SenseTime is moving part of its AI infrastructure story above the ground. At WAIC, the company and ADA Space said they will build a “SenseTime compute constellation,” starting with four compute satellites this year.
The roadmap is concrete enough to test: four satellites in 2026 for early network validation, a space-ground hybrid commercial compute backbone in 2028, and a thousand-satellite constellation with more than 10,000P of total compute by 2030.
The first job is not a data center in space
The useful way to read the plan is narrower. Most remote-sensing workflows still collect data in orbit, downlink it, and process it on the ground. That burns bandwidth and time. Space compute puts a first layer of filtering, compression, recognition and inference closer to the sensor, then sends back results or high-value slices.
That matters for wildfire detection, maritime monitoring, disaster response and weak-network regions. It does not mean a satellite replaces a GPU cluster. It means some decisions may happen before raw data reaches a terrestrial data center.
The ground stack remains central
SenseTime also said it will work with 15 domestic chip companies to build a Token Operations Center and five domestic compute clusters above the ten-thousand-card scale within two years. Caixin and EET China both reported the 15-company ecosystem and the five-cluster plan.
The architecture is therefore three-layered: ground clusters for bulk training and inference, orbit nodes for real-time preprocessing near data sources, and a scheduler deciding where each task should run.
Yang Fan said domestic compute cards in certain Prefill workloads inside heterogeneous domestic inference clusters had reached more than twice the cost performance of NVIDIA H-series cards, and more than 200% better performance than before tuning.
The scope of that quote is important. It is about selected Prefill tasks in a specific heterogeneous inference setup. It should not be read as a blanket claim about all AI training or all inference.
The verification points are engineering points
ADA Space supplies the aerospace track record. The Paper cited filing materials saying the company launched China’s first AI application satellite in 2018 and its first AI large-model compute satellite in 2024; it also raised RMB 3.55 billion in 2026 at a post-money valuation of about RMB 11.554 billion.
The remaining tests are practical: when the first four satellites launch, what compute, power and latency numbers are disclosed, how tasks are split between orbit and ground, and whether customers receive minute-level results instead of demo videos.
Space compute is a bold infrastructure pitch. It will only become infrastructure if the task bill closes: faster, cheaper or more reliable than sending everything back to earth.
Sources: Caixin, The Paper, EET China, CocoLoop, Zhidx; checked the four-satellite first batch, 2028 backbone, 2030 thousand-satellite and 10,000P plan, 15 domestic chip partners, five ten-thousand-card clusters, daily token call volume, and ADA Space satellite and financing figures.