SemiAnalysis Publishes a Map of China’s AI Data-Center Footprint

SemiAnalysis has published its China Datacenter Model, a research tool that maps who owns Chinese data-center facilities, who uses them and how much capacity is operating, under construction or planned. The release puts hyperscalers, colocation providers, telecom carriers and selected AI labs into a single infrastructure picture.

The problem the model addresses is not simply counting buildings. AI capacity can sit in facilities that a company constructed itself, capacity it leases from another operator, or campuses whose owners and major tenants are different companies. SemiAnalysis says its hyperscaler views combine self-built and leased facilities for ByteDance, Alibaba, Tencent, Baidu, Huawei, Kuaishou, JD, Meituan and other companies.

From a building owner to the workload inside

SemiAnalysis says the model covers more than 60 Chinese data-center operators. For leading colocation providers, it lists the facilities they own and identifies key tenants behind major campuses. That distinction makes the model less a directory of real estate than a way to trace the relationship between the companies operating a site and the customers relying on it.

The model separates four infrastructure views

• Hyperscalers: combined self-built and leased data-center capacity.
• Colocation operators: owned facilities and key tenants at major campuses.
• Telecom carriers: major facilities owned by China Mobile, China Telecom and China Unicom.
• Selected AI labs: capacity estimated from installed chip counts and expressed as chip-implied megawatts.

The carrier layer has its own history

The inclusion of the three national telecom carriers matters because they have historically been central to the country’s data-center estate. SemiAnalysis says China Mobile, China Telecom and China Unicom previously operated more than 60% of China’s data-center projects. The model places those carrier-owned facilities alongside commercial colocation campuses and the infrastructure tied to large technology companies.

That structure is useful for a market in which ownership alone can obscure where computing is actually being used. A facility held by a colocation company may support a hyperscaler tenant; a hyperscaler’s footprint can also extend beyond the buildings it owns. By presenting both arrangements, the model treats leased capacity as part of a company’s infrastructure position rather than an afterthought.

An estimate where direct facility records are not the input

The AI-lab portion uses a different method. Rather than presenting a facility inventory for those selected labs, SemiAnalysis estimates capacity from installed chip counts and expresses the result in chip-implied megawatts. A megawatt is a measure of power; “chip-implied” signals that this part of the model is an estimate derived from hardware counts.

The next question is how readers will use those different kinds of entries together. The model offers status categories for operating, under-construction and planned capacity, while its AI-lab figures are chip-based estimates. Its value will rest on making the connections among owners, tenants, construction status and implied demand easier to examine without collapsing those categories into one number.