China plans 1,000 AI satellites, starting with nine in 2027
China's state aerospace group and Zhejiang province have put a number on orbital AI. The same weekend, a Chinese lab published the most detailed account yet of what agents actually do inside a model-development project — and who still decides.
A state-backed constellation called Lingjing — the name Qian Xuesen gave to virtual reality in 1990 — made its global debut at the Digital Trade Expo in Hangzhou: about RMB 30 billion to put 1,000 AI satellites in orbit, with the first nine launching in the first quarter of 2027. The operator is Hangzhou Siwei Space Intelligent Technology, registered in August with Zhejiang's provincial, municipal and district state capital as the largest shareholder and China Aerospace Science and Technology Corporation second — four levels of state money behind a company that did not exist two months earlier. Each satellite carries an on-board processing unit running Tianxuan Qianhe 2.0, described by its builders as China's first self-controlled, high-resolution, global-scale space AI foundation model, trained by a CASC research institute on remote-sensing data from China Siwei and released in Hangzhou on 4 September.
The published specifications are specific. A visible-light multispectral camera resolves 0.5 metres from 500 kilometres, with enough spectral bands to tell rice from wheat in orbit; visible, thermal-infrared, hyperspectral and radar payloads overlap so that night and cloud cover stop being excuses; and priority regions get a revisit every three to five minutes, against the hours-long gap a single low-orbit pass normally leaves. The intended workflow is the inverse of today's: instead of "sense in space, compute on the ground," the constellation is built to decide in orbit and downlink conclusions rather than raw pixels. The model also tasks across the fleet, handing a continuing observation to the next satellite when one finishes its pass.
Read the plan and the contract separately. Only the first nine satellites are in production, no launch provider has been named, and no orbit altitude or inclination has been published — the 1,000 is a stated target, not a manifest. What is already real is the stack around it: ZJLab's Three-Body Computing Constellation, 16 satellites with 25 AI models deployed and validated in orbit, is the largest space-computing constellation flying, and Lingjing's builders say they will open satellite interfaces and an on-orbit compute platform so universities and companies can fly their own industry models — the first candidate named is MUMO, a city-traffic model from a Hangzhou-listed data company. We covered the commercial end of this race last week — Orbital inference starts selling: Yusuan Xinghe ships in October.
Shanghai AI Laboratory published a study of the project that produced its own agentic model, and the finding worth keeping is not the benchmark table. Atria Dawn Preview is a 744-billion-parameter mixture-of-experts model built by post-training Zhipu's GLM-5.2, released under MIT with a 256K context, and the paper reports it leading five of 16 benchmarks — including AutomationBench and CyberGym — while trailing badly on others: 59.6% on SWE-bench Pro against Claude Opus 5's 74.7%, and 1583 on GDPval against 1768. Those scores are the lab's own, and no outside evaluator has rerun them. The collaboration data is the paper's real contribution. Across four weeks and 769 task records from 56 participants, the median number of agent actions per human input rose from 11 to 28.5, and roughly a third of completed AI-assisted tasks — 151 of 455 — were rated infeasible without AI by the participants themselves, spread across 27 of the 56 rather than a handful of power users.
None of that turned into autonomy. Humans made the final call on goals and scope in 93.4% of decisions and on methods and parameters in 85.5%, with agents proposing the method 55.4% of the time. When work hit trouble, 76% of cases moved forward through human intervention — mostly adding context (35.2%) or diagnosing the problem and switching approach (34.7%) — and full human takeovers were 0.7%. The authors' warning is the part to carry: participants routinely ran agents in autonomous modes to avoid interrupting long runs, a boundary drawn for convenience rather than decided, and when every choice rests on a chain of agent work longer than any human can review, oversight slides into rubber-stamping. It matches what we found when the original authors graded the two papers their agents were asked to write — frontier agents flunked two unpublished NeurIPS papers.
What to watch: whether Lingjing's first nine satellites get a named launch vehicle and a date that holds, and whether anyone outside Shanghai AI Lab reruns Atria's five benchmark wins.
If an agent runs 28.5 actions for every decision you make, which of those decisions do you still actually understand? Tell us in the comments.
