Key takeaways

  • Google's Project Suncatcher design flies 81 TPU satellites in a 1 km cluster at about 650 km, where solar panels can be up to 8 times more productive than on Earth.
  • Google's Trillium TPU survived a radiation dose nearly three times what a five-year mission would deliver, but silent bit flips remain a problem for training.
  • Google's paper estimates launch would need to fall from about $3,600 to under $200 per kilogram, around 2035, to rival data center power costs on Earth.

Space data centers are satellites that carry AI chips instead of cameras or radios, powered by solar panels and linked to each other by lasers so a whole cluster works like one computer. The best documented design is Google's Project Suncatcher, laid out in a Google Research paper and blog post in November 2025. Its pitch is cheap, nearly constant solar power. Its open problems are heat, radiation, the links between satellites and the price of getting to orbit.

The idea stopped being only a paper this year. Google's first prototype satellite launched on October 1, 2026 and has since answered prompts from orbit. On November 19, Google announced that Blue Origin's New Glenn rocket will carry its first heavy-lift Suncatcher launch, targeted for the first quarter of 2029.

Why put AI compute in space

The case rests on electricity. AI data centers need enormous amounts of it, and on the ground that means grid connections, power plants and permits.

In orbit, the sun does the work. Google says a solar panel in the right orbit can be up to 8 times more productive than the same panel on Earth, and it produces power almost continuously, so the satellite needs far fewer batteries. Suncatcher's satellites would fly a dawn-dusk sun-synchronous orbit, a path that rides the line between day and night so the panels almost never fall into Earth's shadow.

Cooling is the second argument, though it is less simple than it sounds. Starcloud, an Nvidia-backed startup, treats deep space as a heat sink with no limit and says its design would save the water that cooling towers on Earth evaporate. Starcloud also projects energy costs 10 times lower than on land, launch included. That is the company's projection, not a measured result.

How Suncatcher works

Google's design is a tight cluster of small satellites rather than one giant station. The illustrative version in the paper has 81 satellites flying in a formation with a 1-kilometer radius at about 650 km altitude. Neighbors sit only 100 to 200 meters apart, and the gap swings between those numbers twice per orbit.

Each satellite carries TPUs, Google's own AI chips. A September 2026 Google explainer says future satellites will each carry dozens of them. In the Blue Origin announcement, Google says each satellite will hold about as much compute as one data center rack.

The satellites fly that close for a reason. Training a large model means chips constantly swapping data, and Google's chips on Earth do that over very fast cables. In space the cables become laser beams, and a laser link's received power falls with the square of the distance. Flying satellites a few hundred meters apart makes very high bandwidth possible with off-the-shelf telecom optics.

The engineering problems

Heat has nowhere to go

On Earth, fans and water carry heat away. In a vacuum there is no air, so the only way to shed heat is to radiate it as infrared light from a surface. Google says the team uses heat pipes to move heat from the chips to radiators, and tested the setup in a thermal vacuum chamber.

The paper gives no radiator sizes or temperatures, and the authors list thermal management for power-dense chips as one of the critical open challenges. This is the weak spot in the "space is cold" argument: space is cold, but it is also a near-perfect insulator.

Radiation flips bits

Above the atmosphere, chips are hit by charged particles. Google put its Trillium TPU, its sixth-generation cloud chip, in a 67 MeV proton beam to find out what happens.

The results were better than expected for long-term damage. The high-bandwidth memory, the most sensitive part, began to show irregularities only after a dose of 2 krad(Si), almost three times the 750 rad(Si) a shielded chip would absorb over a five-year mission. No chip failed outright from accumulated dose up to 15 krad(Si).

The short-term effects are the harder part. A single particle can silently flip a bit in a calculation. The paper measured about one such event per 17 rad in the chip's logic and on-chip memory. For a model answering one request per second, that works out to roughly one bad answer in 3 million. The authors say the effect on training still needs study, because a corrupted value in training can spread through the model. Host computers crashed about once every 400 to 450 rad, and failed chips cannot be swapped by a technician.

Commercial laser links between satellites run at 1 to 100 gigabits per second. A cluster of AI chips needs thousands of times more. The paper proposes 24 wavelengths through a 10 cm aperture for 9.6 terabits per second in both directions, and more by using several beams side by side. A bench demonstrator reached 800 gigabits per second each way with one pair of transceivers. Pointing has to hold to about a microradian, which Google compares to hitting a coin-sized target from miles away while both ends are moving.

Launch is still too expensive

Everything depends on what it costs to lift a kilogram. The paper cites about $3,600 per kilogram for a reusable Falcon 9 today. Using SpaceX's historical trend, prices fall about 20% each time the total mass launched doubles. Reaching under $200 per kilogram by around 2035 would take roughly 1,800 Starship launches, about 180 a year.

At $200 per kilogram, the paper estimates a satellite like Starlink's v2 mini would cost about $810 per kilowatt per year to launch and run, against $570 to $3,000 per kilowatt per year for electricity at US data centers. The authors call that roughly comparable, and the comparison leaves out the chips and the rest of the hardware. They say the projection is highly sensitive to its assumptions and note that Starship reuse is still unproven.

Who is building space data centers

  • Google launched its first Suncatcher prototype, built with Planet, on SpaceX's Transporter-18 rideshare on October 1, 2026. It ran Google's open Gemma 4 model and sent back a haiku about space. Two linked satellites are planned for 2027 to test the laser links, and the New Glenn launch is targeted for early 2029.
  • Starcloud says its Starcloud-1 satellite took an Nvidia H100 GPU to orbit in November 2025 and trained a small language model there. It has described a future 5-gigawatt orbital data center with solar and cooling panels about 4 km on each side.

How much to trust the idea

The physics works on paper, and the early hardware tests are encouraging. Google's chips survived a radiation dose well past a five-year mission, and a prototype is running a language model in orbit.

But the economics depend on a launch price that does not exist yet, the heat problem is unsolved at scale, and nobody has trained a large model across a cluster of satellites. Google itself calls the paper "not intended as a comprehensive economic feasibility study." The 2027 two-satellite test of the laser links is the next real data point.

Frequently asked questions

How are space data centers cooled?

There is no air in orbit, so heat can only leave as infrared radiation from radiator surfaces. Google moves heat from its TPUs to radiators with heat pipes and has tested the setup in a thermal vacuum chamber, but its paper lists thermal management as a critical open challenge.

What is Project Suncatcher?

Project Suncatcher is a Google research moonshot to run machine learning on clusters of solar-powered satellites carrying TPUs and linked by lasers. Google announced it in November 2025 and launched its first prototype satellite on October 1, 2026.

Which companies are building space data centers?

Google is testing Project Suncatcher with Planet and has booked a Blue Origin New Glenn launch targeted for the first quarter of 2029. Starcloud, backed by Nvidia, says it flew an Nvidia H100 GPU to orbit on Starcloud-1 in November 2025.

Do space data centers make economic sense?

Not yet. Google's paper estimates that at about $200 per kilogram to orbit, launch and operation could be roughly comparable to data center electricity costs, but launch costs about $3,600 per kilogram today and the estimate leaves out chip and hardware costs.

Sources

  1. Towards a future space-based, highly scalable AI infrastructure system designGoogle (arXiv)
  2. Exploring a space-based, scalable AI infrastructure system designGoogle Research
  3. Google partners with Blue Origin on Project Suncatcher heavy-lift launchGoogle
  4. Learn about Google's Project Suncatcher to put ML infrastructure in spaceGoogle
  5. Google's Project Suncatcher prototype satellite is in orbitGoogle
  6. Google receives response from Project Suncatcher prototype satellite in spaceGoogle
  7. How Starcloud Is Bringing Data Centers to Outer SpaceNVIDIA
  8. Starcloud-1Starcloud