From grasslands to plateaus, a nationwide race for AI computing power is reshaping the country’s digital infrastructure
By Huang Xinyi
Deep in the grasslands of Inner Mongolia, construction sites hum with activity. Across the Qinghai plateau and the industrial heartlands of Heilongjiang, a new battle is being fought — not over land or resources, but over the computing power that will fuel the AI era.
Recent visits to computing centers in Ulanqab in Inner Mongolia, Haidong in Qinghai and Harbin in Heilongjiang show projects advancing rapidly across the country. In particular, infrastructure designed for AI model training is moving west, creating a new computing map stretching across China from east to west and north to south.
‘Orders are coming before construction starts’
Ulanqab has emerged as a major AI computing hub, attracting data center projects from companies including Huawei, Apple, Alibaba (9988.HK), Kuaishou (1024.HK) and Tencent (0700.HK).
DeepSeek is planning to build a large AI data center in Ulanqab, potentially adding about 1 gigawatt of computing capacity, with some capacity expected to come online by late 2027 or early 2028. Industry sources say that DeepSeek already leases data center capacity in the city.
UCloud’s Ulanqab Intelligent Computing Center covers about 140,000 square meters and is designed to accommodate roughly 12,000 server racks of different power configurations for large-model training and inference. UCloud (688158.SH) Chairman and CEO Ji Xinhua said demand was already extremely strong. “The utilization rate of completed data centers is very high, and orders are coming in even before new data centers have started construction,” he said.
Two thousand kilometres west, on the Qinghai plateau, the race is equally intense. Abundant clean energy and a cool climate have made the province China’s first pilot region for coordinating green electricity and computing power.
By the first half of 2026, Qinghai had built 49,760 standard racks and deployed computing capacity totaling 28,116petaflops (quadrillions of floating-point operations per second), covering general-purpose, intelligent and supercomputing.
China Mobile’s (0941.HK) Highland Big Data Center in Haidong, phase one of which opened in 2015, is the earliest and largest green facility on the Tibetan plateau. It has 7,153 standard racks and total IT power of 18 megawatts. It hosts more than 300 AI models. Since 2025, China Mobile Qinghai has invested more than 2 billion yuan ($297 million) to develop a “2+8+X” computing infrastructure network integrating computing, intelligent computing, storage and model resources.
The second phase of the Highland center, with a total investment of 2.5 billion yuan, is nearing completion. Its first stage will provide 14,000 racks, 35 MW of IT power and 32,400P of computing capacity.
Heilongjiang, which entered the computing race later, is also moving quickly. Harbin’s Digital Longjiang Intelligent Computing Center and China Mobile’s Harbin intelligent computing cluster, with more than 10,000 AI accelerators, began operating in 2024. More projects are now under construction, with total planned capacity exceeding 8,000 petaflops.
Power supply becomes a decisive factor
Cheap electricity is a critical reason companies are moving to these regions, especially Ulanqab. It offers four key advantages: low power costs, a cold climate that reduces cooling requirements, proximity to Beijing and access to green electricity that can meet the requirements of international customers such as Apple, according to UCloud’s Ji.
“The ultimate output of computing power is tokens, and the ultimate destination of tokens is electricity,” Ji said. Alibaba Cloud Global Data Center General Manager Wang Chaoyang said electricity in Ulanqab costs about 0.32-0.35 yuan per kilowatt-hour, compared with 0.60-0.90 yuan in eastern and southern China. For a 1-GW data center, the annual electricity bill can differ by as much as 5 billion yuan.
China is also making progress in replacing foreign AI chips. Alibaba’s T-Head Wuzhen AI chips had shipped 560,000 units by April, and Wang expects domestic chip production to rise steadily, with locally developed chips accounting for an increasing share of computing centers.
China Mobile’s Harbin computing center has already achieved 100% domestic AI chip deployment. Qinghai is also attracting chipmakers including Enflame Technology, Biren Technology and Moore Threads.
Ji said domestic GPUs have reached the “usable” stage and are improving rapidly, although they still lag leading foreign products in overall performance. He expects the industry to accelerate significantly.
High-power centers are in demand
As conventional data centers evolve into AI data centers, power density is rising sharply. Wang said a single rack in an AI data center can now consume 100 kilowatts, while 1,000-kilowatt racks could emerge within two years.
This is creating a structural imbalance: some low-power data centers have relatively high vacancy rates, while high-power facilities are attracting strong demand, often securing orders as soon as projects are launched.
At the same time, bottlenecks are spreading beyond chips. Memory, electricity, optical fiber and optical modules are all becoming constraints. A 100,000-accelerator cluster can require hundreds of thousands of kilometers of optical fiber even though China produces about 600 million kilometers annually.
Training moves west, inference moves east
The industry is increasingly settling on a geographical division of labor. Wang said large training clusters of 1,000, 10,000 or even 100,000 accelerators have moved west of the Hu-Huanyong Line, a geographical dividing line running roughly from northeastern to southwestern China that separates the country’s densely populated east from its sparsely populated west. At the same time inference capacity — using the trained model to answer questions and generate images — is moving east toward the main centers of AI demand in the Yangtze River Delta, Greater Bay Area and Beijing-Tianjin-Hebei region.
For now, Ji said, computing scarcity is a greater constraint than latency. Concentrating resources in large clusters allows capacity to be centrally scheduled and reused efficiently, avoiding the fragmented idle capacity that can arise when computing is distributed at the edge.
As edge computing improves, some processing could eventually move closer to users. For now, however, China’s AI computing map is being defined by a clear trend: train in the west, infer in the east.
Source:
STAR Market Daily
