Orient Space Chain and DFH Satellite jointly unveiled the "Space String" compute constellation plan on September 25, aiming to build a schedulable computing infrastructure in orbit. The two companies describe it as China's first space-based computing infrastructure designed for both global and deep-space use.
The plan moves forward in three phases: G1 demonstration satellites, G2 standard satellites, and G3 flagship satellites. G1 is expected to launch in the fourth quarter of 2027; timelines and per-satellite compute specs for G2 and G3 have not been disclosed.
Two layers, more than a thousand satellites
The constellation is split into a business layer and a computing layer.
The business layer is planned to deploy more than 720 data satellites, also called inference satellites, which handle data acquisition and specific tasks — for instance, processing Earth-observation data on the spot. The computing layer is planned to deploy more than 360 compute satellites, also called training satellites, arranged in layers along a dawn-dusk orbit to support the business layer. A dawn-dusk orbit tracks the boundary between Earth's day and night side, so satellites there stay in sunlight almost continuously, giving them a steadier power supply than typical low-orbit satellites.
The two layers connect through inter-satellite laser links, letting business satellites call on the computing layer's resources as needed. Officials describe the goal as gradually building a "schedulable, on-demand" space computing service covering space-based data processing, model deployment, and coordinated compute resources.
In terms of division of labor, Orient Space Chain (Dongfang Xinglian) is responsible for overall design and operations, while DFH Satellite (Diwei Er) supplies AI capabilities and handles international market expansion.
How it compares to similar plans
Putting AI compute into orbit is not a new idea — several players have pursued it over the past two years, each with a different approach.
China's earlier mover is the "Three-Body Computing Constellation." In May 2025, Guoxing Aerospace and Zhejiang Lab launched 12 computing satellites in a single mission to form a network; according to the plan announced at the time, the full constellation is ultimately meant to reach 2,800 satellites. Its approach puts compute on every satellite and networks them together in orbit.
Overseas, SpaceX unveiled the design of its AI1 satellite this June — a single satellite with a 70-meter wingspan and 120 kilowatts of routine compute power, essentially treating one large satellite as a data center. Earlier still, Starcloud sent an Nvidia H100 chip into orbit for a demonstration in late 2025.
What sets "Space String" apart is splitting inference and training across two satellite types: a large number of satellites stay close to the data source, while a smaller number concentrate on supplying compute, coordinated over laser links. This layered approach means observation satellites don't need to carry heavy compute themselves, but it demands more bandwidth and stability from the inter-satellite links.
What hasn't been disclosed
For now, the plan is described only at the architecture and satellite-count level. The compute, power consumption, and cooling approach for individual compute satellites, along with the bandwidth of the inter-satellite laser links, have not been made public.
Cooling is an unavoidable challenge for space-based compute. Without air convection in a vacuum, heat generated by chips can only be radiated away through panels — the higher the compute, the larger the radiator needs to be, which increases a satellite's size and weight and ultimately its launch cost. Training satellites, which handle concentrated computing, will likely face this problem more acutely than inference satellites, and neither company has said how it plans to address it. The launch cadence and funding arrangements for more than a thousand satellites also remain undisclosed.
Going by the timeline, the G1 satellites due in the fourth quarter of 2027 are only demonstration units — still far from the planned constellation of over a thousand satellites. How large a model the demonstration satellites can run once in orbit, and how long they stay stable, will be among the first data points for judging whether this architecture can scale further.
Sources: IT Home, CocoLoop, joint release materials from Orient Space Chain and DFH Satellite; verified the planned counts of over 720 inference satellites and over 360 training satellites, and the G1 demonstration satellite launch window.