SenseTime launched the Galaxy project in collaboration with around 20 partners to expand China’s domestic AI chip infrastructure.
In a keynote speech titled “Intelligent Transformation and Symbiosis,” Yan Huang, co-founder and president of the company’s Large Devices Business Group, explained SenseTime’s closed loop connecting chip-level technology, ecosystem partnerships, and commercial deployment of domestic AI computing power.
In parallel with the Galaxy project, SenseTime has signed a space computing agreement with satellite manufacturer Guoxing Aerospace and research partnerships with five institutions, including the Shanghai Institute of Artificial Intelligence for scientific computing applications.
Yang framed the timing around three converging trends: increasing demand for tokens across enterprise deployments, industrial AI adoption catching up with consumer use cases, and domestic chip commercialization reaching a stage where intelligent computing centers built on Chinese silicon can ramp up.
But whether that window is as open as SenseTime claims depends largely on numbers that the company hasn’t independently verified.
Token throughput numbers are marked with a large asterisk
According to SenseTime, its large-scale device platform currently processes an average of 2.42 trillion tokens per day, and the company predicts that number will increase 25x to 10 trillion tokens per day by the fourth quarter of 2026. This is a forecast, not a measurement, and enterprise buyers evaluating SenseTime’s infrastructure should treat it as such until quarterly numbers begin to be released.
Cost-benefit claims associated with that growth are similarly self-reported. According to SenseTime, the company’s heterogeneous hybrid inference technology increases model FLOP utilization by 85 to 152 percent on domestic mainstream chips, with the company making inference cost-effective 1.25 times that of Nvidia’s H-series parts.
Compared to domestic homogeneous inference setups, SenseTime claims a 2.5x increase in token output at comparable cost, and the company says that this leap takes the optimized hybrid inference cluster beyond what the industry previously considered the lowest threshold for profitability of domestic computing power.
None of these numbers come from third-party benchmarks, and the gap between a vendor’s optimized test cluster and a customer’s production environment, which has uneven data pipelines and slow firmware updates, tends to ease these numbers.
Adaptability claims and multichip problems
Domestic AI chips have historically struggled with fragmented software stacks. Models trained for one architecture often require rework to run on another. SenseTime says it has addressed this by building a full-stack adaptation layer that spans models, frameworks, operators, toolchains, and hardware, with the goal of allowing customers to migrate workloads between domestic chip vendors without major rewrites.
The company cites two application examples. For AI4S long-sequence protein prediction workloads, SenseTime says fusion operator optimization reduces overall prediction time by a factor of three. AIGC video generation claims a multi-card parallel acceleration rate of 93% for domestic chips running DiT models, in addition to what it describes as a zero-cost migration of mainstream AI development tools.
These numbers are the kind of things that read well in sandbox testing, and become even more important when stress testing against real customer pipelines running mixed hardware generations.
Energy index gets new benchmark name
SenseTime has introduced a metric called Tokens Per Watt, which the company positions as an alternative metric for measuring AI data center efficiency, alongside a computing power collaboration agent that handles resource scheduling, power price forecasting, and energy storage optimization across an eight-level data system with five decision chains.
SenseTime claims that by combining computing, electricity pricing, and automatic scheduling, it has increased token output per unit of electricity by 80 percent, average electricity prices are 10 percent lower than comparable regional data centers, and compute load forecasting accuracy is 96 percent.
Rather than accepting them at face value, these claims are worth keeping an eye on over the coming quarters. The accuracy of power price arbitrage and load forecasting tends to perform differently when the system runs through a full seasonal cycle with real demand fluctuations, rather than the conditions under which vendors typically run pilots.
An impressive roster of partners ranging from chip manufacturers to component suppliers
The Galaxy Project’s stated ecosystem includes domestic chip vendors Cambricon, Muxi, Hygon, Huawei Ascend, Moore Threads, Sunrise, Biren Technology, component partner Xizhi Technology, and infrastructure companies such as Silicon Motion, Qijing Technology, Zhongke Jiahe, Qingcheng Jizhi, Sophon Information, and Jiliu Technology.
SenseTime said the plan includes building one “token factory,” building five computing clusters it calls “10,000 calorie” scale, collaborating across 10 technology directions, and supporting 200 AI startups.
“Domestic production is not just replacing individual chips, but a collaborative effort across the entire chain of China’s innovation capabilities, from chips and components to infrastructure and application scenarios,” Yang said.
Bets on space, light and quantum computing look further ahead
Beyond short-term infrastructure, SenseTime outlined its work on optical computing for data center efficiency, applications of quantum computing in AI optimization, and a space computing partnership with Guoxing Aerospace to build what the companies are calling the SenseTime Space Computing Constellation.
SenseTime’s plan is to launch its first satellite in 2026 and aim to build thousands of computing satellites and tens of thousands of petabytes of computing power by 2030.
Yang argued that the value goes beyond its inherent capabilities, framing space-based computing as a way to extend the scope of China’s AI services to vulnerable network environments such as maritime operations and disaster response, and in turn support China’s international exports of AI.
The 2030 goal is five years away, and there is no precedent for deploying satellite computing on this scale with a timeline.
Physical infrastructure spans from Shanghai to Riyadh
On the ground, SenseTime says its Shanghai facility operates the country’s first data center rated at the Intelligent Computing level known as “5A” and processes more than 20 trillion tokens daily across more than 20 industries. The Yancheng site launched with an initial capacity of 3,000 petaflops focused on energy, manufacturing, and low-altitude economy applications.
SenseTime is building what it calls the region’s largest domestic intelligent computing center in Hong Kong, with a target of 40,000 petaflops by 2030. The company is also planning what it calls China’s first overseas domestic computing cluster in Saudi Arabia, positioned as a full-stack domestic computing base for the Middle East.
On the research front, SenseTime’s partnership with Shanghai AI Institute, Beijing Zhongguancun Academy, Shenzhen Hetao Academy, Shanghai Algorithm Innovation Institute, and Shanghai Jiao Tong University AI School aims to build a shared platform across computing, tools, and modeling capabilities for life sciences, materials science, and manufacturing research. Yang calls AI for Science “an important vehicle for paradigm innovation in basic research” and links the effort to China’s broader “artificial intelligence+” policy push.
SenseTime’s prediction of 10 trillion tokens per day by Q4 2026 is a number that will be tracked against what the company reports when the quarter actually ends.
SEE ALSO: Kimi K3 weightless model: China’s biggest AI is betting on memory, not compute
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber Security & Cloud Expo. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
