Chinese tech firms race to build AI computing capabilities in space

BEIJING -- A simple query to a remote-sensing satellite yields an intelligible summary such as "a slightly parched wheat field in need of watering," generated through fully autonomous analysis in space, offering a clear glimpse of the power of a Chinese in-orbit AI system.

As AI fuels an insatiable appetite for computing power, a new field is emerging in the technological race as intelligent computing capabilities are pushed into space. In early November, SpaceX rocket orbited the Starcloud-1 satellite mounted with Nvidia GPUs, thrusting space-based computing into the spotlight.

Chinese tech firms have entered the race early, betting that the final frontier can overcome the Earth-bound limits of energy, space and cooling that increasingly constrain the growth of AI.

The latest to join the competition is Zhongke Tiansuan (Comospace), founded in 2024. It has already logged over 1,000 days of on-orbit operation with its Aurora 1000 space computer aboard a Jilin-1 satellite, and its next-gen Aurora 5000, featuring a home-grown high-performance GPU, is set for in-orbit trials next year. They are pushing a project to build a "space supercomputer" in low Earth orbit.

For Chinese tech entrepreneurs, space-based computing offers a clear edge as terrestrial AI hits power, data and land bottlenecks.

"Orbital edge computing moves AI directly to the source of data filtering petabytes of daily satellite imagery and traffic before the narrow downlink chokes," Liu Yaoqi, CEO of Zhongke Tiansuan, told Xinhua.

Benefits also include the global reach of low-orbit constellations and sunlight that powers computing facilities almost for free, Liu added.

BROADER PLANS

Tiansuan's cosmic computation quest mirrors the country's ambitious blueprint.

Last week, Beijing municipal authorities unveiled a plan to construct a massive, centralized space data center in a dawn-dusk orbit 700 to 800 kilometers above Earth. The project, led by an innovation consortium, targets a system with a power capacity exceeding one gigawatt.

The first tech-demo satellite, Chenguang-1, is slated for lift-off in late 2025 or early 2026. The computing power of the experimental satellite is roughly equivalent to a single ground server.

"Its scale is modest, but we are taking this first small step," said Zhang Shancong, president of Beijing Astro-future Institute of Space Technology (BAIST), which leads the project.

Deployment is expected to roll out in three waves, culminating by 2035 in a megawatt-scale orbital data center whose computing capacity is expected to surpass that of China's entire current ground-based fleet, according to the BAIST.

In eastern China, Hangzhou-based Zhejiang Laboratory has already put up a 12-sat mini computing constellation dubbed "Three-Body," carrying an 8-billion-parameter space-borne AI model. Two of the satellites carried X-ray polarimeters that pool their on-board compute power to catch fleeting gamma-ray bursts in real time.

Once its planned 1,000-plus satellites are in orbit, the constellation will crunch 100 quintillion operations per second, according to the lab.

"With a computing constellation, part of the data can be processed in space and delivered straight to users," said Li Chao from Zhejiang Laboratory.

OVERCOMING HURDLES

To knit together its scattered orbital computers, China is already testing inter-satellite laser links that deliver high-speed, rock-solid, low-latency chatter. Beijing-based startup Laser Starcom is among the firms crafting the cutting-edge terminals.

On its pair of Guangchuan-01/02 satellites lifted off by a Zhuque-2 rocket last November, the company has closed a 400 Gbps laser link in low-Earth orbit.

"Laser links are the bedrock that breaks the communication bottleneck and lets space-based computing fly," said Wu Shaojun, Laser Starcom's founder.

The challenges of space computing also include maintaining stable operation in extreme environments with high radiation, as well as dissipating heat from components in the vacuum of space, where there is no air convection, according to Liu from Zhongke Tiansuan.

The Tiansuan team managed to address radiation-induced failures such as computational errors and system crashes in industrial-grade chips by implementing redundant designs, error correction and recovery protocols. They are also experimenting with fluid-loop cooling in orbit to transfer heat from high heat-flux chips.

In Liu's view, the roadmap for space computing begins with intelligent remote sensing to overcome data and bandwidth bottlenecks. It then moves to smart communications, boosted by large satellite networks, to increase capacity and reduce latency.

The evolution culminates in in-orbit AI, providing real-time computational support for terrestrial services, such as perception for autonomous driving, drone traffic management, cross-border logistics scheduling and maritime navigation assistance, Liu said.

In his vision, the future fisheries management may feature a "Fish Finder" app. This tool would integrate real-time satellite imagery, environmental data, AIS signals, and on-orbit AI processing to guide fishing vessels directly to "where the largest schools are located."