Recently, the 20th China IDC Industry Annual Conference (IDC2025) and Digital Infrastructure Technology Exhibition (DITExpo) were held in Beijing. Lin Hai, General Manager of the Intelligent Computing Center of SenseTime's Large Device Business Group, delivered a keynote speech titled "From Stability Guarantee to Efficiency Leadership - SenseTime's Collaborative Innovation Practice in Large Device Computing".

Lin Hai pointed out that from a global perspective, computing infrastructure is facing a new round of "computing power efficiency" strategic competition, including major technology giants continuously improving computing power efficiency through self-developed AI accelerators, data center liquid cooling, and renewable energy power collaboration. At the same time, represented by Palantir's "Chain Reaction" platform, computing power scheduling, power forecasting, and chip resource management are integrated into the same system, attempting to build a "national level AI infrastructure operating system". As a leading AI infrastructure in China, SenseTime's exploration in the field of computing power collaboration has provided a Chinese style "computing power energy" collaborative construction paradigm through systematic design, which will give China the technological initiative in the next stage of AI infrastructure competition.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Industry leading: Global promotion of "computing power energy" synergy, competition in the new era has begun

Lin Hai mentioned that recently, Palantir Technologies Inc. (NYSE: PLTR), a well-known big data company in the United States, officially released a product called "Chain Reaction" and positioned it as an "operating system for artificial intelligence infrastructure in the United States". This measure indicates that technology powerhouses represented by the United States have begun to systematically promote the integrated layout of "computing power chip energy" at the national strategic level, aiming to build an independent, controllable, and efficient collaborative national level artificial intelligence infrastructure system.

In the context of this global trend, SenseTime, as a pioneer in China's artificial intelligence infrastructure field, based on its long-term operation practice of its own 10000 card intelligent computing center (AIDC), officially released its independently controllable "Computer Computer Collaborative Intelligent Dispatch Platform" in July 2025, fully integrating computing power and power data. This is not only a technological project for enterprises to reduce costs and increase efficiency, but also a strategic issue related to whether China can take the initiative in infrastructure development in the era of artificial intelligence.

The exploration of SenseTime's equipment in the field of computer and electronics collaboration is a Chinese style innovation path driven by core technology and validated by practical results. Not only has it established a technological system that is in line with international advanced concepts, but it has also achieved significant economic and social benefits in practical operation, forming a differentiated Chinese experience and providing support for building a green, efficient, and safe development foundation for China's artificial intelligence industry.

Connect global data and build a complete system based on deep integration of computing and electronics for SenseTime's large-scale equipment

With the explosive demand for AI applications and the promotion of the "dual carbon" goal, electricity is becoming a key bottleneck restricting the expansion of computing power and green operation. Lin Hai pointed out that the new generation of intelligent computing centers is not simply traditional data centers that stack servers and compete for scale, but rather aims to achieve "deep collaboration between computing power and energy".

However, there is a common problem in the current industry where "model task data, cluster load data, and underlying power dispatch data" are disconnected from each other, and the settlement mechanism is not penetrated, resulting in data silos and difficulty in coordinating operating modes. To solve this structural problem, SenseTime focuses on a deeper level of infrastructure operation logic and has proactively built an overall architecture system of "IaaS+MaaS+computing power collaboration". It penetrates the entire data chain from bottom level wind, fire, water, electricity to top-level model tasks, and uses a "source network load storage" full chain computing power collaboration platform to achieve precise matching of computing power and energy.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Among them, with the energy big model as the core, breaking down data silos and achieving intelligent prediction and high-frequency scheduling; At the same time, we will work together with CATL to create a large-scale energy storage system with intelligent control capabilities, providing flexible and stable power support for computing load fluctuations.

Empowering Energy Big Models: Innovative Algorithm Architecture for More Accurate Load Forecasting

Shangtang's self-developed energy big model adopts a multimodal MoE architecture, trained on massive industry knowledge texts, energy structured data, computing power monitoring indicators and other data, and fully integrates the energy industry knowledge base. It can accurately predict the energy demand of complex scenarios and make accurate decisions:

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

? Open up the data loop and achieve precise mapping throughout the entire chain: Create a "computing power" mapping model, and use the unique "energy block" data model and "computing power consumption model" to connect the entire chain data from AI training tasks, computing power platforms, server hardware, and distribution systems, achieving precise mapping of "task computing power consumption".

Innovative "energy blocks" for precise prediction in complex scenarios: By binding energy intrinsic data, user energy consumption characteristics, energy balance rules, and other information with the computing power server as "energy blocks" as the basic token, and optimizing and adjusting the algorithm architecture based on multiple rounds of pre training results, the prediction accuracy and generalization ability in complex scenarios are comprehensively improved.

High frequency dynamic scheduling to achieve optimal energy balance: high-frequency predictions are made every 15 minutes, and decision correction iterations are made every 5 minutes. Based on real-time energy status and load forecasting, the optimal scheduling strategy is automatically generated and executed through cross system linkage to achieve accurate computing power prediction, load prediction, strategy generation, and correction.

Overall, the energy big model can predict the trend of computing power load in advance, and integrate factors such as electricity price signals, green electricity ratio, energy storage status, and grid demand for cross system joint dynamic solving, achieving active scheduling of "computing with electricity consumption and computing with electricity dynamics", and transforming data centers from "rigid loads" to "adjustable resources". At present, the accuracy of energy demand prediction based on energy block models has reached over 88%, and the accuracy of decision-making has reached over 93%. With the continuous iterative optimization of algorithms and energy storage devices, the prediction accuracy will reach an industry-leading level of 90% to 95%, and the decision accuracy will exceed 95%.

Intelligent energy storage system online: millisecond level response, solving the problem of peak fluctuations

On the energy storage side, SenseTime and CATL have jointly developed a new energy storage system with a scale of 17.88MW/35.776MWh, and endowed it with intelligent management capabilities. It has been specially designed for peak fluctuation scenarios of large model training and inference, which can effectively cope with the instantaneous power gap and peak fluctuations of computing power clusters during high load stages, while participating in peak shaving and valley filling and electricity market transactions, directly creating economic value.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Create a "power buffer pool" for the intelligent computing center: This system has millisecond level response capability and can provide instantaneous power support during the start-up of a 10000 card cluster and sudden load increases, effectively responding to the impact of computing load fluctuations and ensuring stable operation of the cluster.

Seasonal scheduling to achieve a balance between safety and efficiency: In winter, spring, and autumn when PUE is low, the traditional two charging and two discharging mode is adopted; In the high PUE and low redundancy summer, switch to the intelligent scheduling mode driven by the energy big model to ensure efficient operation of the system within the safety red line.

Shanghai Lingang AIDC becomes a model in China: the first to run a two-way closed-loop system of 'computing with electricity usage, and computing with electricity movement'

Thanks to the system level collaborative optimization of computing and electricity, Shanghai Lingang AIDC, the first 5A level intelligent computing center in China built and self owned by SenseTime, can not only automatically optimize computing power scheduling according to changes in computing power load, but also predict electricity demand through energy big models, intelligently control energy storage systems to achieve peak shaving and valley filling, continuously reduce energy consumption and electricity costs while ensuring stable operation, and successfully achieve a two-way closed-loop of "computing with electricity consumption and electricity with computing". Thanks to its forward-looking practices and innovative achievements in the field of computer collaboration, SenseTime's Shanghai Lingang Intelligent Computing Center has won the "2025 China IDC Industry Computer Collaboration Pioneer Award".

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

At present, Shanghai Lingang AIDC has achieved an overall PUE reduction of 1.267, a 3% reduction in PUE compared to the design value, an annual energy saving of over 10 million kWh, an annualized electricity cost savings of 7%, and a carbon reduction of 3000 tons, bringing significant economic and social benefits and becoming a "model room" for green intelligent computing centers.

In the future, SenseTime will use energy big models and system level algorithms as the core engine to deepen industrial cooperation, create a next-generation AI infrastructure base for the big model era, and provide sustained momentum for industry cost reduction, efficiency improvement, and green development.

Recently, the 20th China IDC Industry Annual Conference (IDC2025) and Digital Infrastructure Technology Exhibition (DITExpo) were held in Beijing. Lin Hai, General Manager of the Intelligent Computing Center of SenseTime's Large Device Business Group, delivered a keynote speech titled "From Stability Guarantee to Efficiency Leadership - SenseTime's Collaborative Innovation Practice in Large Device Computing".

Lin Hai pointed out that from a global perspective, computing infrastructure is facing a new round of "computing power efficiency" strategic competition, including major technology giants continuously improving computing power efficiency through self-developed AI accelerators, data center liquid cooling, and renewable energy power collaboration. At the same time, represented by Palantir's "Chain Reaction" platform, computing power scheduling, power forecasting, and chip resource management are integrated into the same system, attempting to build a "national level AI infrastructure operating system". As a leading AI infrastructure in China, SenseTime's exploration in the field of computing power collaboration has provided a Chinese style "computing power energy" collaborative construction paradigm through systematic design, which will give China the technological initiative in the next stage of AI infrastructure competition.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Industry leading: Global promotion of "computing power energy" synergy, competition in the new era has begun

Lin Hai mentioned that recently, Palantir Technologies Inc. (NYSE: PLTR), a well-known big data company in the United States, officially released a product called "Chain Reaction" and positioned it as an "operating system for artificial intelligence infrastructure in the United States". This measure indicates that technology powerhouses represented by the United States have begun to systematically promote the integrated layout of "computing power chip energy" at the national strategic level, aiming to build an independent, controllable, and efficient collaborative national level artificial intelligence infrastructure system.

In the context of this global trend, SenseTime, as a pioneer in China's artificial intelligence infrastructure field, based on its long-term operation practice of its own 10000 card intelligent computing center (AIDC), officially released its independently controllable "Computer Computer Collaborative Intelligent Dispatch Platform" in July 2025, fully integrating computing power and power data. This is not only a technological project for enterprises to reduce costs and increase efficiency, but also a strategic issue related to whether China can take the initiative in infrastructure development in the era of artificial intelligence.

The exploration of SenseTime's equipment in the field of computer and electronics collaboration is a Chinese style innovation path driven by core technology and validated by practical results. Not only has it established a technological system that is in line with international advanced concepts, but it has also achieved significant economic and social benefits in practical operation, forming a differentiated Chinese experience and providing support for building a green, efficient, and safe development foundation for China's artificial intelligence industry.

Connect global data and build a complete system based on deep integration of computing and electronics for SenseTime's large-scale equipment

With the explosive demand for AI applications and the promotion of the "dual carbon" goal, electricity is becoming a key bottleneck restricting the expansion of computing power and green operation. Lin Hai pointed out that the new generation of intelligent computing centers is not simply traditional data centers that stack servers and compete for scale, but rather aims to achieve "deep collaboration between computing power and energy".

However, there is a common problem in the current industry where "model task data, cluster load data, and underlying power dispatch data" are disconnected from each other, and the settlement mechanism is not penetrated, resulting in data silos and difficulty in coordinating operating modes. To solve this structural problem, SenseTime focuses on a deeper level of infrastructure operation logic and has proactively built an overall architecture system of "IaaS+MaaS+computing power collaboration". It penetrates the entire data chain from bottom level wind, fire, water, electricity to top-level model tasks, and uses a "source network load storage" full chain computing power collaboration platform to achieve precise matching of computing power and energy.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Among them, with the energy big model as the core, breaking down data silos and achieving intelligent prediction and high-frequency scheduling; At the same time, we will work together with CATL to create a large-scale energy storage system with intelligent control capabilities, providing flexible and stable power support for computing load fluctuations.

Empowering Energy Big Models: Innovative Algorithm Architecture for More Accurate Load Forecasting

Shangtang's self-developed energy big model adopts a multimodal MoE architecture, trained on massive industry knowledge texts, energy structured data, computing power monitoring indicators and other data, and fully integrates the energy industry knowledge base. It can accurately predict the energy demand of complex scenarios and make accurate decisions:

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

? Open up the data loop and achieve precise mapping throughout the entire chain: Create a "computing power" mapping model, and use the unique "energy block" data model and "computing power consumption model" to connect the entire chain data from AI training tasks, computing power platforms, server hardware, and distribution systems, achieving precise mapping of "task computing power consumption".

Innovative "energy blocks" for precise prediction in complex scenarios: By binding energy intrinsic data, user energy consumption characteristics, energy balance rules, and other information with the computing power server as "energy blocks" as the basic token, and optimizing and adjusting the algorithm architecture based on multiple rounds of pre training results, the prediction accuracy and generalization ability in complex scenarios are comprehensively improved.

High frequency dynamic scheduling to achieve optimal energy balance: high-frequency predictions are made every 15 minutes, and decision correction iterations are made every 5 minutes. Based on real-time energy status and load forecasting, the optimal scheduling strategy is automatically generated and executed through cross system linkage to achieve accurate computing power prediction, load prediction, strategy generation, and correction.

Overall, the energy big model can predict the trend of computing power load in advance, and integrate factors such as electricity price signals, green electricity ratio, energy storage status, and grid demand for cross system joint dynamic solving, achieving active scheduling of "computing with electricity consumption and computing with electricity dynamics", and transforming data centers from "rigid loads" to "adjustable resources". At present, the accuracy of energy demand prediction based on energy block models has reached over 88%, and the accuracy of decision-making has reached over 93%. With the continuous iterative optimization of algorithms and energy storage devices, the prediction accuracy will reach an industry-leading level of 90% to 95%, and the decision accuracy will exceed 95%.

Intelligent energy storage system online: millisecond level response, solving the problem of peak fluctuations

On the energy storage side, SenseTime and CATL have jointly developed a new energy storage system with a scale of 17.88MW/35.776MWh, and endowed it with intelligent management capabilities. It has been specially designed for peak fluctuation scenarios of large model training and inference, which can effectively cope with the instantaneous power gap and peak fluctuations of computing power clusters during high load stages, while participating in peak shaving and valley filling and electricity market transactions, directly creating economic value.

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

Create a "power buffer pool" for the intelligent computing center: This system has millisecond level response capability and can provide instantaneous power support during the start-up of a 10000 card cluster and sudden load increases, effectively responding to the impact of computing load fluctuations and ensuring stable operation of the cluster.

Seasonal scheduling to achieve a balance between safety and efficiency: In winter, spring, and autumn when PUE is low, the traditional two charging and two discharging mode is adopted; In the high PUE and low redundancy summer, switch to the intelligent scheduling mode driven by the energy big model to ensure efficient operation of the system within the safety red line.

Shanghai Lingang AIDC becomes a model in China: the first to run a two-way closed-loop system of 'computing with electricity usage, and computing with electricity movement'

Thanks to the system level collaborative optimization of computing and electricity, Shanghai Lingang AIDC, the first 5A level intelligent computing center in China built and self owned by SenseTime, can not only automatically optimize computing power scheduling according to changes in computing power load, but also predict electricity demand through energy big models, intelligently control energy storage systems to achieve peak shaving and valley filling, continuously reduce energy consumption and electricity costs while ensuring stable operation, and successfully achieve a two-way closed-loop of "computing with electricity consumption and electricity with computing". Thanks to its forward-looking practices and innovative achievements in the field of computer collaboration, SenseTime's Shanghai Lingang Intelligent Computing Center has won the "2025 China IDC Industry Computer Collaboration Pioneer Award".

IDCC 2025: The world enters a period of competition in computing power efficiency, and SenseTime offers a Chinese style solution

At present, Shanghai Lingang AIDC has achieved an overall PUE reduction of 1.267, a 3% reduction in PUE compared to the design value, an annual energy saving of over 10 million kWh, an annualized electricity cost savings of 7%, and a carbon reduction of 3000 tons, bringing significant economic and social benefits and becoming a "model room" for green intelligent computing centers.

In the future, SenseTime will use energy big models and system level algorithms as the core engine to deepen industrial cooperation, create a next-generation AI infrastructure base for the big model era, and provide sustained momentum for industry cost reduction, efficiency improvement, and green development.