Why this class exists
Orbit Trades Grid Power for Launched Kilograms
AI data centers are short of power, cooling and permits on the ground; in a dawn-dusk orbit they have continuous sunlight and deep space as a heat sink.
The case for orbital compute is power without a grid connection and cooling without water. The price is mass: every watt needs solar array to make it and radiator to throw it away, and a radiator's output depends only on its own temperature. SpaceX's AI1 design rejects about 1.4 kW per square meter and reaches about 70 kW of compute per tonne. At that ratio the whole business is a bet on launch cost, which is why it appears on the Heavy Lift edge list. Projected
Flight hardware exists and is small. China's Three-Body constellation has flown 12 laser-linked computing satellites since May 2025. Starcloud flew one NVIDIA H100 in November 2025, and Google's first Suncatcher satellite launched on 1 October 2026. SpaceX merged with xAI in February 2026, filed for up to 1 million satellites, and told investors in its S-1 that orbital data centers may not achieve commercial viability. One industry estimate puts orbital compute at about three times terrestrial cost per watt. Building
Western programs
SpaceX Filed for a Million Satellites; Starcloud and Google Are Flying First
Prototypes in 2025-2027, constellations promised after Starship cadence arrives.
| Program · company | Scale | Status | Rated |
|---|---|---|---|
| Starmind · SpaceX (with xAI) | Up to 1,000,000 satellites filed | FCC filing January 2026 for 500-2,000 km shells; prototypes due early 2027; Gigasat factory in Bastrop County, Texas. The S-1 warns the business may not be commercially viable | Building |
| Starcloud-1 · Starcloud | 60 kg · 1 NVIDIA H100 | Launched 2 November 2025; trained NanoGPT and ran Gemma in orbit in December 2025 | Observed |
| Starcloud-2 · Starcloud | Blackwell GPU · large deployable radiator | Due October 2026; Starcloud has filed for up to 88,000 satellites and raised $250M at $2.3B in August 2026 | Building |
| Project Suncatcher M1 · Google with Planet | 4 Trillium TPUs · ~1 kW | Launched 1 October 2026 on Transporter-18; contact confirmed, no radiation data released yet | Observed |
| Orbital data center nodes · Axiom | 2 nodes | Reached orbit 11 January 2026 aboard Kepler optical relay satellites; ISS node due 2027 | Observed |
| Galactic Brain · Aetherflux | First node | Announced December 2025; first node targeted Q1 2027 | Projected |
| TeraWave · Blue Origin | ~5,400 satellites | Data-center networking constellation filed January 2026 | Projected |
| StarVault · Lonestar | Data storage | First payload on Sidus LizzieSat-4, autumn 2026 | Projected |
Chinese programs
China Has the Only Laser-Linked Computing Constellation in Orbit
ADA Space, Zhejiang Lab, Guoxing Aerospace and CASC are each building toward thousands of computing satellites.
| Program · operator | Scale | Status | Rated |
|---|---|---|---|
| Three-Body Computing Constellation · ADA Space and Zhejiang Lab | 12 satellites · ~5 POPS | Launched May 2025: up to 744 TOPS each, 30 TB storage, 100 Gbps laser links held eight days at 99.99%; 100 satellites targeted by 2027 | Observed |
| Star Computing · Guoxing Aerospace | 2,800 by 2030 targeted | Ran the Qwen3 model in orbit in January 2026 | Observed |
| Chenguang-1 · Orbital Chenguang | Technology test | Launched 24 July 2026 on Lijian-1 toward a gigawatt-class orbital data center | Observed |
| CASC space computing | Gigawatt scale | Planned within the 15th Five-Year Plan; Chinese ITU filings total about 200,000 satellites | Projected |
Design figures
Radiator Area, Not Chips, Sizes Every Design
Published figures for the designs that matter.
| Design | Power | Thermal and mass | Rated |
|---|---|---|---|
| AI1 · SpaceX | 250 kW peak · 175 kW average | 160 m² radiator at ~1.4 kW/m² 70 m wingspan; one NVIDIA Vera Rubin NVL72 rack; ~70 kW per tonne | Projected |
| Starcloud-3 | 200 kW | 3 t, sized for Starship; viable at $500 per kg launch | Projected |
| Suncatcher cluster · Google | 81 satellites | 1 km radius at 650 km; economics assume $200 per kg launch around 2035 | Projected |
| Three-Body satellite | 744 TOPS | Flight hardware; laser-linked | Observed |
No flown GPU or TPU has published single-event-upset data yet. Commercial accelerators are flying unhardened, and the radiation record from Starcloud-1, Suncatcher M1 and the Three-Body satellites will decide how much shielding mass every later design carries.
Where it appears
Orbital Compute Sits on Top of Launch, Relay and Chips
Every edge is a dependency the business cannot control.
| Depends on | Relationship | Breaks if | Rated |
|---|---|---|---|
| Heavy Lift · Starship | Array, radiator and rack mass to orbit | Launch cost stops falling below ~$500 per kg | Building |
| Comms & PNT | Starlink and laser backbones carry workloads to the ground | Ground-station or spectrum limits | Observed |
| Terafab · Tesla, SpaceX, Intel | Chip supply for Starmind; Grimes County, Texas | The framework agreement stays non-binding | Projected |
| The Moon | Radiator and array area as a lunar export | Earth-launched area stays cheaper | Modeled |
Contested
Deliberately Not Claimed
The load-bearing uncertainties, tagged with the confidence they actually earn.
The largest player calls the business unproven. SpaceX's S-1 risk section says orbital data centers may not achieve commercial viability. Filings for a million satellites are spectrum claims, not production plans.
Radiation tolerance of commercial accelerators is unmeasured in orbit. The first data returns from 2026 flights will move every power-per-tonne figure on this page.
Cost parity depends on launch near $200-500 per kg. Google's own model assumes $200 per kg around 2035. Falcon 9 lists near $4,250 today.
Record
Provenance and Data Sources
- Confidence
- Observed for flown, fielded or measured - Building for hardware and programs in development - Projected for announced plans - Modeled for original analysis
- Status date
- Q4 2026
- SpaceX
- FCC filing for up to 1,000,000 satellites (January 2026); SpaceX-xAI merger (2 February 2026); S-1 (April 2026); AI1 design release (June 2026)
- Others
- Starcloud flight and funding announcements; Google Suncatcher M1 launch (1 October 2026); Axiom node launch (11 January 2026); Aetherflux; Blue Origin TeraWave filing
- China
- Zhejiang Lab and ADA Space Three-Body releases; Guoxing Aerospace; Orbital Chenguang launch (24 July 2026)
- Weakest link
- No published radiation data for flown commercial accelerators; launch cost assumptions