At a seminar in Guangzhou on 27 January 2026, executives from Chengdu-based GuoXing Aerospace Technology announced that a general-purpose large language model — Alibaba’s Qwen3 — had been uplinked to and run aboard their inaugural space-based computing satellites. Company leaders said the model handled end-to-end reasoning in orbit: queries sent from ground stations were processed on board and answers returned to Earth in roughly two minutes, part of a live trial the firm described as a first for a general-purpose, large-scale AI model operating on an orbital constellation.
The GuoXing Qwen3 trial
GuoXing presented the demonstration as a technical milestone: the startup said it had transmitted Qwen3 to one of its earlier launches and executed multiple experiments in which payloads performed inference and returned results directly to ground. Company executives emphasized the latency and autonomy gains of processing in space rather than moving raw remote-sensing data down to terrestrial datacentres for analysis. The firm reported round-trip processing times on the order of minutes and framed the trial as a proof point for many of the use cases being proposed for orbital AI — from faster disaster assessment to near-real-time analytics for maritime and agricultural monitoring.
Ambitious constellation plans
GuoXing is not alone in setting big targets. The company outlined a roadmap that scales from its initial 12-satellite cluster launched in May 2025 to a planned network of 2,800 specialized satellites by 2035. That architecture, as described, would consist of roughly 2,400 inference satellites and 400 heavier training platforms in sun-synchronous, dawn–dusk and low-inclination orbits between about 500 and 1,000 kilometres. GuoXing’s public figures include extremely large aggregate compute goals — aiming at orders of 100,000 petaflops for inference and up to 1 million petaflops for training across the eventual constellation — and a phased deployments schedule that targets a 1,000-satellite capability by 2030.
Other Chinese projects and early services
Several other Chinese teams and start-ups have been moving quickly in this space. ADA Space, which launched the first 'space computing' tranche in mid‑2025, has publicised a follow-on 12-satellite cluster it calls Liangxi and says its Star Compute service has delivered on-orbit inference for the Aerospace Information Research Institute (AIRI) of the Chinese Academy of Sciences. Research labs and smaller commercial players have also demonstrated experimental systems: Zhejiang Laboratory has flown a 12-satellite mini-constellation carrying an eight-billion-parameter model for domain‑specific tasks, while firms such as Zhongke Tiansuan report long-running on-orbit operations with earlier space computers and are testing home‑grown high-performance GPUs for future satellites.
Architecture and engineering challenges
Pushing AI into space is not only a systems-integration exercise but also an engineering challenge on several fronts. Satellites must survive a high-radiation, wide‑temperature environment while running power-hungry accelerators. Chinese teams describe mitigations including redundant electronic architectures, error-detection and recovery protocols to tackle radiation-induced faults, and novel thermal management approaches such as fluid-loop heat transport to move waste heat to radiators where it can be dumped by radiation. High-throughput inter-satellite connectivity is a parallel bottleneck: firms are developing laser-based links to shuttle large data volumes between satellites and concentrate compute where it is needed, with company briefings referencing multi‑hundred‑gigabit links for mesh-like constellations.
Why operators want compute in orbit
The pitch for orbital AI rests on several practical advantages. Low Earth orbit constellations sit where much of the raw data originates — from Earth-observation imagers, maritime trackers and other sensors — so on‑board processing can collapse petabyte streams into compact, actionable outputs before any downlink. That reduces the need for costly, high‑bandwidth ground infrastructure and can cut latency for time-sensitive decisions. Operators also point to plentiful solar energy and a cold background for radiative cooling as environmental benefits relative to land-based datacentres; however, those advantages are contingent on solving the engineering hurdles described above and on the economics of launch, replacement and maintenance.
Strategic context: a global race
Announcements from Chinese firms have entered a wider international conversation: private-sector and national actors in the United States and Europe are pursuing related ideas, and Silicon Valley executives have publicly sketched plans that include space-based compute augmentation. Security analysts and space policy observers point out that the same hardware and optical networking that enable commercial AI services in orbit can also be dual-use, raising questions about resilience, national security and export control regimes. Commentators in multiple countries are already treating space AI as a strategic frontier — both because of its commercial potential and because a distributed, space-based compute layer would alter where and how critical national infrastructure and industrial AI workloads are hosted.
Near-term timetable and what to watch
GuoXing said the second and third clusters of its constellation will roll out this year, and company spokespeople continue to publish aggressive multi‑year timetables. Independent start-ups and research labs across China have similarly rapid roadmaps: second‑generation payloads, experimental inter-satellite laser tests, and early customer deployments have already been announced or demonstrated. International watchers will be watching a few concrete indicators closely — the scale and cadence of launches, successful long-duration operation of GPU‑class accelerators in orbit, the reliability of laser crosslinks, and whether in-orbit training at scale becomes technically and economically viable. Equally important will be governments’ policy responses on spectrum, export controls and norms for responsible behaviour in orbit as commercial AI platforms and national security interests converge.
Space-based AI is moving from concept to demonstrator to operational service in months rather than years. That compressed timetable raises familiar trade-offs: speed and scale versus reliability and openness. For now, the headline is straightforward — companies in China say they have run a general-purpose large language model in orbit, and they are backing that demonstration with plans for thousands of compute-optimised satellites. Whether those plans become a transformative new layer of global computing will depend on engineering follow-through, launch economics and the policy choices nations make as commercial ambition meets sensitive national interests.
Sources
- Chinese Academy of Sciences (Aerospace Information Research Institute)
- Institute of Computing Technology, Chinese Academy of Sciences
- Zhejiang Laboratory
- Beijing Astro‑future Institute of Space Technology (BAIST)
- China National Space Administration
