plumbing for liquids, so this will limit the size of what is feasible. The hypothetical structure in Figure 1 is probably too big for a single satellite. Yet Starcloud is talking about 4 km x 4 km solar arrays. As Oren Etzioni, the former CEO of the Allen Institute for AI and a professor emeritus of computer science at the University of Washington says, “Show me the calculations that suggest this is even feasible.”
All of the cabling now seen in a data center will not do well in the shock and Gs of a launch or fit with self-assembly. Starlink already uses lasers to communicate between satellites, beaming through free space rather than using fiber optics. This is what Google and SpaceX propose for larger swarms of interoperating data center satellites. At short distances the data transfer rates will be very high, like fiber optics in the data center, though maintaining alignment will be required.
The orbits of each of the satellites are highly predictable, so the algorithms required should be manageable, and Starlink is already doing it. A bigger challenge will be intercommunication within the satellite between “tiles” that self-assemble post launch – copper connections are already limited in distance to meters on Earth and fiber connections between tiles would require connectors that cut into scarce signal loss budgets for fiber optics. Laser communications in the vacuum of space without fiber is a possibility.
Radiation is much higher in space because of the absence of the Earth’s atmosphere, which protects us on Earth.
Google actually tested their V6e Trillium TPU compute tray and accompanying AMD CPU server tray in a 67 MeV (mega electron volt) proton beam to simulate the operating conditions in sun-synchronous orbit. They used 10 mm of aluminum to provide shielding.
Total ionizing dose (TID) radiation can, over time, cause device failures. Google’s measurements showed that for a five-year TID, the devices should continue to operate reliably. The first to fail were the HBM DRAM devices at 3x the dose expected over five years. Logic devices operated far longer.
The other failure mechanism due to space radiation involves single event effects (SEEs), which are large energetic particle strikes that cause a trail of electron-hole pairs along their strike path. These can cause bit flips in storage elements and possibly multiple storage elements. Again, the HBMs were the most affected due to a combination of the sensitivity of the DRAM storage element combined with the huge number of bits stored by HBMs in a TPU versus SRAM bits on the TPU. They found that for typical inference workloads, the observed rate of uncorrectable ECC (error correcting code) errors for the HBMs was approximately one per year or one per several million inferences. An error in an inference may not significantly affect the outcome since so much data goes into the answer. Google judges this error rate to be acceptable for inference, but training jobs, which are much bigger and longer, might not find this error rate acceptable.
Microsoft, at the AI Infra conference in September 2025, described the operation of its data center dedicated to training runs for OpenAI. The entire data center does one training run at a time for weeks and months. Their average mean-time-to-failure was about two hours, at which point it would take minutes to hours to pinpoint and then fix the problem. For inference, the data center would run hundreds of separate inferences at a time, but the point is that problems happen frequently, and humans have to fix them, or the data center would slowly become inoperable.
A lot of data center subsystems come with redundancy built in. For example, Nvidia’s NVL72 has 72 GPUs and 18 switches, but only 64 GPUs and 16 switches are needed at a time. The others can be swapped in when there is a failure.
In co-packaged optics (CPO), like in Nvidia’s Ethernet switch announced last year, the lasers are pluggable so they can be easily and quickly replaced if they fail.
But this level of human intervention will not be practical in space, and robotic systems are far from advanced enough to do the troubleshooting and repairs a human can.
The availability and cost of electricity are serious concerns for the huge ramp in data center capacity required on Earth. But the cost of electricity is 20% to 30% of the total cost, and ways are being found to circumvent capacity constraints, such as moving data centers to where energy is abundant. Even with “free” (after the cost of solar panels) orbital solar energy, an orbital data center will be much more expensive than terrestrial for a long time due to launch costs. Launch costs are coming down, but solar energy and battery costs are coming down at a faster rate. And there are serious challenges involved in re-designing AI for space, including fitting in launch vehicles, self-assembling, intercommunicating, and reliability.
Niche applications may utilize orbital data centers in the near future, but it will take much longer for the economics of space data centers to become more attractive than terrestrial.
 |
 |
 |
 |
 |
 |
 |
 |
Leave a Reply