--- title: "China Is Putting AI Compute in Orbit. The Business Is Still Satellites" date: 2026-11-03 category: China AI Decoded site: NeuroAI canonical: https://neuroai.site/a/na-decode-orbital-compute-order-book language: en --- # China Is Putting AI Compute in Orbit. The Business Is Still Satellites > Shanghai wants 3,000 satellites by 2030, but the Hong Kong-bound contender earns thin hardware margins selling AI satellites, not tokens from orbit. The two spacecraft that left Jiuquan in mid-October were not sent up to photograph the ground and radio the pictures home. They were sent up to look at the pictures and transmit the conclusion. That is the argument for space-based compute (天基算力): a satellite that can run a model stops being a camera with a radio and becomes a server with a view. The argument is coherent. The revenue has not caught up. ## Key takeaways - A six-department Shanghai plan issued on 26 October 2026 targets 3,000-plus satellites in orbit and RMB 80 billion (≈US$11.3bn / HK$88bn) of core industry revenue by 2030. - ADASpace held orders for 114 undelivered AI satellites worth RMB 3.475 billion (≈US$489m) as of 20 October 2026, per its Hong Kong prospectus. - AI compute satellites produced 32.0% of ADASpace's 2025 revenue at a gross margin of 7.6% — hardware economics, not service economics. - ADASpace's top five customers were 91.1% of first-half 2026 revenue, which makes delivery timing, not demand, the swing factor (prospectus). - ADASpace's own 星算 (Star Compute) plan calls for 2,800 AI satellites by 2035; the order book meant to fund it covers 114. ## What Shanghai actually committed to On 26 October 2026, six Shanghai departments issued an action plan for the satellite-internet industry covering 2027 to 2030. The targets, restated by Xinhua at a Songjiang conference the next day: more than 3,000 satellites in orbit, core industry revenue above RMB 80 billion (≈US$11.3bn / HK$88bn) and more than 100 demonstration applications by 2030 (Shanghai Municipal Commission of Economy and Informatization, 26 October 2026). Read the units. A satellite-internet plan counts satellites and industrial revenue. It does not count tokens, floating-point throughput or utilisation — the three numbers China's terrestrial compute policy uses. Nothing in the published targets says how much inference is meant to run in orbit, or who buys it. A satellite target is a launch target. Reading it as demand for orbital compute repeats the error of reading construction targets as utilisation. ## The company at the front of the queue To see what the sector sells today, read the filings of the company closest to a listing. ADASpace (国星宇航), a Chengdu satellite maker founded in 2018, passed the Hong Kong exchange's listing hearing on 29 October 2026 and filed with China's securities regulator the next day. Its prospectus states that as of 20 October 2026 it held orders for 114 undelivered AI satellites worth RMB 3.475 billion (≈US$489m / HK$3.8bn) — roughly RMB 30.5 million (≈US$4.3m) per unit, a price for a manufactured object rather than a rate for a service. The revenue mix agrees. AI compute satellites brought in RMB 225 million (≈US$31.7m) in 2025, up 189% on 2024 and equal to 32.0% of group revenue, at a 7.6% gross margin. Group revenue was RMB 703 million against a net loss of RMB 256 million; losses from 2023 to 2025 exceed RMB 572 million. A 7.6% margin is a contract manufacturer's margin, not an operator's — and the valuation argument depends on becoming the second thing. ## Why orbit, and what it actually fixes The physical case is real but narrower than the marketing. A satellite in low orbit has sunlight without land, water or a grid, and radiates heat into space; what it lacks is power per kilogram and a fast way to get data down. The bottleneck on-board inference addresses is downlink. A remote-sensing satellite collects far more imagery than its radio link can return; most is discarded or delayed. Running a model on board and sending a result instead of a scene compresses the product to what a narrow link can carry. The commercial version is small. On 14 October 2026, two compute satellites built by ADASpace for Yusuan Xinghe (宇算星河), a Guangzhou company incubated with backing from the listed transport-technology firm PCI (佳都科技), reached orbit from Jiuquan. According to the company, they run remote-sensing models on board and return results as tokens rather than imagery, with customers paying for analysis (China Securities Journal, 2 November 2026). Three constraints decide whether that becomes a business: - **Power and heat.** Every watt spent on inference is a watt not spent on instruments, and waste heat needs radiator area that adds mass. - **Downlink and ground stations.** Results still need a radio window and a licensed ground segment; orbital compute moves the bottleneck, it does not remove it. - **Launch cadence.** Revenue arrives when a rocket does. ADASpace lost two experimental satellites in a January 2026 failure and says it plans at least 14 more satellites in orbit this quarter. ## The gap that matters ADASpace's stated plan is 2,800 AI satellites by 2035. As of its filing it had built 26 AI compute satellites and completed 15 missions, including a 12-satellite constellation launched in May 2025 that it describes as the first of its kind. Between 26 and 2,800 sits the investment case, and it is financed by selling satellites rather than compute. That shows in the customer list: the top five were 91.1% of first-half 2026 revenue. It shows in a gross margin of 7.6% on that line in 2025 — what happens when a product is priced deal by deal. None of this makes orbital compute a bad idea. It means the thing being valued — a recurring, high-margin compute service — is not yet the thing being reported. ## Honest limitations ADASpace's figures come from its Hong Kong prospectus as summarised in Chinese financial media; the filing is in Chinese and I have not read it in full, so segment definitions may differ. The Shanghai plan does not define whose satellites count toward 3,000. No throughput or utilisation data for on-orbit inference has been independently verified, and the token claim is the operator's own. The 2,800 figure is a corporate plan with no published funding schedule. ## What readers can do now - Price the order book, not the launch count: divide contracted value by satellite number, then track whether new orders are compute or imaging satellites. - Ask any orbital-compute vendor for two revenue lines, hardware sales and recurring compute service. If the second cannot be produced, value it as a satellite manufacturer. - Follow the Shanghai plan's 2027 milestone, not the 2030 headline, and check whether planned ground-station capacity matches the satellite count. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-decode-orbital-compute-order-book Free to quote with attribution and a link to the original.