SenseTime and the Galaxy Project: A Revolution in AI Chips in China

SenseTime and the Galaxy Project: A Strategic Alliance for AI

SenseTime, a major player in the field of artificial intelligence, has recently unveiled its ambitious Galaxy Project. This initiative is part of a collaborative effort with nearly 20 partners aimed at strengthening the AI chip infrastructure in China. During a conference titled "Intelligent Transformation and Symbiosis," Yang Fan, co-founder of SenseTime and president of the business group for large devices, presented an integrated vision where chip technology, strategic partnerships, and commercial deployment converge to boost locally produced AI computing power.

In parallel, SenseTime has signed a spatial computing agreement with Guoxing Aerospace, a satellite manufacturer, and established research partnerships with five institutions, including the Shanghai Artificial Intelligence Laboratory. These collaborations aim to explore applications of scientific computing, thereby enhancing the Chinese technological ecosystem.

Yang Fan highlighted three major trends currently converging: the increasing demand for tokens in enterprises, the growing adoption of industrial AI compared to consumer-oriented applications, and the maturity of domestic chips enabling the rapid creation of intelligent computing centers based on Chinese silicon. However, these promising prospects rely on figures that SenseTime has yet to have independently verified.

Ambitious but Unverified Forecasts

SenseTime claims that its large-scale device platform currently processes an average of 2.42 trillion tokens per day. The company forecasts a dramatic increase in this figure, aiming for 10 trillion tokens per day by the fourth quarter of 2026. It is crucial to note that these figures are projections and not measured results. Companies interested in SenseTime's infrastructure should therefore await the publication of quarterly figures to assess the reality of these ambitions.

SenseTime's claims of profitability also rely on its own statements. The company asserts that its heterogeneous hybrid inference technology improves the utilization of model FLOPs by 85 to 152% on traditional domestic chips, with an estimated inference profitability 1.25 times that of Nvidia's H series chips. In comparison with domestic homogeneous inference configurations, SenseTime claims that its technology allows for a 2.5 times increase in token production at equivalent cost. This would place hybrid inference clusters beyond the minimum profitability threshold for national computing power.

However, these figures have not been verified by third parties, and the gap between results obtained in an optimized test environment and those in a real production environment can be significant.

Chip Adaptability and Multi-Chip Challenges

Domestic AI chips have often faced challenges related to a fragmented software stack, requiring adjustments to function across different architectures. SenseTime claims to have developed a comprehensive adaptation layer that encompasses models, frameworks, operators, toolchains, and hardware. This innovation aims to enable clients to migrate their workloads between different chip vendors without requiring significant rewrites.

Two examples illustrate this advancement. In the context of a long-sequence protein prediction workload AI4S, SenseTime optimized the fused operators, reducing overall prediction time by a factor of three. In the field of AIGC video generation, the company claims a 93% multi-card parallel acceleration ratio for domestic chips running DiT models, with cost-free migration for traditional AI development tools.

However, these promising results need to be confirmed in real production environments, where challenges related to mixed hardware generations may prove more complex.

A New Metric for Energy Efficiency

SenseTime has introduced a new metric, Tokens per Watt, to evaluate the efficiency of AI data centers. This measure is accompanied by a Computing Power Collaboration Agent that manages resource scheduling, electricity price forecasting, and energy storage optimization. The company claims to have improved token production per unit of electricity cost by 80%, with average electricity prices 10% lower than those of comparable regional data centers, and a 96% accuracy in forecasting computing load.

These claims will need to be closely monitored in the coming quarters, as electricity price arbitrage and load forecasting accuracy can vary significantly once a system is subjected to a full seasonal cycle with genuine demand volatility.

A Diverse Partner Network

SenseTime's Galaxy Project relies on a diverse ecosystem of partners, including domestic chip suppliers such as Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren Technology. The project also includes component companies like Xizhi Technology and infrastructure firms such as Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology.

SenseTime plans to build a "token factory," five "10,000 calorie" scale computing clusters, work in ten technological directions, and support 200 AI startups. Yang emphasized that domestic production is not limited to replacing individual chips but involves a collaborative effort across the entire chain of China's innovation capabilities.

Prospects for Spatial, Optical, and Quantum Computing

Beyond short-term infrastructure, SenseTime is exploring cutting-edge technologies such as optical computing to enhance data center efficiency, quantum computing applications for AI optimization, and a spatial computing partnership with Guoxing Aerospace. Together, they aim to create the SenseTime Spatial Computing Constellation, with the first satellite launch planned for 2026 and computing capacity reaching several tens of thousands of petabytes by 2030.

Yang explained that spatial computing could extend the reach of Chinese AI services in low-network environments, such as maritime operations and disaster response, thereby supporting China's AI exports internationally. However, this ambitious goal remains to be realized, as large-scale satellite computing deployments have yet to be precedent.

Physical Expansion from Shanghai to Riyadh

On the physical front, SenseTime operates the country's first 5A intelligent computing data center in Shanghai, capable of processing over 20 trillion tokens per day across more than 20 industries. Another site in Yancheng has been launched with an initial capacity of 3,000 petaflops, targeting sectors such as energy, manufacturing, and low-altitude economy.

In Hong Kong, SenseTime is building the largest national intelligent computing center in the territory, aiming for 40,000 petaflops by 2030. The company also plans to establish the first national computing cluster abroad in Saudi Arabia, intended to serve as a complete national computing base for the Middle East.

In the research domain, SenseTime collaborates with the Shanghai AI Laboratory, the Zhongguancun Academy of Beijing, the Hetao Academy of Shenzhen, the Shanghai Algorithm Innovation Research Institute, and the AI School of Shanghai Jiao Tong University. Together, they aim to build a shared platform for research in life sciences, materials science, and manufacturing. Yang described AI for science as a key lever for paradigm innovation in fundamental research, aligning this initiative with China's broader "Artificial Intelligence +" political strategy.

SenseTime's forecast of 10 trillion tokens per day by the fourth quarter of 2026 remains a target to closely monitor to assess the realization of these ambitions.