Orbital data centres
Also known as Space data centres, Orbital compute, Space-based data centres
Orbital data centres are proposed computers in orbit, powered by sunlight and cooled by radiators that face deep space. A few small flight demonstrations have been reported, but no independently verified commercial orbital data centre has yet shown sustained service or a competitive cost.

In plain language
An orbital data centre is a computer placed on a satellite. Sunlight can power it, and the cold of space can help carry away its waste heat, while laser links can connect it to other satellites and to the ground. The phrase covers several different ideas. A satellite that processes its own pictures before sending them down is one. A large cluster of satellites that sells cloud computing to customers on Earth is a much harder and more speculative project.
Most of the evidence so far is small. A commercial computer flew to the International Space Station and ran a short image-analysis task, with remote help from a ground team. Companies have reported launching graphics processors and running small AI tasks in orbit, and Google has launched a test satellite carrying its own AI chips. None of these shows that a full data centre in orbit can run reliably for years, or that it can compete with ordinary data centres on cost.
Technical explanation
Power is the first constraint. Solar arrays can supply electricity in suitable orbits, but the array, batteries, power conversion and pointing must be sized for a continuous load, including the periods when the satellite is in Earth's shadow. In vacuum there is no air to carry heat away. Heat leaves a spacecraft by conduction to radiators, which emit it as infrared radiation towards deep space. Radiator area, emissivity, temperature and the heat arriving from sunlight and Earth all set how much heat can be rejected. Every watt a processor uses eventually becomes heat that the spacecraft must reject.
Radiation and reliability are the second constraint. Processors in orbit can suffer single-event upsets, latch-up and cumulative total ionising dose. Google reports that its Trillium TPU design survived a total ionising dose equivalent to a five-year mission without permanent failure, under the conditions it tested, and that bit-flip errors were also characterised. A component test does not qualify a whole system, however, because memory, interconnects, power supplies, storage, optics and software must also survive. Optical inter-satellite links can move data between spacecraft without a ground relay, but acquisition, pointing, handover and ground gateways still limit end-to-end service.
Economics form the third constraint. A fair comparison counts delivered compute rather than peak accelerator specification, and includes launch, manufacture, utilisation after eclipses and outages, replacement cadence, ground systems and disposal. Two 2026 arXiv preprints model this from different angles. Turyshev's system model gives 34 to 59 kg per kW for a 1 MW high-sunlight case once fixed spacecraft mass is included. Van Berkel's network-focused model finds low Earth orbit inference may be feasible, but training the largest models in orbit is unlikely to be competitive with terrestrial facilities. Both are preprints built on stated assumptions that still need testing against measured service.
What the evidence says
Each statement is typed by the strength of its evidence and linked to the source that supports it.
Demonstrated Shown in a published experiment, test or shipping product.
In vacuum there is no convection. A spacecraft's heat leaves through thermal radiation from its radiators, and the rate depends on radiator area, infrared emissivity and the temperature difference to deep space.
Sources: [5] · 7.0 Thermal ControlReported, unverified Stated by a company or team but not independently confirmed.
AWS reports that a Snowcone edge device ran a short machine-learning image-analysis workload aboard the International Space Station, with 2 CPUs, 4 GB of memory and 14 TB of storage, and that a ground team diagnosed a file-extension mismatch and uploaded a revised model during the mission.
Sources: [1] · Company mission account · medium confidenceStarcloud reports that Starcloud-1 launched in November 2025 carrying an NVIDIA H100 GPU, which it describes as the first H100 in orbit.
Sources: [2] · medium confidenceStarcloud reports that Starcloud-1 became the first satellite to run a version of Gemini in orbit, and the first to train a small nanoGPT model in orbit, both in December 2025. The page gives no independent benchmark, availability figures or training comparison.
Sources: [2] · low confidenceGoogle reports that a Project Suncatcher prototype built with Planet launched on Transporter-18, that contact was confirmed and the spacecraft was operating as expected, and that TPU tests for spaceflight, radiation and thermal conditions were to follow.
Sources: [3]In a company-authored preprint, Google reports that its Trillium TPU design survived a total ionising dose equivalent to a five-year mission without permanent failure, and that bit-flip errors were characterised. The abstract gives no bit-flip results.
Sources: [4] · Abstract · medium confidenceProposed A design or plan that has not yet been built or tested.
Google's preprint projects that launch cost to low Earth orbit may reach US$200 per kilogram or less by the mid-2030s, based on a learning-curve analysis.
Sources: [4] · Abstract · low confidenceTuryshev's model gives, for a 1 MW high-sunlight case, a base case of about 5,640 square metres of beginning-of-life solar array and 2,500 square metres of radiator, and 34 to 59 kg per kW once fixed spacecraft mass is included. These are model outputs, not flown hardware.
Sources: [6] · medium confidenceVan Berkel's preprint finds that running AI inference in low Earth orbit may be feasible, but that training frontier-scale language models in orbit is unlikely to be competitive with terrestrial data centres under its modelled assumptions.
Sources: [7] · Abstract · low confidenceLimitations and open problems
- Most flight evidence is company-reported. The sources reviewed give no independent benchmark, uptime record or delivered-service data for any orbital compute node.
- The 2022 Snowcone workload relied on a ground team that connected remotely to diagnose and reconfigure the device. It does not show autonomous operation of a spacecraft node.
- The Starcloud-1 page gives no performance metrics, availability figures or training comparison, so its milestones cannot be set against terrestrial hardware.
- Google's radiation result covers tested components under stated conditions. It does not show that a complete orbital system survives five years.
- Heat-rejection needs depend on heat load, radiator temperature, emissivity and end-of-life coating degradation. The sources reviewed give no general figure for data-centre heat rejection in orbit.
Possible future uses
Ideas, not predictions. None of these has been demonstrated.
- A sovereign or defence compute service in which customers accept a high cost per compute-hour for resilience and independence from terrestrial infrastructure.
- Processing placed near optical relay nodes, so that several spacecraft can share data and return only selected results.
- Megawatt-class constellations that sell general cloud computing to terrestrial customers. This depends on launch costs, utilisation, lifetime and communications all improving together.
Timeline
- 21 Jul 2022AWS publishes an account of a Snowcone edge device running an image-analysis model aboard the International Space Station. Sources: [1]
- Nov 2025Starcloud reports that Starcloud-1 launched carrying an NVIDIA H100 GPU. Sources: [2]
- 22 Nov 2025Google Research posts a company-authored preprint proposing a space-based AI infrastructure design. Sources: [4]
- 29 Apr 2026Turyshev submits a system-level model of orbital data centre constraints and economics to arXiv. Sources: [6]
- 15 Jul 2026Van Berkel submits a model of space-based AI compute network limits to arXiv. Sources: [7]
- 1 Oct 2026Google reports, on 1 October 2026, that its Project Suncatcher prototype satellite launched and that contact was confirmed, with TPU spaceflight tests planned. Sources: [3]
Futures this could help bring about
Glossary
- Radiator
- A surface that gives off heat as infrared radiation. In orbit, radiators are the main way a spacecraft gets rid of waste heat.
- Single-event upset
- A change of state in a memory cell or circuit caused by a single charged particle from space.
- Total ionising dose
- The cumulative radiation energy absorbed by a component over its life, which can cause gradual permanent damage.
- Sun-synchronous orbit
- An orbit that keeps a fixed angle to the Sun, so a satellite passes over a given place at about the same local time each day.
Sources
- company announcement
How We Sent an AWS Snowcone into Orbit
Amazon Web Services, 21 Jul 2022. Accessed 21 Oct 2026. - company announcement
Starcloud-1 mission page
Starcloud, Dec 2025. Accessed 21 Oct 2026. - company announcement
Project Suncatcher prototype satellite update
Google, 1 Oct 2026. Accessed 21 Oct 2026. - preprint
Towards a future space-based, highly scalable AI infrastructure system design
Blaise Agüera y Arcas, Travis Beals, Maria Biggs et al.. arXiv, 22 Nov 2025. Accessed 21 Oct 2026. - government report
7.0 Thermal Control (State-of-the-Art of Small Spacecraft Technology)
NASA SmallSat Institute, 7 May 2026. Accessed 21 Oct 2026. - preprint
Orbital Data Centers: Spacecraft Constraints and Economic Viability
Slava G. Turyshev. arXiv, 29 Apr 2026. Accessed 21 Oct 2026. - preprint
The Cost and Network Limits of Space-Based AI Compute
Kees van Berkel. arXiv, 15 Jul 2026. Accessed 21 Oct 2026.
Revision history
3 changes since this profile was created.
What it took to get here
The inventions this technology was built on, oldest foundations at the top. Lines show what each one required.
- 2022Orbital data centresResearch and prototypesRequires: Space-based solar power, Reusable launch vehicles. Being researched.