STAR.VISION's 1,080-satellite Space.IDC plan leans on $1.49B in Chinese bank credit

STAR.VISION Aerospace (地卫二空间技术) used the Global Digital Trade Expo on 25 September to unveil what it calls the Space.IDC Computing Constellation, or Space String Computing Constellation. The money behind it is the part worth reading twice. China in Space, which reported the plans on 29 September, says the company holds strategic credit agreements with the Industrial and Commercial Bank of China and the Agricultural Bank of China worth 10 billion Yuan, or 1.49 billion US dollars as of 28 September. At least 1 billion Yuan, about 149 million dollars, is earmarked for a "Phase One." The company gave no further detail on the drawdown schedule.

The constellation itself is large on paper. About 1,080 spacecraft would make up the main system, fully deployed by 2035, with roughly 720 for inference tasks and 360 for general compute. Tasks would arrive from Earth or from other satellites that lack on-board pre-processing.

Three spacecraft, three power budgets

STAR.VISION outlined three designs. The G1 verification craft is meant to deliver 40 to 60 PetaFLOPs from 11.5 kilowatts of solar generation. The G2 standard unit moves to 160 to 310 PetaFLOPs on 48 kilowatts. The G3 flagship is quoted at 460 to 1000 PetaFLOPs on 140 kilowatts.

Several engineering questions went unanswered at the unveiling. The company did not share orbits, inter-satellite links, radiator size or mass, and it has not filed the international regulatory paperwork that would normally fill those gaps. STAR.VISION did say it has booked an undisclosed number of launches on Orienspace's Gravity-1 and Galactic Energy's Ceres-2 for the G1 craft, but gave no target month or quarter. The company has prior form here: it has been putting AI systems on remote sensing satellites since at least mid-2023, and it claims two satellites built for Uzbekistan and Indonesia have already verified parts of the coming constellation.

It is also not alone. China in Space notes that ADA Space and Zhejiang Lab's Three-Body Computing Constellation already has about a dozen satellites in orbit, that Orbital Chenguang has major financial backing, and that Chaozhisuan Beijing Technology has well-linked partners, both with demonstration satellites working. State-owned firms are pursuing large computing spacecraft with little public noise. STAR.VISION disclosed agreements around chips and manufacturing with, among others, the Hangzhou Institute for Advanced Study at UCAS, the Hangzhou Research Institute of Xidian University, Hongqing Technology, Baidu AI Cloud and MinoSpace, plus unspecified support from several Zhejiang zones and the cities of Xichang and Chongzhou.

Autonomy moves on a slower track

On the ground, the harder question is what all this compute would actually decide. IEEE Spectrum reported on 29 September that NASA's Jet Propulsion Laboratory used Anthropic's Claude models last December to help plan two Mars drives for the Perseverance rover, with humans checking the route before upload. In May, NASA and IBM placed a compressed AI model on the International Space Station and a satellite to spot floods and clouds from orbit. In July, astronauts on the ISS tested a large language model for maintenance questions.

"Autonomy often does not have a deterministic outcome, which means that how you got into a certain situation changes the behavior," Robert Ambrose, former chief of NASA's Software, Robotics, and Simulation Division, told IEEE Spectrum. "So if you come into the same situation but from different paths, the outcome could be different. And so engineers hate that."

That caution is the counterweight to the orbital data centre pitch. In-orbit computing promises solar power without weather and heat rejection into vacuum, but it also moves the reliability problem off Earth. Elsewhere, the same week produced a more mundane reminder of how hard automated interpretation remains: Labqoat tested eight models on 2,000 wildlife trace photos and found the best, Claude Opus 5.5, identified the correct species 37.15% of the time, with only 51.75% of photos resolved to the right family.

Meanwhile the model market keeps churning. A stealth reasoning system listed on OpenRouter as space-bunny-alpha advertises a one-million-token context and a free preview, with its own published benchmark table placing it behind GPT-6 Sol and Qwen3.8 Flash on GPQA Diamond and MMLU-Pro. None of that settles whether a satellite should act on its own conclusions.