SenseTime launches Galaxy Project to scale China’s AI chips

SenseTime announced the Galaxy Project during a keynote titled “Intelligent Transformation and Symbiosis.” The initiative brings together almost 20 domestic partners to expand AI chip infrastructure, link chip-level technology with ecosystem partners and enable commercial deployment of domestically produced AI computing power.

The company signed a space computing agreement with satellite maker Guoxing Aerospace and set up research collaborations with five institutions: the Shanghai Artificial Intelligence Laboratory, Beijing Zhongguancun Academy, Shenzhen Hetao Academy, the Shanghai Algorithm Innovation Research Institute and Shanghai Jiao Tong University’s AI school. The project aims to support scientific computing work in life sciences, materials science and manufacturing.

SenseTime reported that its large-scale device platform currently processes about 2.42 trillion tokens per day and projected growth to 10 trillion tokens per day by the fourth quarter of 2026. The company described a closed loop connecting chip design, software toolchains and application deployment, and said the Galaxy Project will build one “token factory,” five large computing clusters at what it called “10,000-calorie” scale, advance ten technology directions and support 200 AI startups.

On performance, SenseTime presented results for its heterogeneous hybrid inference technology. The company reported improvements in Model FLOPs utilization on mainstream domestic chips of 85–152 percent, an inference cost-effectiveness ratio of 1.25 times compared with Nvidia H-series parts, and a 2.5-times increase in token output over domestic homogeneous inference setups at equivalent cost. The firm noted these figures are its own measurements and require independent benchmarking in customer environments.

To address software fragmentation across Chinese AI chips, SenseTime said it has developed a full-stack adaptation layer that covers models, frameworks, operators, toolchains and hardware. The company cited applied tests: a long-sequence protein prediction workload where fused operator optimizations reduced prediction time by threefold, and an AI-generated-content video test where domestic chips running DiT models reached a 93 percent multi-card parallel acceleration ratio. SenseTime described a zero-cost migration path for mainstream AI development tools pending real-world customer trials.

SenseTime introduced a new energy-efficiency metric called Tokens Per Watt and unveiled a Computing-Power Collaboration Agent to manage resource scheduling, electricity price prediction and energy storage optimization across an eight-level data system with five decision chains. In pilot results, the company reported an 80 percent increase in token output per unit of electricity cost, average power prices 10 percent below comparable regional data centers, and 96 percent accuracy in computing load prediction, and said these results should be tracked over full seasonal cycles.

The Galaxy Project roster includes chip designers and component suppliers such as Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Biren Technology and Sunrise, and infrastructure partners including Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information and Jiliu Technology.

SenseTime outlined physical infrastructure plans that include a Shanghai data center rated at what the company calls “5A” intelligent computing level, a Yancheng site with an initial 3,000 petaflops of capacity, a Hong Kong center targeting 40,000 petaflops by 2030, and a planned full-stack domestic computing cluster in Saudi Arabia. The space computing plan calls for a first satellite launch in 2026 and an eventual build-out of thousands of satellites and computing capacity measured in the tens of thousands of petabytes by 2030.

Yang Fan, co-founder and president of SenseTime’s Large Device Business Group, presented the Galaxy Project as a collaborative effort across chips, components, infrastructure and application scenarios. SenseTime’s forecast of 10 trillion tokens per day by Q4 2026 is a discrete target the company highlighted for future reporting.