Zhejiang Lab set out the state of its Three-Body Computing Constellation on 29 July: 12 satellites now in orbit with a combined 5 quadrillion FLOPS, running 20 AI models between them, including an 8-billion-parameter model for remote sensing. The plan is roughly 100 satellites by 2027 and about 1,000 by 2032.
The argument is bandwidth, not ambition
The case for computing in orbit is unglamorous. A single remote-sensing satellite generates around 0.1 petabytes per day. A constellation of 3,000-plus would therefore produce on the order of 300 petabytes daily, requiring something like 100,000 ground servers to ingest and process — assuming the downlink capacity to get it to the ground at all, which does not exist. Running the model where the sensor is turns a bandwidth problem into a compute problem, and ships answers instead of pixels.
What that changes
For applications where latency is the product — disaster response, maritime monitoring, crop and flood assessment — the difference between analysis in orbit and analysis after downlink is the difference between a usable answer and a historical record.
What is not disclosed
Neither the power budget nor the thermal envelope of the compute payloads is given, and both are the binding constraints on orbital hardware. Nor is there detail on which models run where, how they are updated, or who the users are. The 5-quadrillion-FLOPS figure is a combined constellation number, not a per-satellite one.
The wider pattern
Every gigawatt-scale AI announcement this month has been an argument about where compute can physically go: to El Paso, to a Texas oilfield, to whichever grid has spare interconnection. This one answers the same question by leaving the grid entirely — and if the roadmap holds, by a factor of eighty within six years.