China's plunge into AI has created a huge amount of excess data center capacity.

According to investigations by Chinese media, data center players have built too much capacity, lack the skills to carry out complex AI processes, and have overestimated demand.

China is planning to establish 50 intelligent computing centers by 2025, increasing its computing capacity by a third over four years – although according to one estimate, 70 new centers are already under construction.

In contrast to CPU-based legacy data centers, AI data centers integrate GPU and other types of chips for simultaneous processing of huge workloads. They consume as much as four times the energy of traditional data centers.

A report by AI Technology Review late last year complained the market had "blindly advanced" data center construction and had over-estimated demand, resulting in idle servers and in some cases the closure of intelligent computing centers.

It estimated an intelligent data center with 1,250 servers supporting large model training and inference could cost up to 1 billion Chinese yuan (US$141 million) a year to run, mostly in power costs and depreciation.

Storage, not compute

A big part of China's push into high-powered computing has been its East-West Computing Initiative, in which data generated in the high-income cities of the east can be processed by data centers in western provinces, taking advantage of low costs, clean energy and cheap land.

That was the plan, but the reality is that many of these western China facilities are operating at well under capacity, according to an investigation published by the official Xinhua news agency.

Demand from eastern seaboard customers for data services in the west was weak, with relatively few mature commercial applications, the study found. Instead of providing compute services, the main use case in the western China facilities was storage.

The new data centers are unable to support customers with low latency and complex processing requirements. 

Interviewees also blamed high network charges, with a 1Gbit/s dedicated line costing RMB160,000 ($22,500) a month, eliminating the advantages of the lower rental and power costs.

Another problem is the poorly architected AI data centers, which require intricate coordination across compute, storage and networking.

Sha Kaibo, a vice president of Tencent Cloud, compared it to buying a premium sports car but not understanding its advanced features and lacking a professional team to maintain it, with the result that "you can only drive it on ordinary roads."

This is not a problem confined to data centers. China's energy storage market is another case in point, also suffering from over-capacity and low usage.

Of course, the rest of the world is greatly exercised about China's over-invested manufacturing sector and whether it plans to dump its surplus abroad. 

The underlying issue is that over-capacity is endemic in the top-down economic system. The central government identifies priority sectors in which it offers easy credit and handouts, attracting eager enterprises and local governments with little regard for market conditions.