Orbital inference starts selling: Yusuan Xinghe ships in October
In-orbit AI stopped being a demonstration this week. One company put a launch date and a price list on it, and a second put the hardware on orbit to make it work.
A Guangzhou startup says its first compute satellites launch in October, and customers have already paid in advance for the models they will run. Yusuan Xinghe, incubated by listed smart-transport firm PCI Tech (Jiadu Technology 600728), used the second Guangdong AI Application Matchmaking Conference in Guangzhou on Tuesday to unveil the Pazhou Intelligent Compute Constellation — a set of satellites built to keep a large model running seven days a week in orbit rather than shipping raw imagery back to a ground station. Founder and CEO Wang Kai, a Tsinghua–Caltech PhD and formerly PCI Tech's chief AI scientist, says the company will procure and own the satellites itself: 24 of them, forming what it calls an "on-orbit AI lab."
The two things the first batch is meant to prove are the two nobody has proved yet. Whether a large model can run continuously for weeks in orbit, and whether several satellites can split the work across laser inter-satellite links instead of each running a copy alone — distributed inference with a 500-kilometre gap in the middle. The commercial pitch is deliberately ordinary: token APIs for platforms, and a chat interface for transport, insurance and finance customers who want to ask questions about the ground rather than download pictures of it. Wang told 36Kr the team already runs a multimodal model on orbit aboard ADA Space hardware, which it claims was the first commercial mission completed by a model deployed on a space compute satellite. Treat "first" as the company's framing; the October launch is what will settle it.
The hardware case landed two days earlier: a Chinese remote-sensing satellite that reads its own images before sending anything home. Chaozhisuan-1 rode a Lijian-1 Y18 rocket out of the Dongfeng commercial space zone on September 20 as part of a nine-satellite rideshare, carrying a four-metre visible-light camera and an AI compute payload. Its operator, Chaozhisuan, says the satellite performs image parsing and target recognition on orbit and downlinks only the high-value results, cutting response time from the hours a conventional downlink-then-process chain takes to minutes. The same launch carried Pengcheng Nanke-1, a 150-kilogram test satellite from Galaxy Space, Peng Cheng Laboratory and SUSTech that packs a 5G non-terrestrial base station, a 100 Gbps laser link and an onboard compute payload.
The reference point for scale is the constellation ZJLab launched in May 2025 — 12 satellites at up to 744 TOPS each, about 5 PFLOPS in total. One satellite processing its own pixels does not change the economics of orbital compute; a network that never has to downlink the raw frame does.
iFlytek's industrial arm shipped version 3.5 of its Lingyang industrial model with three agents aimed at the power grid. Announced at the World Manufacturing Conference in Hefei on September 21, the release bundles compute-power coordination, power trading and zero-carbon campus agents — the model already spans wind, solar, hydro, thermal, nuclear and storage. The company says the upgrade consolidates scattered vertical capabilities into agents that plan and execute on their own, and claims zero-shot and few-shot generalisation across the industrial time-series data that makes up more than 70% of plant telemetry. It is the least glamorous of the three stories and the one most likely to touch a working engineer: the model's job is to move energy scheduling from a human reacting to a meter reading to software deciding before the reading arrives.
What to watch: whether Yusuan Xinghe's October batch actually flies and then runs unattended for weeks, and whether the "results only" downlink model produces customers who renew rather than pilots who demo.
If a satellite can answer a question about the ground without sending you the picture, what does the ground-station business become? Tell us in the comments.
