SenseTime Mobilizes Nearly 20 Partners for Ambitious AI Chip Infrastructure Expansion in China
SenseTime, a prominent AI company, has unveiled its “Galaxy Project,” a sweeping initiative designed to bolster domestic AI chip infrastructure within China. The project has garnered support from a coalition of nearly 20 partners, signaling a significant push towards self-reliance in AI computing power.
During a keynote address, the company’s co-founder and president of its Large Device Business Group outlined a comprehensive strategy focused on a closed-loop system. This system aims to seamlessly integrate chip-level advancements, robust ecosystem partnerships, and the commercial deployment of AI computing capabilities engineered in China.
Beyond the Galaxy Project, SenseTime has forged strategic alliances, including a space computing agreement with satellite manufacturer Guoxing Aerospace and a research collaboration with several academic institutions, such as the Shanghai Artificial Intelligence Laboratory, to advance scientific computing applications.
The timing of these initiatives appears strategically aligned with three key trends: a surge in demand for token processing across enterprise deployments, the growing adoption of industrial AI mirroring consumer-facing applications, and the maturation of domestic chip commercialization, enabling the rapid establishment of intelligent computing centers powered by Chinese silicon.
However, the extent to which this ambitious vision is achievable hinges on performance metrics that are yet to undergo independent verification.
**Token Throughput Claims Face Scrutiny**
SenseTime reports that its large-scale device platform currently processes an average of 2.42 trillion tokens daily. The company projects a substantial 25-fold increase to 10 trillion tokens per day by the fourth quarter of 2026. Investors and enterprise buyers will want to see sustained, verifiable quarterly figures to validate this ambitious forecast.
Similarly, the claimed cost-effectiveness of this projected growth warrants careful observation. SenseTime asserts that its heterogeneous hybrid inference technology achieves an 85–152 percent increase in Model FLOPs Utilization on mainstream domestic chips. Furthermore, the company claims its inference cost-effectiveness surpasses Nvidia’s H-series components by a factor of 1.25x.
When compared to domestic homogeneous inference setups, SenseTime posits a 2.5x improvement in token output at an equivalent cost. This, they argue, elevates optimized hybrid inference clusters beyond previous industry benchmarks for profitability in domestic computing power. It’s crucial to note that these claims are vendor-reported and lack independent third-party benchmarking. The practical performance in real-world customer environments, with their inherent complexities and variable updates, may differ from optimized test conditions.
**Addressing Adaptability and the Multi-Chip Challenge**
A persistent hurdle for domestic AI chips has been the fragmentation of their software stacks, often requiring significant rework for models trained on one architecture to function on another. SenseTime claims to have developed a full-stack adaptation layer encompassing models, frameworks, operators, toolchains, and hardware. The objective is to enable seamless workload migration across different domestic chip vendors without extensive re-engineering.
The company highlights two specific applications demonstrating this adaptability. In an AI4S long-sequence protein prediction task, SenseTime reports that fused operator optimization reduced prediction time by a factor of three. For AIGC video generation, they claim a 93 percent multi-card parallel acceleration ratio for domestic chips running DiT models, along with reportedly zero-cost migration for mainstream AI development tools. These performance figures will be more telling when rigorously tested against diverse customer pipelines operating with mixed hardware generations.
**Introducing a New Benchmark for Energy Efficiency**
SenseTime has introduced “Tokens Per Watt” as a novel metric for evaluating AI data center efficiency. Complementing this, they unveiled a Computing-Power Collaboration Agent designed for resource scheduling, electricity price forecasting, and energy storage optimization across an eight-level data system with five decision chains.
By integrating compute, electricity pricing, and automated scheduling, SenseTime claims an 80 percent increase in token output per unit of electricity cost. They also report average power prices 10 percent below comparable regional data centers and 96 percent accuracy in computing load prediction. These metrics are best assessed over several quarters, as the real-world performance under fluctuating demand and seasonal variations may differ from vendor-controlled pilot conditions.
**An Extensive Partner Network**
The Galaxy Project’s ecosystem includes a robust roster of domestic chip manufacturers such as Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren Technology. Component partners like Xizhi Technology and infrastructure firms including Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology are also part of the initiative.
SenseTime’s roadmap encompasses the construction of a “token factory,” five large-scale computing clusters, joint development across ten technological areas, and support for 200 AI startups.
As Yang stated, “Domestic production is not simply about replacing individual chips, but rather a collaborative effort across the entire chain of China’s innovation capabilities, from chips and components to infrastructure and application scenarios.”
**Long-Term Bets on Space, Optical, and Quantum Computing**
Beyond immediate infrastructure development, SenseTime is exploring advanced computing paradigms. This includes optical computing for enhanced data center efficiency, quantum computing applications for AI optimization, and a pioneering space computing partnership with Guoxing Aerospace to establish the SenseTime Space Computing Constellation.
The company envisions launching its first satellite in 2026, with the ultimate goal of deploying thousands of computing satellites and achieving tens of thousands of petabytes in computing capacity by 2030. This ambitious undertaking is framed not only as a technological advancement but as a means to extend the reach of Chinese AI services into challenging environments, such as maritime operations and disaster response, and to bolster China’s international AI exports. The 2030 target is a significant undertaking, with no direct precedent for satellite computing deployments of this magnitude.
**Global Physical Infrastructure Footprint**
On the terrestrial front, SenseTime’s Shanghai facility is noted as housing the nation’s first data center certified at an “5A” intelligent computing level, handling over 20 trillion tokens daily across more than 20 industries. A Yancheng site has commenced operations with an initial capacity of 3,000 petaflops, focusing on energy, manufacturing, and low-altitude economy applications.
In Hong Kong, SenseTime is developing what is set to be the territory’s largest domestic intelligent computing center, targeting 40,000 petaflops by 2030. The company also plans to establish China’s first overseas domestic computing cluster in Saudi Arabia, positioning it as a comprehensive domestic computing hub for the Middle East region.
Furthermore, SenseTime’s research collaborations with institutions like the Shanghai AI Laboratory are aimed at creating a shared platform for compute, tooling, and model development, particularly for advancements in life sciences, materials science, and manufacturing. Yang emphasized the role of AI in scientific discovery as a critical driver for innovation in basic research, aligning with China’s broader “Artificial Intelligence+” policy objectives.
The projected 10 trillion tokens per day by Q4 2026 remains the key performance indicator to monitor as SenseTime’s progress unfolds.
Original article, Author: Samuel Thompson. If you wish to reprint this article, please indicate the source:https://aicnbc.com/23958.html
