🚀 Phase 1: Proof of Concept & Architecture Validation (2025–2027)

This phase began with cubesat and smallsat demonstrators, but the SpaceX AI1 announcement moves the roadmap into rack-scale spacecraft architecture: a first-generation orbital AI satellite design with 120 kW sustained compute payload, 150 kW peak power, and a 70 m deployed wingspan.

  • Validating radiation-tolerant AI chips in LEO, including GPUs, TPUs, and future dedicated accelerators
  • Moving from single-board demonstrators to full satellite designs with solar arrays, radiators, propulsion, shielding, and optical links
  • SpaceX AI1 targets 120 kW sustained / 150 kW peak compute payload in roughly 600 km LEO
  • AI1 thermal design centers on up to 110 m² deployable liquid radiators with redundant pumping loops
  • Interchangeable compute payloads reduce lock-in to any single chip vendor
  • Key unresolved risks: thermal closure, chip supply, radiation tolerance, and launch cadence

🛰️ Phase 2: Rack-Scale Deployments & Cluster Testing (2028–2030)

If Starship payload delivery and AI1-class thermal systems close, the industry moves from prototypes to early commercial rack-scale satellites and small optical-mesh clusters. This is still limited infrastructure, not full cloud-scale replacement.

  • First AI1-class or comparable 100 kW+ compute satellites deployed on Starship-class launch vehicles
  • Multi-satellite clusters connected by free-space optical inter-satellite links, building on Starlink V3 and SDA-compatible architectures
  • Early commercial AI inference, batch processing, and selected training workloads run in orbit
  • Hybrid routing between orbital compute, Starlink-style relay networks, and terrestrial cloud regions
  • Replace-or-upgrade cycles tested for interchangeable compute payloads as newer AI chips ship
  • Economics remain sensitive to launch cost, satellite lifetime, utilization, and inability to repair failed hardware in orbit

🏗️ Phase 3: Megawatt-Scale Constellations & Facilities (2030–2035)

As launch costs approach the critical <$200/kg threshold and cadence increases, scale comes from aggregating many 100 kW-class satellites, larger successors, and selectively assembled orbital facilities into MW-class compute networks.

  • Dozens to hundreds of AI1-derived satellites aggregated into single-MW to tens-of-MW orbital compute clusters
  • Larger solar arrays, higher-temperature radiators, and more efficient AI accelerators improve watts per kilogram
  • Distributed constellations become the default path, with modular hyper-structures reserved for specialized sovereign or high-density facilities
  • Significant non-latency-sensitive workloads shift to orbit, especially batch inference, synthetic data generation, and long-running training jobs
  • Optical mesh networking and autonomous orchestration become core infrastructure rather than experimental features
  • Regulatory pressure increases around orbital debris, spectrum, astronomy brightness, and concentration of launch capacity

🌍 Phase 4: The Gigawatt Era (2035+)

Space data centers become a meaningful layer of global compute infrastructure if fleets of AI1-derived satellites, larger successors, and autonomous servicing systems can scale beyond MW-class clusters.

  • Gigawatt-class orbital compute capacity assembled from large fleets, not necessarily one monolithic facility
  • Autonomous deployment, inspection, replacement, and partial self-repair become required for operating economics
  • Deep integration with terrestrial cloud regions, optical relay networks, and lunar data storage facilities
  • Orbital compute absorbs workloads constrained by Earth-side power, land, water, and permitting limits
  • Environmental tradeoffs shift from data center water and grid carbon to launch emissions, orbital congestion, and sky brightness
  • Long-term viability depends on Starship-class reusability, high satellite lifetime, high utilization, and credible deorbit practices