Preface: While global tech giants fiercely compete for multi-billion dollar AI chip orders, China's computing power infrastructure has undergone a quiet transformation. From blindly pursuing scale expansion to focusing on industry application implementation, from the consumer internet to the new frontier of embodied intelligence, a revolution in computing power, driven by robotics and liquid cooling technology, is reshaping the industrial landscape.

Author | Fang Wensan, Image Source | Network

From Frenzy to Rationality: A Strategic Shift in Computing Power Construction

The global competition for AI computing power once fell into the trap of an "arms race." The US "Stargate" plan invested $500 billion in building a new generation of AI infrastructure, while China's "East Data to West Computing" project spurred the construction of over 60 smart computing centers. Amidst the frenzy of scale expansion, the issue of underutilization in some smart computing centers gradually emerged.

Li Bin, Senior Vice President of Sugon, noted: "The scarcity of computing power is a dynamic phenomenon. In the past, construction focused excessively on hardware scale, neglecting application implementation and software ecosystems." This realization is sparking a fundamental shift in construction logic—moving from "build first, think later" to demand-driven production.

UCloud's Ulanqab Smart Computing Center exemplifies this trend. Demand for the 2,300 12kW high-power cabinets delivered in its third-phase project is robust, with the core focus being on precisely meeting customer needs. Sugon's Intelligent Computing Product Business Unit has shifted from "building roads" to vertical industry integration, emphasizing the practical utilization of computing power.

At the policy level, structural optimization is also being encouraged. The Central Cyberspace Administration Office clearly mandates "strengthening the integration of general computing, smart computing, and supercomputing to reduce the cost of computing power use." The National Data Administration is deepening reforms related to data elements and reducing the computing power threshold for small and medium-sized enterprises through various avenues, transforming computing power from a luxury into an inclusive resource.

The Rise of Embodied Intelligence: Robots Reshape the Geography of Computing Power

As the market continues to discuss the demand for large-scale model computing power, a new force is quietly redefining the computing power landscape. The robot economy is driving unprecedented growth in computing power requirements.

Tian Feng, Dean of the Institute of Fast Thinking and Slow Reflection, predicts: "Intelligent agents such as robot dogs, humanoid robots, and drones will fundamentally alter the regional competition landscape for computing power."

Sugon has received storage infrastructure orders from embodied intelligence enterprises like Zhiyuan Robotics and Unitree Technology, signaling the beginning of this transformation.

Embodied intelligence enterprises bring three core demands: massive perceptual data processing, distributed object technology to establish multi-modal data retrieval capabilities on a scale of tens of billions; algorithm training performance, supporting private cloud platforms for rapid iteration; and optimization of migration costs, shifting from public clouds to customized private clouds. These demands are reshaping the geography of computing power.

Robot spinning workshops in Xinjiang, autonomous driving logistics fleets in Tibet, and drone plant protection networks in Yunnan, Guizhou, and Sichuan—these non-traditional hotspots for computing power have given rise to new edge computing hubs due to the deployment of robot clusters.

China Mobile's planned "centralized + distributed" reasoning computing power network is well-timed, with its "central training, edge inference" system perfectly aligned with the computing power characteristics of the robotics industry.

Liquid Cooling Revolution: A Key Technology for Sustainable Computing Power Development

Liquid cooling technology is not new, but the explosion of AI computing power has propelled it onto the fast track of industrialization. IDC data shows that the Chinese liquid-cooled server market will continue to grow rapidly in 2024, with a market size of $2.37 billion, representing a 67.0% increase compared to 2023. Among these, the market share of cold plate solutions has further increased. IDC predicts that from 2024 to 2029, the annual compound growth rate of China's liquid-cooled server market will reach 46.8%, with a market size of $16.2 billion in 2029.

Experts state that while liquid cooling technology has been around for a long time, its recent surge in popularity coincides with the widespread adoption of artificial intelligence. Due to the high computing power demands of AI, particularly the current primary method of supplying computing power centered around accelerators, their thermal sensitivity is exceptionally high.

At the same time, the demand for computing power related to large AI models is immense, and the combination of these two factors has reciprocally accelerated the rapid development of the entire liquid cooling infrastructure. Liquid cooling will also be a core development trend for AI computing power infrastructure in the future.

As AI chips with a single card power consumption exceeding 1200W become the norm, traditional air cooling has reached its technical limit. Liquid cooling technology has leaped from a laboratory solution to a cornerstone of AI computing power infrastructure.

Statistics indicate that 40% of data center electricity consumption is dedicated to cooling, and the limitations of traditional air cooling methods in large-scale cooling are becoming increasingly apparent. In contrast, liquid cooling technology, with its liquid thermal conductivity 25 times that of air and cooling capacity reaching 1000 to 3000 times that of air, has become the new industry favorite.

Three major technical routes are reshaping the form of data centers: cold plate liquid cooling with a PUE value of 1.15, saving 40% of energy, applicable in the renovation of existing computer rooms, reducing the cost of a single cabinet by 30%; immersion liquid cooling with a PUE value of 1.05, saving 40% of electricity, suitable for high-power density clusters, with larger scales yielding more significant advantages; and spray liquid cooling, offering chip-level precise temperature control, applicable to cutting-edge scientific research computing, but with a higher initial investment.

The Essence of Industrial Transformation: Reconstructing Value Logic

As the construction of computing power shifts from scale orientation to effect orientation, the entire ecological chain is being reshuffled.

Operators are the first to adjust their strategies. China Mobile is constructing an ultra-large-scale "computing power factory," conducting forward-looking research on a smart computing center with 100,000 cards, and forming a smart computing system that is "centrally intensive and edge ubiquitous."

Its Wuhu cluster in the Yangtze River Delta hub has established the country's first "four-in-one" computing power scheduling platform, achieving hundreds of millions of computing power invocations daily and increasing resource utilization by 35%.

Small and medium-sized enterprises have obtained new market entry tickets. The "distillation model" solution pioneered by the 360 Group enables ordinary computers to run high-performance AI models; the open-source ecosystem lowers the threshold for using large models, combined with optimized computing power leasing, reducing the computing power costs of small and medium-sized enterprises by over 60%.

The service model has undergone a fundamental transformation. Traditional IDC resource-based services are being upgraded to task-based services, with Model as a Service (MaSS) becoming the new standard.

The national hub computing power scheduling platform utilizes intelligent matching algorithms to automatically allocate molecular simulation computing tasks of Shanghai biomedical enterprises to the Gansu clean energy cluster, reducing computing costs by 25% while meeting carbon neutrality requirements.

Conclusion: The Co-evolution of Industry and Machines Rewrites Market Rules

When tens of thousands of sensors in a Xinjiang robot spinning factory process data through Sugon's immersion liquid cooling cluster, and when the Chengdu drone logistics network relies on the Ulanqab Smart Computing Center to achieve millisecond-level path optimization, AI computing power has transitioned from abstract technical indicators to a hematopoietic system for embodied intelligence.

The off-grid smart computing center designed by Qinhuai Data has begun to harness energy from nuclear fusion experimental reactors, with the cooling liquid flowing in the liquid cooling pipes dissipating 150 watts of heat per square centimeter.

And within these liquid-cooled server clusters, a new generation of robot training models is iterating and upgrading at a speed surpassing human capability.

When the technological wave recedes, only the computing power ecosystem rooted in applications can transcend cycles, which may be the most crucial lesson left to the industry by the AI revolution.

In the next decade, the winners in the intelligent world will be pragmatic innovators who lay liquid cooling pipes into workshops, integrate computing power nodes into pastures, and truly bring robots to life.

Content sourced from: 36Kr Friends: Smart Computing Center Shifts from Scale Priority to Application Orientation, Robots Spur New Demand for Computing Power; Tian Feng Said: Robots Spur New Demand for Computing Power