Nine months of orbital testing. 94 percent accuracy. Government data, not independently verified — but further along than anything the US has publicly demonstrated in orbit. The bigger problem: American AI refuses nearly all military queries.
image from Gemini Imagen 4
China has barred Manus chief executive Xiao Hong and chief scientist Ji Yichao from leaving the country following Meta's announced acquisition of the autonomous AI agent startup for over $2 billion. This marks the first time Beijing has used exit bans directly against executives involved in a deal with a major US technology firm, serving as an enforcement mechanism while China's Ministry of Commerce conducts an ongoing review of the acquisition that began in January 2026. Manus had already relocated its headquarters to Singapore in mid-2025 and replaced state-linked investors with American venture capital before the exit restrictions were imposed.
China used exit bans against Meta deal executives. That's the story.
On March 26, Beijing barred Manus chief executive Xiao Hong and chief scientist Ji Yichao from leaving China. The two were summoned by the National Development and Reform Commission alongside Meta executives. No charges were filed. No trial was offered. They simply cannot leave.
This was not a random regulatory move. Meta announced in December 2025 that it would acquire Manus, the autonomous AI agent built on Anthropic Claude 3.5 Sonnet and refined Alibaba Qwen models, in a deal valued at over $2 billion, according to reporting by Euronews and eWeek. Manus had already wound down its Chinese operations, relocated its headquarters to Singapore in mid-2025, and replaced state-linked investors with American venture capital, as AINvest reported. The company had done exactly what Beijing has been trying to prevent: built something commercially viable and sold it to the Americans. The exit bans are the punishment. They are also a first. According to eWeek, this marks the first time Beijing has used exit restrictions to directly target executives involved in a deal with a major US technology firm.
China's Ministry of Commerce launched a formal review of the Meta-Manus acquisition in January 2026, as Entrepreneur Loop reported. The review is ongoing. The exit bans serve as the enforcement mechanism while it plays out. What happens to the deal if the review concludes no? The executives remain in China. That's not a legal proceeding. It's a hostage situation with extra steps.
The SpaceNews op-ed published March 31 by retired Lieutenant General David Wood and retired Major General Erin M. Dick was titled Agentic AI: The Future of Space Warfare. It was not a news article. It was an opinion piece from two defense consultants who have a financial interest in defense spending increasing. The piece cited their own analyses, along with a Georgetown CSET report and a Space Systems Command article, as if these sources had equal weight. They do not.
More importantly, the generals buried the lede. The real story is not that AI will change space warfare in the abstract. The real story is that China is already running AI in orbit, and that the United States has a structural problem with commercial AI that makes it unreliable for exactly the use cases the generals say matter most.
Nine months of orbital testing
The China Aerospace Science and Industry Corporation completed nine months of on-orbit testing for the Three-Body Computing Constellation in February 2026, according to SatNews. Twelve satellites launched May 14, 2025 aboard a Long March 2D from Jiuquan Satellite Launch Center, as SpaceNews reported. Those twelve satellites form the initial tranche of a planned 2,800-satellite network called the Star-Compute Program, a joint effort between ADA Space and Zhejiang Laboratory, also confirmed by the Chinese government. The constellation delivers 5 peta operations per second of combined computing power, 30 terabytes of onboard storage, and inter-satellite laser links rated at up to 100 gigabits per second, per SpaceNews. The launch marked the world's first space computing constellation and the first deployment of ADA Space's Star Computing plan, according to official Chinese government reporting.
On those satellites, China successfully deployed and executed an 8-billion-parameter large language model designed for remote sensing analysis, per SatNews. The model ran entirely on-orbit. The satellites demonstrated the ability to classify astronomical phenomena and terrestrial infrastructure types with 94 percent accuracy without ground intervention. For specific reconnaissance tasks, the system reduced the volume of data transmitted back to Earth by a factor of 1,000. You don't compress a 1,000-to-1 ratio without meaningful onboard intelligence making decisions about what matters and what doesn't. This is not a technology demo. This is a deployed system with demonstrated military utility.
ADA Space, legally Chengdu Guoxing Aerospace Technology Company Limited, was established in 2018 and filed for a Hong Kong initial public offering in January 2025, per SpaceNews. Zhejiang Laboratory, in Hangzhou, was founded in September 2017 as a collaboration between Zhejiang provincial government, Zhejiang University, and Alibaba Group, also per SpaceNews. These are not garage startups. They have state backing, institutional depth, and a demonstrated orbital capability that the United States currently cannot match with anything it has publicly acknowledged.
Meanwhile, US commercial AI refuses to do military work
The EdgeRunner artificial intelligence study, published on arXiv as paper 2603.10012, tested 24 leading frontier large language models against military-use queries authored by United States Army veterans including a Special Forces veteran of 20 years, with no AI assistance in constructing the benchmark. The results were not close. Hard rejection rates reached 98.2 percent across frontier models, per the arXiv study. The military-tuned EdgeRunner 20B model showed a 96.7 percent refusal rate before a targeted modification process improved it to 29 percent, at the cost of a 2 percent performance drop on other military tasks, according to the same study. EdgeRunner chief executive Tyler Saltsman described the situation plainly: the models are so overly sensitive they simply will not be helpful, as National Defense Magazine reported.
This is the data infrastructure gap the SpaceNews op-ed either did not understand or chose not to mention. When American commercial AI companies build refusal into their products as a safety feature, and when China deploys purpose-built models directly onto satellites with no refusal layer, the gap between the two countries is not just about model quality. It is about whether the model will even respond when a warfighter needs it.
The Georgetown Center for Security and Emerging Technology reviewed thousands of open-source requests for proposal published by the People's Liberation Army between January 2023 and December 2024, per the CSET China Military AI Wish List report. The PLA is not asking politely for AI systems on long procurement timelines. The requests specify acquisition timelines of three to six months, per National Defense Magazine. They are requesting space targeting capabilities for targets on land and in space. CSET research fellow Sam Bresnick put it directly: China has made surprisingly specific requests for space targeting capabilities, adding more supporting evidence to warnings about China's offensive space ambitions, per the same National Defense Magazine reporting.
General Chance Saltzman, Chief of Space Operations for the US Space Force, said publicly that the next era of space warfighting will be driven by artificial intelligence, autonomy, and maneuver operations, per Air and Space Forces Magazine. The Space Force has outlined a three-tier AI framework: Enterprise AI for general tools, Functional AI for profession-specific applications, and Mission-Specific AI for custom warfighting systems, per Space Systems Command. The service's annual appropriation has grown from $15 billion in fiscal year 2021 to nearly $40 billion in fiscal year 2026, per Air and Space Forces Magazine. Space Systems Command Data and Artificial Intelligence Officer Bartley Stewart has described $300 billion in annual private AI investment available for national defense partnership, per the same Space Systems Command article. The money is there. The procurement architecture is not moving fast enough to matter against a PLA that issues RFPs with six-month timelines.
The AI Diffusion Rule proposed by the Biden administration would have created a tiered export system for AI chips that many international partners viewed as second-tier status. The Trump Commerce Department withdrew it on March 13, 2026, per the Bureau of Industry and Security press release and Reuters. Under Secretary for Industry and Security Jeffery Kessler said the rule would have stifled American innovation and saddled companies with burdensome new regulatory requirements. Whatever the merits of that decision, it also removed a tool that would have restricted the flow of advanced AI computing hardware to adversaries. The regulatory floor under which China can acquire American AI chips just got lower.
There are two separate stories here that the SpaceNews op-ed conflated into something vague about agentic AI and space warfare.
The first is geopolitical enforcement. Beijing has used exit bans against the chief executive and chief scientist of a company that sold itself to Meta. This is a new instrument of statecraft. It does not look like anything in the international legal framework for cross-border technology transactions. It looks like what it is: a government treating its citizens as assets that cannot leave without permission, deployed in response to a transaction it chose not to approve. Every technology company considering an acquisition involving Chinese operations or Chinese founders now has to factor this into their deal structure. That is a significant change in the environment for global tech dealmaking.
The second is operational capability. China has demonstrated AI inference running on satellites in low Earth orbit, with onboard decision-making that reduced data transmission by three orders of magnitude for reconnaissance tasks. The Three-Body Computing Constellation is not theoretical. It is tested. The 94 percent autonomous classification figure comes from nine months of on-orbit operation, not a lab result or a computer simulation. If the PLA is already running LLMs in space that classify infrastructure and reduce bandwidth by 1,000x, the question for the United States is not whether to develop similar capability. It is how quickly and from where.
The SpaceNews op-ed by retired generals David Wood and Erin M. Dick talked around both of these stories. It cited the CSET PLA procurement data and the EdgeRunner study without foregrounding what they actually show. It called for US action without noting that the American commercial AI ecosystem is structurally unwilling to serve the use cases that matter most in a conflict. It described China's space AI program as part of a general trend without reporting the specific orbital numbers that make it concrete.
Two retired generals writing an opinion piece for a defense trade publication are not bad sources. They are just not the story. The exit bans are the story. The satellites running LLMs at 94 percent classification accuracy without ground contact are the story. The 98.2 percent refusal rate from commercial AI when asked to help with military tasks is the story. Anyone covering this space and missing those three facts has not done the reporting. They have done the briefing.
Story entered the newsroom
Research completed — 0 sources registered. The SpaceNews piece is an op-ed by two retired generals (Wood, Dick) synthesizing secondary sources — not original reporting. The story conflates thre
Draft (952 words)
Reporter revised draft based on fact-check feedback (936 words)
Approved for publication
Headline selected: Manus Already Left China Before Beijing Trapped Its Founders
Published (1671 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Space & Aerospace · 9h 15m ago · 5 min read
Space & Aerospace · 19h 48m ago · 3 min read
