NodalZen

Blog · News

Data centers in space, the state of play

Where orbital data centers stand in late 2026, what has actually flown, what is only planned, and the hard problems of heat, radiation and launch cost.

Putting AI computers in orbit sounds like science fiction, but by late 2026 real chips are running in space. This post separates what has happened from what is only planned, and the problems still in the way.

Why anyone wants to do this

The pitch rests on two ideas. First, sunlight in space is strong and, in the right orbit, almost constant. Google's researchers wrote that a solar panel in the right orbit can be up to eight times more productive than on Earth, and can produce power nearly all the time, so it needs fewer batteries (Google Research, November 2025). Second, space has no neighbors waiting in a grid connection queue and no local water supply to share.

What has actually flown

Starcloud-1. Starcloud, a startup based in Redmond, Washington, launched its first satellite in November 2025 carrying an Nvidia H100, a data-center-class GPU (Starcloud). Nvidia described the satellite as weighing 60 kilograms, about the size of a small fridge (Nvidia, October 2025). Starcloud says that in December 2025 the satellite ran an open Google language model in orbit and trained a small model called nanoGPT, which it describes as the first time a language model was trained in space (Starcloud).

Google's Project Suncatcher prototype. Project Suncatcher is a Google research effort exploring fleets of solar-powered satellites carrying Google's TPU chips and linked by laser. On October 1, 2026, Google said its first prototype satellite, built with Planet, had launched on SpaceX's Transporter-18 rideshare mission and was operating as expected. Over the following weeks the team planned to collect data on how the chips cope with the stress of launch, radiation and temperature extremes (Google, October 2026).

What is planned or proposed

  • More Suncatcher satellites. When it announced the project in November 2025, Google said two prototype satellites built with Planet were slated to launch by early 2027 to test the chips and the laser links between satellites (Google Research, November 2025).
  • Starcloud-2. In March 2026 Starcloud announced a $170 million funding round at a $1.1 billion valuation and said Starcloud-2 would launch later in 2026 to run commercial workloads (Via Satellite, March 2026). We could not confirm from the company that it has launched as of this writing. The same report says Starcloud has filed with the US Federal Communications Commission (FCC) to deploy up to 88,000 satellites.
  • SpaceX. On February 4, 2026, the FCC accepted for filing an application from SpaceX for a system of up to one million satellites operating as orbital data centers, at altitudes from 500 to 2,000 kilometers (Benton Institute, February 2026). Accepting a filing is the start of a review, not an approval.

Hard problem one: heat

On Earth, data centers dump heat into air or water. In space there is no air, so the only way to get rid of heat is to radiate it away as infrared light, the same invisible glow you feel from a warm stove. That needs large panels called radiators, and the more computing you add, the bigger they must be. Google lists thermal management among the open challenges (Google Research, November 2025). Nvidia's description of Starcloud's long-term plan for a 5-gigawatt orbital data center mentions solar and cooling panels about 4 kilometers wide and 4 kilometers long (Nvidia, October 2025).

Hard problem two: radiation

Above the atmosphere, chips are hit by fast-moving particles that can flip bits of memory or slowly wear the chip out. Google tested its Trillium TPU in a particle beam. The most sensitive part, the high-bandwidth memory, showed problems only after a dose nearly three times what the team expects over a shielded five-year mission, and there were no hard failures up to a much higher dose (Google Research, November 2025). The accompanying paper reports that the chips survived a dose equal to a five-year mission without permanent failures (arXiv, November 2025).

Hard problem three: launch cost

Every kilogram has to be lifted into orbit. Google's analysis suggests launch prices could fall below about $200 per kilogram by the mid-2030s, and that at that level, the launch and running cost of a space data center could be roughly comparable to the energy cost of an equivalent one on the ground (Google Research, November 2025). That is a forecast made in 2025, not a price anyone pays today. Starcloud has projected that energy in space could be 10 times cheaper than on land, including launch, according to Nvidia (Nvidia, October 2025). Both are company projections.

Other open questions

Google also lists the need for very fast laser links between satellites, keeping tightly packed satellite formations under control, fast links to the ground, and reliability when nobody can visit to swap a broken part (Google Research, November 2025).

The bottom line

As of November 2026, a handful of data-center-class chips are running in orbit as experiments, and large constellations exist on paper. Whether orbit becomes a real home for AI computing depends on radiators, radiation tolerance and launch prices, none of which is settled yet.

Try it in NodalZen

The Orbit tool lets you plan a cluster in space and see how solar power, radiators and radiation shape it. Compare the result with the same cluster on the ground in Size.

Sources

Learn the basics Open the planner