China's first AI computing satellite fleet takes flight

Traditional Earth-observation satellites have a glaring bottleneck: they take crystal-clear pictures, then wait hours until they pass over a ground station to dump the data for processing. In a wildfire or flood, that delay can be fatal.

Now, Chinese engineers are putting AI data centers directly into orbit.

A Long March 2D rocket blasted off carrying 12 AI-powered satellites, kicking off the ambitious "Star-Compute Project" (also dubbed the Three-Body Computing Constellation). Developed jointly by ADA Space (Guoxing Aerospace) and Zhejiang Lab, the initiative aims to deploy an eye-popping 2,800 AI satellites, with orbit and spectrum allocations already cleared by the International Telecommunication Union (ITU).

Running 8B models in the cold of space

These first 12 spacecraft serve as operational prototypes for the broader constellation, and the specs are surprisingly beefy.

Each satellite packs onboard AI processors delivering 744 TOPS of compute, pushing the initial batch to a combined 5 POPS. More strikingly, the fleet carries AI models with up to 8 billion parameters alongside Earth-observation sensors and X-ray polarization detectors designed to spot gamma-ray bursts.

To share the heavy lifting, the satellites link up with 100Gbps laser inter-satellite communications. Instead of piping raw, gigabyte-heavy images down to Earth, the satellites can talk to each other, run inferences in orbit, and send back finished insights in real time.

Clever hardware, big-tech backers

Launching nearly 3,000 satellites requires serious logistical flexibility, so ADA Space split the spacecraft into two distinct form factors.

Type 1 satellites use a flat, stackable design so dozens can pile onto a single heavy rocket. Type 2 satellites are standard CubeSats designed to hitch rides as secondary payloads whenever a rocket has spare margin.

According to ADA Space's Hong Kong IPO filings, the goal is to weave space-based compute with ground networks into a unified AI infrastructure by Stage III. Rather than shouldering the burden alone, the venture has partnered with Chinese tech giants including Alibaba Group, Kepu Cloud, and iSoftStone.

Why it matters

Edge computing has officially reached low Earth orbit. For disaster management, remote sensing, and low-altitude drone tracking, crunching pixels on the satellite cuts turnaround times from several hours to near real-time.

Strategically, shifting from "sense in space, compute on Earth" to orbital AI tackles the brutal bandwidth constraints of satellite downlinks. Pumping 100Gbps laser meshes between satellites essentially turns a constellation into a distributed floating data center.

That said, we think the physical hurdles are steep. Dissipating heat in a vacuum while shielding power-hungry AI accelerators against cosmic radiation is notoriously difficult. Running an 8B model reliably in orbit is an impressive engineering showcase, but ADA Space still has to prove that deploying and maintaining 2,800 satellites makes economic sense beyond the initial launch hype.