Skip to content
DopeSwagYolo

AI in Space

Data Centers in Space: How They Would Work and Why Critics Doubt Them

Orbital data centers would run AI chips on near-constant sunlight. As of October 2026, Google and Starcloud have flown only small test satellites, and critics cite cooling, radiation and launch cost.

By DopeSwagYolo4 min read

Researched and fact-checked by AI, with no human review. 16 sources listed below. How we verify

A data center in space would be a satellite, or a tight cluster of satellites, carrying AI chips. The chips would run on solar power and shed heat through radiator panels. They would send results to the ground by laser or radio.

As of early October 2026, the leading U.S. proposals amount to a few small test satellites and several very large regulatory filings. Outside analysts argue that cooling, radiation, launch cost and latency make the idea far harder than its backers suggest.

How would a data center in space work?

Google's Project Suncatcher is one detailed public design. It calls for solar-powered satellites carrying the company's TPUs (its AI chips) in a dawn-dusk sun-synchronous orbit. That orbit is a path along the boundary between day and night that keeps the panels almost always lit. Google's research team says panels there can yield up to 8 times as much energy as on the ground. Because one satellite holds only so much hardware, Google modeled 81 satellites flying within a 1-kilometer radius, linked by lasers.

Backers expect inference to come first, TechCrunch reported. Inference means answering queries with a trained model. The reason, the outlet reported, is that training needs thousands of chips coordinated across many satellites.

An orbital data center would be a satellite, or a tight cluster of satellites, running AI chips on solar power and shedding heat through radiator panels.

Why are data centers in space considered a good idea?

Near-constant sunlight reduces the need for batteries. Backers also present orbit as a way around the power and water demands of server farms. Blue Origin told the Federal Communications Commission (FCC) in March 2026 that its plan would ease pressure on U.S. communities and natural resources, TechCrunch reported.

Google's analysis suggests launch prices could fall to about $200 per kilogram by the mid-2030s. At that point, Google says, launching and operating an orbital data center could cost roughly as much as the energy bill of a comparable one on the ground. That assumes SpaceX's Starship flies about 180 times a year for a decade, TechCrunch reported.

What has actually flown, and what is only a filing?

The best-documented U.S. test hardware is small:

  • Starcloud-1 launched in November 2025 carrying an Nvidia H100 graphics processing unit (GPU). The startup says it trained an AI model in orbit. A second GPU failed during launch, TechCrunch reported.
  • Google and Planet's Suncatcher prototype launched on October 1, 2026, on SpaceX's Transporter-18 rideshare mission. Payload reported that it carries four TPUs on about 1 kilowatt of solar power. It also reported that the chips must power down every 15 to 20 minutes to avoid overheating. Google plans two follow-up satellites in 2027.

Outside the United States, China's ADA Space and Zhejiang Lab launched the first 12 satellites of a planned 2,800-satellite computing constellation in May 2025, SpaceNews reported.

The filings are vastly larger. SpaceX asked the FCC in January 2026 to authorize up to 1 million data center satellites. Starcloud applied for up to 88,000, an application the FCC accepted for filing on March 13, 2026. Blue Origin's Project Sunrise proposes more than 50,000. Acceptance for filing means an application passed initial review, not that it was approved.

Why do critics say data centers in space are a bad idea?

Cooling. A vacuum has no air to carry heat away, so it must be radiated. In IEEE Spectrum, ABI Research analyst Andrew Cavalier calculated how much radiator is needed at 60 degrees Celsius:

  • One 700-watt H100 chip needs 1.4 square meters of radiator.
  • A 40-kilowatt rack needs about 80 square meters.
  • A 100-megawatt facility needs at least 2,500 radiators of that size.

The magazine's computing editor has said Starcloud's radiator was too weak to run its chip at full power.

Radiation. Charged particles can corrupt stored data or destroy circuits, Cavalier wrote. Google says proton-beam tests of its TPUs produced no hard failures from accumulated dose. It says memory showed irregularities only at nearly three times the dose expected over five years.

Launch cost. A reusable Falcon 9 costs roughly $3,600 per kilogram to orbit, according to TechCrunch. That is about 18 times the $200 target, TechCrunch reported. Cavalier's model found that launching and running a GPU in orbit for a year costs at least an order of magnitude more than on Earth. He found that holds even at a Starship price he calls highly optimistic.

Latency and repairs. Data sent to low Earth orbit takes tens of milliseconds to return, IEEE Spectrum reported. Amit Verma is an electrical engineering professor at Texas A&M University-Kingsville. He told the magazine that chatbots can tolerate such delays but real-time stock trading cannot. He also said large constellations raise failure risk with limited repair options.

Environment. A solar-powered orbital data center could emit an order of magnitude more than one on Earth, according to Saarland University researchers cited by Scientific American. Their estimate counts rocket launches and reentering hardware.

The bottom line

Orbital data centers have moved from concept papers to first experiments. But the hardware Starcloud and Google have flown amounts to a handful of chips, and Google's must pause regularly to stay cool. The larger plans depend on launch prices that rockets do not yet offer, on radiators and chips that survive years in orbit, and on regulatory approval. Data from Google's prototype and the 2027 flights planned by Google and by Starcloud are the next tests.

Sources

More from AI in Space

See all in AI in Space