Orbital Inc. emerged from stealth in mid‑April and announced plans to run AI inference from a mesh of small satellites in low Earth orbit, arguing that terrestrial power limits make space the next frontier for scalable compute. "There simply isn’t enough capacity here [on Earth], and the only way is up," founder and CEO Euwyn Poon told reporters, and the startup says its system could shift some inference demand off terrestrial grids and toward abundant orbital solar energy.
The company’s hardware concept centers on fridge‑sized satellites, each housing a GPU server rack, solar panels roughly the size of a tennis court, and radiative cooling panels of comparable area. Orbital’s long‑term target is as many as 10,000 satellites delivering roughly 100 kilowatts apiece. To test the approach, the firm plans a prototype launch in 2027 aboard a SpaceX Falcon 9 to validate GPU operations in orbit and run commercial inference workloads.
Orbital is explicitly focused on inference rather than model training and intends to distribute work across independent GPU nodes rather than tightly coupled clusters. In the proposed architecture, a user request would route from an Earth data center to a ground station, be relayed to a satellite, passed among satellites via optical interlinks to an available GPU for processing, and then returned through the same chain; ground stations would only contact satellites within line‑of‑sight.
The team acknowledges substantial engineering challenges. Every watt collected by the solar arrays becomes heat that must be shed through large radiative coolers because the absence of atmosphere forces thermal design to rely on radiation to space. Radiation in low Earth orbit can induce bit flips and accelerate hardware degradation in GPUs, and routine physical maintenance will be far more difficult and costly than for terrestrial data centers — constraints that shape choices around cooling area, shielding, redundancy, and error mitigation.
Orbital joins a small field exploring space data centers: Starcloud has already run a similar test, and proposals such as SpaceX’s AI Sat Mini signal industry interest. Orbital says its differentiator is a deliberately distributed, inference‑focused fleet and an effort to match small satellites to inference workloads to reduce per‑unit launch and deployment cost. The startup plans to court "big model labs" such as OpenAI and Anthropic via direct API token access and enterprise deals that could move inference traffic into its orbital network.
For builders and operators the trade‑offs are clear. Moving inference into orbit could offload peak electricity demand and tap abundant solar input, but it imposes limits on continuous availability, enforces latency windows tied to ground‑station passes, requires optical‑link scheduling, and adds the overhead of radiation hardening and large radiator sizing. Orbital still must validate practical GPU reliability in orbit and demonstrate whether distributed inference across optical interlinks and intermittent ground contacts can meet commercial service requirements.
Sources
Replies (0)
No replies in this topic yet.
