Welcome to Issue 16, where China opens its first space-computing hub
The story this week is where space data gets interpreted: China opened infrastructure to run compute in orbit, and three research teams pushed their models out to the telescope, the satellite, and the rover.
On the Space4AI side, the thesis got a physical address. Shanghai's Songjiang district inaugurated what it calls China's first dedicated space-computing industry hub, convening aerospace, semiconductor, AI, and cloud players around in-orbit data processing — and spotlighting billion-parameter language models already running on an orbiting constellation.
On the AI4Space side, the models are learning to watch objects in space and read the ground beneath a rover. A UK institute released a foundation model that learns normal satellite behaviour from sunlight reflected off orbiting objects and flags anomalies for human review. A separate team published an unsupervised anomaly detector built to run on the satellite itself, and a third posted a lightweight terrain-vibration model for autonomous rovers working under tight compute and power budgets.
Both sides are moving interpretation to the edge — compute toward orbit, models onto the hardware that gathers the data.
Specifics below.
Did you know?
We have recently launched the Bits & Orbits Insights
Check out the growing set of independent dashboards, trackers, and maps on the
AI × Space frontier:
Space4AI
Shanghai's Songjiang district opens China's first space-computing industry hub, spotlighting 8B-parameter LLMs running on Three-Body Computing Constellation
On August 31, 2026, Shanghai's Songjiang district inaugurated China's first dedicated space-computing industry hub at an inaugural space-computing forum.
The dedicated centre in the city's southwestern Songjiang district brought together aerospace developers, semiconductor makers, artificial intelligence specialists, cloud vendors, as well as researchers and investors to advance in-orbit data processing.
The forum is the most visible institutional moment yet in a buildout that has been underway for more than a year. A European Space Policy Institute brief published earlier that month, titled "Constellating Compute: China's Domestic Policy Towards Orbital Data Centre Leadership," restates the Europe-dependence warning we covered in Issue 12.
The hub anchors a supply chain that already reaches orbit.
The Three-Body Computing Constellation, led by government-affiliated Zhejiang Lab in partnership with Chengdu-based ADA Space, put its first 12 satellites into low Earth orbit in May 2025 — a launch we covered in Issue 7 — and each carries an 8-billion-parameter AI model capable of 744 trillion operations per second, running on commercial off-the-shelf processors inside a 6U CubeSat platform.
The constellation uses free-space optical inter-satellite links running at up to 100 gigabits per second, and on July 29, 2026, its developers announced that two satellites had maintained a continuous laser communications link for 192 hours, 8 minutes, and 45 seconds — just over eight days — at 99.99% uptime across distances up to 1,000 km, with zero data loss.
The rationale is a downlink bottleneck.
Li Chao, director of the space-based computing system research center at Zhejiang Lab, describes an inefficiency in which up to 90% of the data generated by satellites is currently lost and never effectively processed.
Space computing inverts the traditional "data pipe" architecture: an onboard AI model processes raw imagery in orbit and transmits only the answers, such as confirmed ship coordinates, infrastructure damage reports, and crop stress patterns, and the Three-Body Constellation demonstrated a 94% autonomous classification accuracy rate for astronomical phenomena and terrestrial infrastructure without ground intervention, according to its developers.
Songjiang concentrates the rest of the industrial chain.
The district is home to SpaceSail, the Shanghai government-backed firm building Qianfan, China's main low-Earth-orbit challenger to Starlink, which has launched over 200 satellites to date and aims to deploy a 15,000-satellite mega-constellation by the end of 2030.
SpaceSail raised approximately ¥6.976 billion (about $1.04 billion) in August 2026 in a round that explicitly barred all foreign capital and left at least 80% of equity in state hands.
The district also hosts satellite manufacturers including Gesi Aerospace and Xuntian Qianhe, as well as the Shanghai Aerospace Precision Machinery Research Institute.
China's lead is uneven.
By comparison, Starcloud placed an NVIDIA H100 GPU into orbit in November 2025 without distributed networking — a flight we covered in Issue 8. Aethero's Phobos satellite delivered roughly 157 teraflops for customers including Booz Allen Hamilton and the U.S. Air Force Research Laboratory. SpaceX's AI1 satellite — rated at 150 kilowatts of peak compute using NVIDIA Rubin GPUs — remains a design with no prototype launches until 2027; we covered the Starmind AI1 partnership in Issue 12.
All performance figures cited by Chinese programs — the 94% accuracy, the 744 TOPS per satellite, and the 8-day laser link duration — come from Chinese state or state-affiliated sources, and no Western institution has independently audited the constellation's performance claims.
AI4Space
Alan Turing Institute's DARe releases self-supervised foundation model for satellite behaviour monitoring
It is described as the first AI model to predict anomalies and satellite motion by learning from the way light is reflected off objects in space.
The model works much like a large language model, but trained on satellite brightness readings rather than text.
It is first trained on large quantities of telescope-gathered "light curves" to learn normal patterns of satellite behaviour, then fine-tuned with curated simulation data from Strathclyde's Aerospace Centre of Excellence and industry partner GMV. In operation, the model is fed real-time or recent light curves from ground-based observatories and flags anomalies for human experts to investigate.
In testing, it identified unusual or unexpected light curves 88% of the time and could distinguish between behaviours such as spinning versus tumbling, distinctions that are essential for in-orbit servicing and extending satellite life.
The work forms part of AI4 Space Safety and Sustainability (AI4S3), a UK Space Agency International Bilateral Fund consortium spanning Five Eyes countries, with partners including the Turing, the University of Arizona, MIT, the University of Waterloo, and industry partners GMV, Columbiad, LMO, and Zendir.
The tool targets an increasingly crowded environment.
Over 4,000 new satellites launched in 2025, compared with 159 worldwide launches in 2000, and SpaceX's Starlink alone has more than 10,000 satellites in orbit with aims to reach 42,000.
Nockles, senior author on the paper, said that while in-orbit collisions are rare their consequences are catastrophic and can severely disrupt infrastructure supporting communications, GPS, and precision timing for financial markets.
Next steps include researching multimodal systems that go beyond light curves to incorporate radar data, hyperspectral data, and satellite orbit data.
Peak-AAE unsupervised orbit anomaly detector published for on-satellite edge deployment
On 2 September 2026, four researchers at the Shandong Key Laboratory of Space Debris Monitoring and Low-Orbit Satellite Networking, Qingdao University of Technology, published a lightweight orbit anomaly detection method built around a Peak Adversarial Autoencoder (Peak-AAE) in Aerospace, a journal of the Multidisciplinary Digital Publishing Institute (MDPI).
The unsupervised method is trained on historical semi-major axis telemetry and is optimised for edge deployment on satellite hardware, eliminating manual feature engineering and reducing false-positive rates versus a baseline AAE.
The pipeline pulls two-line element data from Space-Track, derives the orbital semi-major axis from mean motion, and segments it with a sliding window before reconstruction. The squared difference between reconstructed and original data forms a residual sequence, and a threshold-free Automatic Multi-Scale Peak Detection (AMPD) algorithm flags anomalous points on that residual. A peak-probability generation module — three parallel 1-D convolutions at kernel sizes 3, 5 and 7 — computes a per-point peak probability, which drives a non-peak loss term during adversarial training. That term raises reconstruction accuracy at peak regions, where the authors say conventional VAE, GAN and Transformer reconstructors otherwise confuse the onset and end phases of a short-duration anomaly and misfire.
For on-orbit inference the model is compressed from roughly 48,000 parameters to about 9,300, and from 201 KB to 52 KB, with the maximum accuracy degradation held to 2.85%. Validation used two navigation satellites, BEIDOU-3 G2 and IRNSS-1A, against baselines including VAE, USAD, the Anomaly-Transformer and the original AAE. Across those tests the compressed Peak-AAE reached a minimum precision of 90.00% and a minimum recall of 84.74%, and the authors report a ground simulation of the system running on an edge device.
RoughSense: lightweight point-cloud + IMU model for real-time rover terrain-vibration prediction
On September 3, 2026, researchers from the University of Luxembourg and IRL Georgia Tech–CNRS in Metz posted RoughSense to arXiv, a lightweight method for real-time, vibration-aware traversability mapping aimed at autonomous space rovers operating under tight communication, compute, and power budgets.
The system fuses a LiDAR point cloud with inertial measurement unit (IMU) feedback. An initial vibration proxy is estimated from terrain geometry by applying RANSAC to local point-cloud patches produced by a SLAM algorithm, while the IMU supplies direct observations of the vibration the rover experiences as it drives. The geometric prediction is then corrected online using Recursive Least Squares, letting the system adapt its terrain estimate to the measured rover response.
The authors frame the work around underground operations, where limited communication and onboard resources make reliable terrain assessment hard. They evaluate the approach across three settings: a lunar analogue environment, an outdoor field, and an underground mine.
Till next time,
Meta-beat Column of this week
Read also about the AI Pipeline that sits at the core, producing this Newsletter, including its ups and downs of this week: