SenseTime has launched the Galaxy Project with nearly 20 partners to accelerate AI infrastructure built around Chinese chips. The initiative connects components, computing centres, software adaptation and commercial deployment in an effort to turn domestic hardware into a usable multi-vendor ecosystem.
Its ambition is broad, extending from near-term inference clusters to optical, quantum and orbital computing. The central claims are substantial—and largely await independent verification.
Galaxy tries to solve China’s fragmented stack
Domestic accelerators have struggled with incompatible software. A model prepared for one architecture may require extensive engineering to run on another, weakening the practical value of having multiple chip suppliers.
SenseTime says its full-stack adaptation layer spans models, frameworks, operators, toolchains and hardware. The goal is to move workloads between domestic vendors without major rewrites. Partners include Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise and Biren Technology, alongside component and infrastructure companies.
The project plans one “token factory,” five large computing clusters, collaboration across ten technical areas and support for 200 AI startups. This ecosystem approach recognises that replacing an Nvidia accelerator alone does not create an alternative platform; drivers, compilers, networking, orchestration and developer tools must mature together.
SenseTime cites a threefold reduction in prediction time for a long-sequence protein workload and a 93% multi-card parallel acceleration ratio for video-generation models. It also claims that common AI development tools can migrate at no additional cost. Real customer pipelines spanning hardware generations will provide a more demanding test than optimised demonstrations.
Throughput and energy claims require outside measurement
SenseTime says its large-device platform currently processes 2.42 trillion tokens each day and forecasts 10 trillion by the fourth quarter of 2026, describing that as a 25-fold increase despite the figures not appearing arithmetically aligned. The 10-trillion target is a projection, not a measured result.
Its heterogeneous inference system reportedly improves Model FLOPs Utilisation by 85% to 152% on mainstream domestic chips. The company places cost effectiveness at 1.25 times Nvidia H-series systems and token output at 2.5 times comparable homogeneous domestic clusters for the same cost.
No independent benchmark accompanies these numbers. Buyers should seek reproducible tests using their own model mix, context lengths, latency targets and utilisation patterns. A data centre running continuously differs from a demonstration cluster kept near an ideal load.
SenseTime also proposes Tokens Per Watt as an efficiency measure. Its scheduling agent combines workload allocation, power-price prediction and energy storage, with claimed gains of 80% in token output per unit of electricity cost, electricity prices 10% below regional peers and 96% load-prediction accuracy.
Tokens Per Watt can be useful only when token quality and workload are comparable. A small model generating unnecessary output may look efficient while doing less valuable work. Energy accounting should include networking, memory, cooling and idle capacity, not just accelerator consumption.
Physical expansion mixes practical sites with distant bets
SenseTime points to a Shanghai facility serving more than 20 industries, an initial 3,000-petaflop site in Yancheng and plans for a 40,000-petaflop Hong Kong centre by 2030. It also intends to build an overseas domestic-computing cluster in Saudi Arabia.
Research partnerships with five institutions target life sciences, materials and manufacturing. These projects could provide valuable workloads for testing whether the hardware abstraction performs outside generative media.
The longer-term vision includes optical and quantum systems plus a satellite constellation with a first launch planned in 2026 and thousands of computing satellites envisaged by 2030. Space-based processing could help weak-network environments, but deployment at that scale has little precedent and should be separated from near-term procurement decisions.
Galaxy’s strongest idea is coordination across the stack. China’s chip challenge is not a shortage of individual designs; it is making diverse components behave like a dependable platform. The measure of progress will be verified throughput, application portability and total energy cost at customer sites. Until those results appear, SenseTime has presented a serious roadmap and partner network—not yet proof that domestic infrastructure matches the economics it claims.
