Adaspace将AI送上卫星，IPO后计划出售太空算力

Editor’s note: This article is republished from Jumpstart, a TechNode Group publication. At an algorithms competition in Guangzhou, a team working on traffic analysis sent its model somewhere unusual: to a satellite already in orbit. Jiazhihuihang, part of Chinese transport technology company PCI Technology, worked with Adaspace to run the model in space. The satellite could process imagery before sending the analysis back to Earth. If the job is to extract information about a road network, there is less reason to transmit every pixel before getting started. For Adaspace, the experiment pointed to a different kind of business. A company that builds satellites for customers could also give them somewhere to run their software. The Chengdu company expects its shares to begin trading in Hong Kong on November 13. Its IPO will help fund the engineering needed to take that idea beyond individual demonstrations.

From a Payload to a Constellation

Adaspace began in 2018 with a smaller engineering problem: how to make AI work aboard a satellite in the first place. Its early work centred on payloads, the equipment a satellite carries to perform its mission. In December that year, it launched its first AI application satellite. The early satellites used onboard computing to help with their own work. For an Earth observation satellite, that might mean analysing some of the data it collects before sending anything back. The next step was to design satellites that could take on computing tasks beyond their own operations. That changed what Adaspace had to build. Running more demanding software requires more electrical power and generates more heat, which the spacecraft has to dissipate. Putting several such satellites into orbit also makes the shape of the spacecraft a practical concern: a flat design lets multiple units stack together for launch. The route to more computing power in space ran through changes to the satellite itself. In May 2025, Adaspace launched 12 computing satellites into a single orbital plane. A second group of eight followed in October 2026. Between those launches, the company reported that it had deployed Alibaba’s Qwen3 model on a satellite already in orbit. That test illustrated what makes the hardware more useful once launched: engineers on the ground can give it new software and ask it to do something different.

What Customers Buy Today

Adaspace has built a business around both the spacecraft and the work that can be done with satellite data. On one side, it designs and tests satellites for customers, coordinates launches with third parties and can help operate the spacecraft after delivery. On the other, it uses satellite data to build tools for customers on the ground. One example in the prospectus is a system that helps a digital content company turn satellite data into landscapes for games and film. The balance between those businesses has changed. Adaspace reported that satellites and related services accounted for 57.7% of revenue in the first half of 2026, up from 0.6% in 2023. By October 20, it had customer orders for 114 AI satellites still to be delivered, with a combined contract value of RMB3.48 billion. What counts as delivery helps explain how this business works. A satellite can be complete in the factory and still be months away from customer acceptance. Depending on the contract, part of the revenue may be recognised only after the spacecraft is launched, tested in orbit and handed over. A launch delay can therefore push revenue into a later reporting period even when the hardware is ready. Selling access to computing would give Adaspace a different relationship with customers. The traffic team brought a model and a task. A service built around that kind of use would need to handle the next task, and the one after that, on a schedule customers can plan around.

What Hong Kong Will Fund

Under its Star-Compute plan, Adaspace wants to link satellites so that they can share computing capacity, storage and software. The company aims to launch 100 satellites under the plan by the end of 2028. The much larger figure of 2,800 refers to the satellites covered by the spectrum and orbital resources supporting the plan. The proposed use of IPO proceeds shows where the work is now. Adaspace plans to allocate about 36% to research and development and another 24% to a Shenzhen centre for satellite design, development and testing. Those investments serve the business it already has—building and delivering spacecraft—as well as the network it wants to operate. The traffic experiment showed why a customer might send a model to space in the first place. Adaspace’s next step is to make it easy enough for that customer to come back with another job.