40 Satellites, One Console: Beijing Unveils the First Space-to-Ground AI Computing Cloud

On October 30, at a "Tiansuan Jiyuan" space computing themed event in Beijing, the Beijing Space Computing Innovation Center launched what organizers describe as the world's first space-ground integrated AI computing cloud platform. According to coverage by People's Daily Beijing, Beijing News and Securities Times, the platform has completed its core functionality and now pools computing resources from more than 40 satellites with ground CPUs, GPUs and NPUs behind a single management and scheduling interface.

The pitch is to make in-orbit compute satellites manageable the way ground servers are. The platform stitches together an end-to-end pipeline — satellite-ground resource aggregation, model upload and publishing, in-orbit inference, and result return — so that an operator gets one entry point for resources, while a model provider gets a one-stop deployment entry supporting container-image uploads, inference API testing and version iteration, with full logging across six stages from task creation to result return.

Architecturally, the platform is organized into five layers: compute infrastructure, resource management and scheduling, inference acceleration, model services, and token operation. Organizers highlight four capabilities: visualized operations showing fleet size, orbital status and in-orbit duration; pooled management of heterogeneous space and ground compute; coordinated scheduling across ground centers, satellites and orbital planes, with preemption based on orbital passes to cope with intermittent links; and one-click model publishing with a closed loop from deployment to inference to result return.

The platform is already interoperable with Baidu Intelligent Cloud's Baike platform, whose model and dataset interfaces it can call for data processing and in-orbit inference. First-batch members include Xingce Weilai, Yiwei Aerospace, CAS Satellite Group, CAS Xiguang Satellite, Yusuan Xinghe, Chaozhisuan and Xinghuo Chuanming, contributing satellites such as BUPT-2 and BUPT-3, AIRSAT-05/08/14/15, several Xiguang imaging satellites, the Pazhou intelligent computing constellation 01 and 02, Chaozhisuan-1 and Yuxing-2 09.

The event also featured two supporting moves: a list of ten space-computing achievements for 2026 spanning onboard compute chips and payloads, thermal and power systems, communications and computing constellations; and a joint announcement with the China Academy of Information and Communications Technology of a public service platform for the space computing application ecosystem, aimed at standards development, testing and industry matchmaking. Wang Shangguang, the center's director, a Beihang- and BUPT-affiliated professor and founder of the Tiansuan constellation, delivered the keynote.

Academic commenters framed the effort as an early step in star-ground collaboration. Tsinghua professor Cui Yong argued compute satellites can filter data in orbit to save downlink bandwidth while ground centers handle training and elastic scaling — "compute nearby in space, with the ground as backstop." Beihang's Wei Xinguo said unified scheduling could accelerate applications such as space situational awareness. All performance and capability claims come from the organizers, and no independent benchmarking exists yet. The platform also competes for attention with very different visions, from Google's Suncatcher orbital prototype to conventional ground clouds — whether token-based billing makes sense for satellites that duck in and out of sunlight remains the open question.