Summary of Key Points
This article discusses the claim by JPMorgan Chase that "80% of data centers in China are idle," clarifying a crucial misconception: Data centers are not equivalent to intelligent computing centers, and a single idle rate cannot generalize the status of all computing infrastructure. The article highlights the real contradiction in China's AI computing capacity: there is a shortage of high-quality, efficient computing power, while there is an excess of low-efficiency resources. Top companies still face shortages of advanced computing power, yet many newly built intelligent computing centers in various regions have low utilization rates due to issues with technology, ecosystem, and mismatched demand. The article also analyzes the shortcomings of the domestic computing ecosystem, the shift in demand from model training to inference, and the risk of treating computing resources as a form of "computing real estate." It concludes that China's AI computing industry has entered a new phase where efficiency, rather than scale, is the key factor.
Breakdown and Interpretation
1. The Misconception of "80% Idle"
JPMorgan Chase's figure of 80% idle capacity includes traditional IDCs (server hosting), cloud computing centers, supercomputing centers, and intelligent computing centers dedicated to AI, which is not accurate. It's like combining the vacancy rates of warehouses, offices, and laboratories and claiming that all buildings are empty—meaningless. What matters is the utilization rate of intelligent computing centers. For example, a certain thousand-watt intelligent computing center in the western region has an occupancy rate of less than 50%, with less than 30% of the installed servers actually in use, and annual operating costs amount to 30 million yuan. There are several ways to measure utilization: whether machines are installed (on-shelf rate), whether they are running (up-time rate), whether GPUs are being fully utilized (GPU usage rate), the actual efficiency of model training (effective training rate), and whether it is profitable (commercial utilization rate). Regardless of the metric, it points to the same issue: the amount of computing power purchased (book value) does not equal the power that can actually be used to generate value.
2. High Paper-Based Computing Power Is Useless
Local governments often boast about the number of intelligent computing centers, the number of cabinets, and the number of GPU cards installed, but these are merely figures on paper. Training large models requires more than just hardware; it's a systematic endeavor. For instance, a cluster with 1000 GPUs may have lower efficiency than a well-optimized cluster with 200 GPUs if the interconnect bandwidth is insufficient or the scheduling system is inefficient. Imagine buying 10 high-end computers without the necessary software or users to operate them; they will remain idle and unable to perform complex tasks together—this illustrates the gap between paper-based computing power and actual usable power.
3. The Shortcomings of Domestic Computing Chips
Even though domestic chips are now being used in some intelligent computing centers, the problem lies not in their capability but in the lack of a supporting ecosystem. After export restrictions on certain chips, engineers must re-optimize algorithms and adapt frameworks to use them, which can be more costly than purchasing imported chips. The international standard for AI computing is NVIDIA's CUDA ecosystem, which provides mature development tools and libraries, making it easier for developers to work with domestic chips. Without such an ecosystem, companies face additional challenges.
4. Changing Demand
The focus in the AI industry has shifted from training large models to daily inference tasks (e.g., AI customer service, image recognition, industry-specific agents). Training power can be located in areas with lower electricity costs (such as the western regions), but inference power needs to be close to users (e.g., in cities for faster responses). The Ministry of Industry and Information Technology's goal of a 1-millisecond latency circle for computing resources emphasizes the importance of placing inference centers near users. Intelligent computing centers built according to traditional training logic in remote areas will likely lack customers, similar to a large factory in the suburbs that serves no nearby residents.
5. Avoid Turning Computing Power into "New Real Estate"
Many regions treat intelligent computing centers as digital infrastructure projects, hoping to receive subsidies and improve their financial profiles. However, AI infrastructure is different from traditional infrastructure. While traditional projects can wait for demand to emerge, AI hardware depreciates quickly, and market needs change rapidly. Currently, there are over 500 intelligent computing centers in operation or under construction nationwide, with 222 projects costing over 100 million yuan each. If we continue the "build first, find customers later" approach, many of these centers will end up as unused "islands of computing power"—similar to vacant shopping malls that incur maintenance costs without tenants.
Conclusion
The real issue with China's AI computing capacity is not whether it exists but whether it can be effectively utilized. The future competition will not be about who has the most GPUs, but who can integrate hardware, ecosystems, electricity, and scheduling systems to turn computing power into productive assets. This will be the defining factor in the next stage of China's AI industry.