Data Centers in Space: Beyond the Hype, a Real Engineering Challenge

October 22, 2026

Innovation posts, News

Data Centers in Space

For the past year, orbital data centers have been making headlines. Most announcements remain projections. But on October 1, Google launched the first experimental satellite for its Suncatcher project, and its researchers published their design in the journal Joule [1, 2].

Why the Idea Is Appealing

On Earth, building AI data centers faces obstacles related to grid connections, water supply, and public acceptance. Many projects are delayed or abandoned. In orbit, on a dawn-dusk sun-synchronous orbit, solar panels receive nearly continuous sunlight: Google estimates up to eight times more energy generated per panel per year than on Earth¹.

The race is underway. Starcloud placed an H100 GPU in orbit in November 2025⁴. SpaceX is seeking authorization for up to one million satellites⁵. Blue Origin, Axiom, and Aetherflux are also advancing their ambitions.La course est lancée. Starcloud a placé un GPU H100 en orbite en novembre 20254. SpaceX demande une licence pour jusqu’à un million de satellites5. Blue Origin, Axiom et Aetherflux ont aussi des ambitions qui se concrétisent.

Why 81 Small Satellites Instead of One Large One?

Rather than a giant satellite, Suncatcher proposes a cluster of small satellites flying in formation. Its reference scenario involves 81 satellites at an altitude of approximately 650 km, spaced 100–200 m apart and connected by lasers¹. No performance results have been published yet. A two-satellite test to validate the optical links is planned for early 2027⁶.

Google’s choice reflects an engineering rationale: distributing computing power across multiple satellites instead of concentrating it in a single structure. This modular approach allows heat to be managed at the individual-unit level, modules to fit within existing launch-vehicle fairings, and a failed satellite to be replaced without losing the entire cluster.

But that flexibility comes at a cost: the satellites must maintain precise formation and reliable optical communications. That is Suncatcher’s challenge: demonstrating that a distributed architecture can operate efficiently in orbit.

Four Major Obstacles

1. Heat. In a vacuum, heat can only be dissipated through radiation, requiring large radiator surfaces and therefore additional mass. Google’s paper acknowledges this as a major challenge but does not quantify radiator area or mass¹. An IEEE Spectrum analysis concludes that operating a GPU in orbit would cost at least ten times as much as on the ground, even under optimistic assumptions⁷. Estimates of thermal-system mass vary by a factor of ten across sources⁸.

2. Launch costs. Google’s economic case assumes launch costs below approximately $200/kg by 2035, compared with around $3,600/kg today on Falcon 9. Under its own model, this would require sustaining roughly 180 Starship flights per year¹. And the comparison concerns energy costs, not the total cost of a data center.

3. Reliability. In orbit, radiation poses a significant challenge for electronic components. While total radiation dose can be managed through component selection and shielding, energetic particles can cause transient errors, including bit flips. Google plans to mitigate their effects through hardware redundancy¹ and software-based error correction. The challenge remains ensuring long-term system reliability without the possibility of direct repairs.

4. Orbital congestion. Nearly all projects target the same dawn-dusk orbit. Applications explicitly filed for this orbit total approximately 111,000 satellites⁹, in a band estimated to represent around 1% of usable low-Earth-orbit volume¹⁰. There are currently about 16,000 active satellites⁹. Google estimates that terawatt-scale computing capacity could fit in this orbital band¹, a claim challenged by 33FG’s analysis¹⁰. SpaceX argues that its autonomous collision-avoidance maneuvers make this density manageable¹¹, but coordination remains an open question. With 81 satellites, Suncatcher is the most modest of the current projects.

Key Takeaways

– Short term: Orbital computing is most compelling for niche applications, such as preprocessing Earth-observation imagery before transmitting it to the ground.

– Long term: A general-purpose, gigawatt-scale data center remains an aspiration for around 2035 or later, contingent on launch costs that have yet to be demonstrated.

– What to watch: Prototype thermal data and error rates, the 2027 test, actual Starship launch pricing, and orbital regulation.

The authors themselves conclude that AI in space is promising but faces serious engineering challenges¹. Google’s main contribution is to put numbers, assumptions, and limitations on the table. That is what separates an R&D project from an investor presentation.

References

  1. Agüera y Arcas, B., Beals, T., Biggs, M., Bloom, J. V., Fischbacher, T., Gromov, K., Köster, U., Pravahan, R., Manyika, J., « Toward a future space-based, highly scalable AI infrastructure system design », Joule 11, 102678 (2027).
  2. Google Research, « Exploring a space-based, scalable AI infrastructure system design ».
  3. Google, « Project Suncatcher: facts ».
  4. DCD, « Starcloud-1 satellite reaches space with Nvidia H100 GPU now operating in orbit ».
  5. Brookings, « Orbital data centers’ feasibility gap is a governance risk » (juin 2026).
  6. Pulse 2.0, « Google launches Project Suncatcher prototype satellite to test AI infrastructure in space ».
  7. IEEE Spectrum, « Why Thermodynamics Rules Future Orbital Data Centers ».
  8. Forum économique mondial,« Why cooling is the real obstacle to space-based data centres » (juin 2026).
  9. Secure World Foundation, « To ODC or not to ODC » (2026).
  10. 33FG, « Where should space data centers live? ».
  11. Forethought, « Will we really put data centers in space? ».

Further Reading