The rapid development of AI technologies in China is not without its excesses. According to Datacenter Knowledge, a large number of data centers for AI systems have been created in the country, but their capacity turned out to be excessive, and they themselves were not best suited for such loads.
According to Chinese media, conducting their own investigations, local operators created too large capacities, overestimating demand. The industry is also negatively affected by the lack of sufficient competencies in creating and managing complex AI infrastructures. AI applications require a different data center architecture that more tightly integrates computing, storage, and networking at the hardware and software levels. In addition, such objects consume approximately four times more energy compared to traditional CPU-based data centers.
Image source: Nadiia Ganzhyi/unsplash.com
At the same time, operating an average AI data center with 1,250 servers in China is quite expensive, up to ¥1 billion ($141 million) per year. The main costs are electricity and depreciation. Last year, AI Technology Review experts concluded that forecasts for the need for a large number of data centers were not entirely correct and the demand was overestimated. As a result, this led to server downtime and, in some cases, even the closure of AI centers.
By 2025, China plans to commission 50 computing centers for artificial intelligence systems; in four years, the country’s computing capabilities should increase by a third. According to some estimates, 70 new centers are already being built. Representatives of Chinese government agencies revealed that over the past year the authorities spent $6.1 billion on the construction of large data center campuses, and another $28 billion was invested in the industry by private companies.
However, the program to move large data centers to the western regions of the country, for example, to Inner Mongolia, where there is a surplus of energy and labor and land are cheaper than in the east, has not fully justified itself. Customers from the densely populated eastern regions prefer to use the new campuses in the west solely for data storage, but not data processing. This is hampered by high access latency, as well as the high cost of connection – about ¥160 thousand ($22.5 thousand) per month for a dedicated channel with a width of 1 Gbit/s.