BY:SpaceEyeNews.
Twelve satellites are no longer just watching Earth and the cosmos. They are computing in orbit. China’s latest test shows a space-based computing network can run AI models onboard, connect satellites through crosslinks, and deliver results without waiting on ground processing.
This matters because space data is exploding. Cameras get sharper. Sensors get faster. Yet downlinks still bottleneck. The new approach flips the workflow: process first, transmit later. That saves bandwidth and time. It also changes what “real-time” can mean in space science.
In this article, we’ll break down what China tested, how the system achieved its results, and why the shift could reshape space infrastructure.
The Three-Body Computing Constellation
China’s experimental system is built around the Three-Body Computing Constellation, developed by Zhejiang Lab with international partners. The long-term vision points to more than 1,000 satellites designed to deliver large-scale processing power in space.
The first milestone came in May 2025, when 12 satellites launched as the initial batch. That kickoff started an in-orbit test campaign that has run for nearly nine months.
Those months were not just “systems check” time. Mission teams verified multiple core functions:
- Crosslink networking between spacecraft
- Distributed computing across the group
- Deployment of AI models in orbit
- Operation and verification of scientific payloads
That list tells you the real goal. This is not a single smart satellite. It is the first real test of a space-based computing network that behaves like a connected orbital platform.
Space-based computing network architecture
A classic satellite pipeline looks simple: collect data, send it down, analyze it on Earth. The problem is scale. Ground stations have limited contact time. Bandwidth remains precious. Raw sensor feeds can overwhelm the pipeline fast.
This project pushes a different architecture: compute where the data is born.
Satellites as computing nodes
Zhejiang Lab’s constellation treats each satellite as a node in an orbital “cluster.” In practice, that means onboard processors handle tasks that Earth used to do. The system can filter, classify, and compress outputs before transmission.
Inter-satellite networking becomes the backbone
A computing cloud needs connectivity. In this test, the team demonstrated inter-satellite links among six spacecraft, a key milestone toward routing and sharing data directly in orbit.
That matters because crosslinks turn a set of satellites into a network. Without them, each spacecraft acts alone. With them, you can move data to the best node, share workloads, and keep operations going even when a satellite has no ground contact.
The 1,000-satellite ambition
China’s government-facing communications framed the goal as scaling to a network of thousands of satellites. One official report described the constellation’s direction as a major step in “space-based computing,” aiming for a large distributed system in orbit.
In short, the architecture tries to build something closer to orbital infrastructure than a traditional sensor fleet.
Ten AI models deployed in orbit
Here’s where this story stops being “interesting” and starts being disruptive: the constellation deployed 10 AI models and applications in orbit.
Two models stand out:
- An 8-billion-parameter remote sensing model
- An 8-billion-parameter astronomical time-domain model
The reporting emphasizes these as among the most capable AI models currently operating in orbit. In other words, this is not a tiny “edge AI” demo. It’s an attempt to run serious-scale models above Earth.
Why “8 billion parameters” in orbit is a big deal
Large models are heavy in every way that matters for satellites. They demand compute. They demand energy. They generate heat. Space systems live on strict power budgets. Thermal management is also unforgiving. This is why most space AI to date has stayed smaller and narrower.
So, if a constellation can run multi-billion-parameter models in orbit and keep them stable, it signals a step-change in what future satellites can attempt.
Updating models in orbit
Chinese reporting also notes that multiple models and algorithms have received updates deployed on orbit. That hints at a more flexible “upload and run” workflow, closer to cloud operations than classic satellite software cycles.
That flexibility is central to the promise of a space-based computing network: you can improve it over time without rebuilding the hardware.
The gamma-ray burst test that grabbed attention
One of the most concrete results in the reporting comes from astrophysics.
Two satellites in the constellation carry cosmic X-ray polarization detectors. These instruments work with an onboard AI model to classify gamma-ray bursts (GRBs) in real time. Reported accuracy: 99%.
Why that result matters for time-domain astronomy
If you follow Euclid and modern survey astronomy, you already know the theme: the sky changes fast. Time-domain science depends on quick identification and quick follow-up. But traditional workflows can lag. Data downlinks take time. Ground processing takes time. Alerts can arrive late.
Here, the constellation approach pushes intelligence upstream. The satellite decides what it is seeing while it is still seeing it. That can reduce the time from detection to classification. It can also reduce how much data you have to downlink.
The bandwidth win
The reporting explicitly says the onboard model sharply reduces the amount of data transmitted to the ground and processed on Earth. That is the practical upside of orbital AI: send conclusions, not raw noise.
This is the kind of change that scales. One event classification is nice. A thousand satellites filtering a planet’s worth of data is a different era.
How they got these results
The article you shared highlights four capabilities proven across the test campaign. Put together, they explain the “how.”
1) Crosslink networking
A connected constellation can share data in orbit. It can also route information toward the satellite best positioned for downlink. In this campaign, controllers demonstrated crosslink networking and, specifically, links among six spacecraft.
2) Distributed computing
Distributed computing means tasks do not live on one satellite. Workloads can move. Processing can spread across nodes. The reporting describes verification of distributed computing across the constellation.
3) AI model deployment in orbit
Deploying 10 AI models is not just about uploading files. It implies system-level support: runtime environments, scheduling, validation, and stable operations under power and thermal limits. The campaign verified that deployment and operation in orbit.
4) Payload operations tied to AI
The GRB example shows tight integration: instruments feed data to onboard AI, which classifies events in real time. That pairing is how you turn sensors into decision systems.
Each piece supports the same conclusion: this is not a single demonstration. It is a system proof.
What’s special about this approach
A lot of space projects look impressive on launch day. This one looks more impressive months later, because the test is about operations.
It targets the real bottleneck
Downlink has always been the quiet limiter. Sensors have improved faster than transmission. Orbital computing attacks that mismatch directly.
It shifts autonomy upward
With onboard inference, satellites can prioritize. They can decide what deserves transmission. Over time, that can make space systems more responsive and less dependent on constant ground intervention.
It treats space as infrastructure
The long-term plan for a huge constellation frames the project as a computing layer, not just an observing layer. China’s official communication around the May 2025 launch described it as a significant move in space-based computing, pointing toward a larger network concept.
This is why many commentators describe it as the early formation of an orbital “cloud.” Even if that phrase is informal, the architecture is moving in that direction.
What the world should learn from it
This is not only a China story. It is a roadmap signal.
Lesson 1: AI is becoming part of satellite design, not an add-on
Space missions used to treat AI as optional software. This project treats AI as mission-critical. That likely becomes the norm for next-gen constellations.
Lesson 2: “Edge computing” in orbit scales better than raw downlink
Sending everything home does not scale. Processing in orbit does. The more sensors you launch, the more you need this model.
Lesson 3: Time-domain science will benefit first
GRB classification is a strong showcase because it rewards speed. Expect more time-domain use cases next: transient alerts, anomaly detection, and rapid prioritization for follow-ups.
Lesson 4: The hardest part is operations, not launch
The project’s value comes from sustained testing: nearly nine months of in-orbit verification across networking, computing, model deployment, and payload operations.
Space-based computing networks will live or die on reliability, not headlines.
Conclusion: a space-based computing network is now real
This test campaign suggests a clear pivot. Satellites can form a space-based computing network that runs large AI models, shares data through crosslinks, and delivers real-time classification for scientific events like gamma-ray bursts.
The bigger message is architectural: orbit is becoming a place where decisions happen, not just data collection. If the long-term plan scales toward a thousand satellites, the result could look like a persistent orbital computing layer that supports astronomy, Earth observation, and new data services.
We are used to thinking of space as a telescope. We may need to start thinking of it as a processor.
Main sources:
SpaceDaily (original article): https://www.spacedaily.com/m/reports/China_tests_AI_satellite_swarm_for_space_based_computing_999.html
Xinhua (English, Feb 13, 2026): https://english.news.cn/20260213/f697fc260d66410398395307dd27443b/c.html
Xinhua (Chinese feature, Dec 30, 2025): https://www.news.cn/tech/20251230/9b207a5480ad4ba0bd50853fac20b508/c.html
Xinhua Zhejiang (Chinese, Feb 13, 2026): https://www.zj.xinhuanet.com/20260213/8387199a33184ccb82fa0d1fc254fcfd/c.html
China government site (May 15, 2025 launch): https://english.www.gov.cn/news/202505/15/content_WS6825452ec6d0868f4e8f28e6.html
Xinhua launch brief (May 14, 2025): https://english.news.cn/20250514/420ee02233394205b0e9aaf4dbc344bb/c.html