Guoxing Aerospace (国星宇航) has disclosed detailed plans for what it calls the world’s first space compute network built specifically to serve “silicon‑based intelligent agents” — autonomous systems that run on conventional semiconductor hardware. At a Beijing seminar on space computing the company laid out a roadmap for a 2,800‑satellite constellation, split between 2,400 inference satellites and 400 training satellites, and said it has already placed an initial cluster into orbit for validation.
The proposed network would operate in low Earth orbit (500–1,000 km), using dawn‑dusk and sun‑synchronous tracks plus low‑inclination orbits, linked by star‑to‑ground and inter‑satellite laser links. Guoxing says the architecture will provide global coverage across land, sea, air and space and deliver low‑latency, high‑reliability on‑orbit computing for applications such as autonomous vehicles, drones and intelligent robots. The firm is explicit about scale: a target of 100,000 P‑level inference capacity and 1,000,000 P‑level training capacity, where Guoxing defines 1P as 1,000 trillion (10^15) operations per second.
Guoxing’s statement frames the move as a shift from the conventional model of routing data down to terrestrial data centres and back up to satellites, to one where computation happens in orbit — a transition the company calls going from “ground compute” to “space compute”. It argues that placing compute in orbit can overcome ground constraints, cut data‑transport costs for distributed, mobile systems and enable real‑time decisioning for widely dispersed agents.
The programme is already underway. Guoxing says the first ‘space compute centre’ (Group 01) launched in May 2025 and completed key technical tests. Two further centres have been manufactured and are scheduled for orbital deployment in 2026. In a milestone it plans to load a large foundation model, Tongyi Qwen3 (通义千问 Qwen3), into the Group 01 centre in November 2025 — a move the company describes as the first ever deployment of a general‑purpose large model to run on an in‑orbit platform.
Guoxing is one of several Chinese private actors racing to assemble large low‑orbit constellations. Zhejiang’s Zhijiang Laboratory leads a separate “three‑body computing constellation” aiming at a thousand satellites and dispatched 12 compute satellites in May 2025. Other private firms are assembling communications and IoT constellations and applying for spectrum: one group has sought ITU rights for up to 10,000 satellites. Guoxing itself says it holds ITU approvals for 3,145 orbital frequency slots and has built and launched dozens of satellites and payloads since 2018.
Technical and commercial hurdles are substantial. High‑performance training in space demands power, thermal management and radiation‑hardened hardware at scales far beyond legacy satellite payloads. Laser inter‑satellite links and in‑orbit model updates must operate reliably through weather, orbital dynamics and contested electromagnetic environments. Cost economics — launching, replacing and operating thousands of compute‑heavy satellites — will be a central test of viability.
Strategic and regulatory implications are just as weighty. An orbiting compute layer that serves civilian silicon‑based agents could also be repurposed for military or intelligence use, intensifying dual‑use concerns and helping explain the keen interest of national actors. The scale and density of new constellations raise questions about space traffic management, orbital debris and spectrum allocation, and will likely spur international debate over norms governing compute in space.
Guoxing’s plan exemplifies an emerging global trend: moving edge compute closer to the physical endpoints that need instant decisions. If successful, space‑based compute could unlock new capabilities for maritime autonomy, beyond‑line‑of‑sight aerial systems, disaster response and remote sensing analytics. Yet success is not guaranteed; the combination of engineering difficulty, capital intensity and geopolitical sensitivity will decide whether orbiting superclusters become a practical complement to terrestrial cloud and edge infrastructure, or remain an experimental niche.
For now, Guoxing’s disclosures signal that China’s commercial space sector is pushing beyond communications and imaging into an integrated compute layer in orbit. That pivot—if mirrored elsewhere—would reshape assumptions about where and how AI workloads run, and create a new frontier for competition over space infrastructure and standards.