**SenseTime’s Galaxy Project: A Bold Push to Build China’s Domestic AI Chip Ecosystem**
Chinese artificial intelligence (MS) company SenseTime has officially launched the Galaxy Project, an ambitious initiative designed to scale domestic AI chip infrastructure within China. The move represents a significant step in the country’s broader strategy to reduce reliance on foreign semiconductor technologies and accelerate the adoption of indigenous AI computing power.
At the center of the launch was a keynote presentation titled **“Intelligent Transformation and Symbiosis”** delivered by Yang Fan, co-founder of SenseTime and president of its Large Device Business Group. In his address, Yang outlined a comprehensive vision centered on creating a **closed-loop ecosystem** that connects chip-level innovation, strategic partnerships, and real-world commercial deployment.
The Galaxy Project is not operating in isolation. SenseTime has also entered into a **space computing agreement** with satellite manufacturer Guoxing Aerospace and established a **research partnership with five leading institutions**, including the Shanghai Artificial Intelligence Laboratory. These collaborations are focused on advancing scientific computing applications, signaling SenseTime’s intent to expand beyond traditional AI training and inference into high-performance and specialized computing domains.
According to Yang, the timing of the Galaxy Project is driven by three key trends:
– Rising demand for AI tokens across enterprise applications
– The growing adoption of industrial AI, moving beyond consumer-facing tools
– The increasing maturity of domestic chip commercialization, enabling faster deployment of intelligent computing centers built on Chinese-designed silicon
However, as with many emerging technology initiatives, the true test will be execution. The window of opportunity that SenseTime describes depends heavily on factors that have yet to be independently verified.
### Token Throughput Figures Come with a Large Asterisk
SenseTime claims its large-scale device platform currently processes an average of **2.42 trillion tokens per day**, with a projected increase to **10 trillion tokens per day by Q4 2026**. While these numbers are impressive on paper, they remain **forecasts rather than verified results**. Enterprise buyers considering SenseTime’s infrastructure should treat these projections with caution until quarterly performance data becomes available.
The company also asserts that its **heterogeneous hybrid inference technology** delivers:
– An 85–152% improvement in Model FLOPs Utilisation (MFU) on mainstream domestic chips
– Inference cost-effectiveness that is **1.25 times better** than Nvidia’s H-series GPUs
– A **2.5x increase in token output** compared to domestic homogeneous setups at the same cost
These claims suggest significant efficiency gains, but they are based on internal benchmarks rather than third-party testing. In real-world deployments, performance often varies due to data pipeline inefficiencies, firmware delays, and workload complexity.
### Adaptability Claims and the Multi-Chip Challenge
One of the long-standing weaknesses of domestic AI chips in China has been **software fragmentation**. Models trained on one architecture often require substantial rework to run on another. SenseTime claims to have addressed this through a **full-stack adaptation layer** that spans models, frameworks, operators, toolchains, and hardware.
The company cites two practical examples:
– In **AI for Science (AI4S)**, fused operator optimization reduced long-sequence protein prediction time by **3x**
– In **AIGC video generation**, SenseTime reports a **93% multi-card parallel acceleration ratio** for DiT models running on domestic chips, along with **zero-cost migration** for mainstream AI development tools
While these results are promising, their impact will only be fully understood under large-scale, production-level workloads involving mixed hardware generations.
### Energy Metrics and a New Benchmark: Tokens Per Watt
SenseTime has introduced a new performance metric called **Tokens Per Watt**, positioning it as a more relevant efficiency benchmark for AI data centers. Alongside this, the company launched a **Computing-Power Collaboration Agent** that handles:
– Resource scheduling
– Electricity price prediction
– Energy storage optimization
SenseTime claims this integrated approach delivers:
– An **80% increase in token output per unit of electricity cost**
– Average power prices **10% lower** than comparable regional data centers
– **96% accuracy** in computing load prediction
These figures are worth monitoring over several quarters. Energy arbitrage and forecasting accuracy behave differently under real seasonal demand fluctuations compared to pilot environments.
### A Broad Partner Ecosystem
The Galaxy Project features an extensive network of domestic partners, including:
– **Chipmakers**: Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, and Biren Technology
– **Component suppliers**: Xizhi Technology
– **Infrastructure firms**: Silicon Motion, Qujing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology
SenseTime aims to establish:
– One “token factory”
– Five computing clusters at “10,000-calorie” scale
– Collaboration across ten technology directions
– 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.”
### Looking Further Ahead: Space, Optics, and Quantum Computing
Beyond immediate chip deployment, SenseTime is investing in longer-term technologies:
– **Optical computing** for data center efficiency
– **Quantum computing** for AI optimization
– A **space computing constellation** in partnership with Guoxing Aerospace
The plan envisions a first satellite launch in 2026, eventually leading to thousands of computing satellites with tens of thousands of petabytes of capacity by 2030. The goal is not only to expand compute capacity but also to deliver AI services to low-connectivity environments such as maritime operations and disaster zones, while supporting China’s AI export ambitions.
### Global Infrastructure Footprint
SenseTime’s computing infrastructure is already geographically diversified:
– **Shanghai**: Hosts China’s first data center rated at “5A” intelligent computing level, processing over 20 trillion tokens daily across more than 20 industries
– **Yancheng**: A 3,000-petaflop facility focused on energy, manufacturing, and low-altitude economy
– **Hong Kong**: The territory’s largest domestic intelligent computing center, targeting 40,000 petaflops by 2030
– **Saudi Arabia**: Plans for China’s first overseas domestic computing cluster, serving as a full-stack base for the Middle East
On the research front, SenseTime has partnered with:
– Shanghai AI Laboratory
– Beijing Zhongguancun Academy
– Shenzhen Hetao Academy
– Shanghai Algorithm Innovation Research Institute
– Shanghai Jiao Tong University’s AI school
These collaborations aim to build a shared platform covering compute, tooling, and models, with a focus on life sciences, materials science, and manufacturing. Yang described AI for Science as “a key lever for paradigm innovation in basic research,” aligning with China’s “Artificial Intelligence+” policy initiative.
—
### FAQ
**What is the Galaxy Project?**
The Galaxy Project is SenseTime’s initiative to scale domestic AI chip infrastructure in China through a closed-loop system connecting chip technology, ecosystem partnerships, and commercial deployment.
**Who are SenseTime’s partners in this project?**
The ecosystem includes domestic chipmakers like Cambricon, Huawei Ascend, and Moore Threads, component suppliers, and infrastructure firms, along with support for 200 AI startups.
**What is Tokens Per Watt?**
It is a new metric introduced by SenseTime to measure AI data center efficiency, focusing on token output per unit of energy consumed.
**Are SenseTime’s performance claims verified?**
Many of the company’s claims, including token throughput and cost-effectiveness, are internally reported and lack third-party verification.
**What role does space computing play in SenseTime’s strategy?**
SenseTime is developing space-based computing capabilities through a partnership with Guoxing Aerospace, aiming to extend AI services to remote and low-connectivity regions.
—
### Conclusion
SenseTime’s Galaxy Project reflects a comprehensive and strategically timed effort to strengthen China’s domestic AI chip ecosystem. By integrating hardware innovation, software adaptability, energy efficiency metrics, and broad industry collaboration, the company is positioning itself as a central player in the nation’s AI infrastructure development. While many of its claims remain unverified and heavily dependent on successful execution, the scale of partnerships and long-term vision behind projects like space computing suggest that SenseTime intends to be more than a chip designer—it aims to define the backbone of China’s AI future. The industry will be watching closely when the promised token throughput and efficiency gains begin to show up in real-world deployments.
