high machines
Impossible. But doable.
Orbital AI data centres. From Europe, for Europe.
Time to 1GW European Orbital Space AI Inference Factory (2026-09-18 + 10y)
European space AI enabling Europe’s AI future
A fleet of AI-hosting satellites. Fully under European control.
European independence
Compute Europe actually controls.
Workload, keys, and downlink stay under European jurisdiction. Without trusted access to hosted intelligence a whole economy can halt in milliseconds — it has happened before (US export-controlled AI access, June 2026). Space is how European models stay European.
The grid cannot host it
~35 GW demand. ~17 GW supply.
Europe’s data-centre load roughly doubles this decade; the grid cannot. FLAP-D queues run 7–10 years. Terrestrial energy and land permits are the bottleneck — in orbit we just add compute nodes with every launch. It is not about what is cheaper per watt, but what is possible.
Unlimited scale in space
One satellite. One node. Then more.
Continuous solar, heat radiated into deep space, no grid queue. ~7,000 satellites per GW. LEO for low-latency inference; Sun–Earth L1 when training needs uninterrupted power. Capacity grows with every launch.
Compute does not saturate
AI using AI using AI.
Growing adoption puts intelligence into every product, business procedure, security and defence aspect. Access to the best intelligence is a business advantage. Tokens are not shared — more capability means more GPUs on the bus, brutally energy-hungry. We go where energy is not a problem.
Planetary competition
USA and China are already building.
Starcloud, Blue Origin, Google on one side; ADA Space and Orbital Chenguang on the other. Starlink’s Starmind filing covers one million satellites. ASCEND was an early kick-off, but not enough. Europe’s tech stack still has gaps (GPU, super-heavy-lift) — pushing forward fills them.
Own the hardware
SaaS is a commodity. The GPU in orbit is not.
Software is reverse-engineered and prompted into existence; the SaaS era is dying. The scarce, earning asset is powered, cooled, sovereign compute that does not come down when the next chip lands. One satellite is one node that keeps earning.
Use cases
Defence
Secure processing and space-domain awareness that never leaves the spacecraft.
Earth observation
Edge AI on the sensor feed — inference at the satellite, not in a foreign region.
Industry
Overflow for European data centres when terrestrial power and permits run out.
Research
Large-scale training when L1 offers undisrupted solar for long-horizon compute.
The base concept
Starcloud-1's orbit — a 45.4° inclined low-earth ring carrying the first NVIDIA H100 in space, lit by the real sun so you can see where it is day and night. The orbits and the terminator are computed for the current time from our published element sets, alongside the lower and upper limits of the dusk-dawn sun-synchronous band.
Loading 3D view…
- Starcloud-1 — 502 km, 45.4°
- Dusk-dawn SSO 600 — 600 km, 97.8°, LTAN 18:00
- Dusk-dawn SSO 800 — 800 km, 98.6°, LTAN 18:00
Three phases
Phase 1
Demonstrator
European Endurosat FAME S bus and one AI GPU. Sun-synchronous orbit for continuous sunlight.
Phase 2
Space AI data centre
Full orbital build-out on ~150 kW NVLink72 compute nodes, still in SSO. Inter-satellite routing — and worldwide satellite internet, because the AI fleet needs global access anyway.
Phase 3
Around the sun
Build-out in solar orbit, focused on Sun–Earth L1.