China’s AI expansion is creating an unusual economic question: where should increasingly power-hungry computing actually live? The answer is beginning to point away from the country’s traditional technology centers and toward places such as Inner Mongolia, Ningxia and Gansu. That shift is not simply a story about governments building more data centers in regions with available land. It reflects a deeper change in the economics of computing, where electricity, cooling, grid capacity, network connectivity and physical space increasingly shape the value of an AI deployment.

China’s “East Data, West Computing” strategy established eight national computing hubs and 10 national data center clusters in 2022, including hubs in Inner Mongolia, Ningxia and Gansu. The policy explicitly considers energy supply, climate, geological conditions, transmission distances and existing industrial infrastructure when deciding where computing capacity should develop. That framework now looks increasingly relevant to AI rather than merely traditional cloud computing. The important question is not whether northwest China can host more servers. It clearly can. The more consequential question is whether the region can turn its physical advantages into a durable cost advantage as AI workloads become larger, denser and more continuous.

AI is turning electricity into a location decision

The economics of AI infrastructure are increasingly tied to electricity. Training large models and running inference at scale require substantial computing capacity, while the servers themselves generate significant heat. That makes the cost and availability of power inseparable from the cost of operating the computing infrastructure. China’s western and northern regions offer a combination that is difficult to reproduce in crowded coastal technology markets: comparatively abundant land, significant renewable-energy resources and, in several locations, cooler or drier conditions that can reduce cooling requirements.

The National Development and Reform Commission identified shortages of land and energy in eastern regions as constraints on data center expansion while pointing to resource availability in western China as an opportunity for additional capacity.  A parcel of inexpensive land has limited value if the electrical connection cannot support a large AI cluster. Cheap electricity has limited value if the network cannot move data efficiently. Renewable generation has limited value if computing facilities cannot obtain reliable power when workloads demand it. The emerging asset, therefore, is not land or electricity by itself. It is the combination.

Northwest China is becoming a systems-engineering play

The strongest case for moving AI infrastructure west is not that one resource is dramatically cheaper. It is that several infrastructure variables can reinforce one another. Ningxia provides a useful example. Zhongwei has a cool, dry climate and significant renewable-energy resources, conditions that have supported its development as a computing hub. Recent government-linked reporting says data centers there can use natural cooling for much of the year, reducing the electricity required for mechanical cooling.

Gansu is developing a similar logic. Qingyang has emerged as one of the national computing hubs, with local projects connecting computing infrastructure to wind and solar generation. Recent reporting also described AI demand as a growing driver of the region’s computing infrastructure. Inner Mongolia brings another version of the model. Horinger New Area near Hohhot has become a significant computing cluster, with multiple data center projects developing around the national computing network. Official Chinese reporting in 2026 highlighted the region’s wind and solar resources as part of the economic rationale for locating computing capacity there. These developments suggest that China is treating compute increasingly like industrial infrastructure rather than a service that must remain physically close to major technology companies.

The network has to make distance economically tolerable

Moving compute away from users creates an obvious problem: distance. AI workloads do not all have identical latency requirements. Some inference applications need rapid responses close to users or enterprise systems. Other workloads, including model training, storage, batch processing and certain forms of analysis, can tolerate greater network distance. China’s national computing strategy explicitly recognizes this difference. Policy documents call for western hubs to handle workloads that can tolerate higher latency while improving high-speed connections between eastern demand centers and western computing capacity. That matters because geography only becomes an economic advantage when networks can make the geography manageable.

China reported that network latency between eastern and western computing hubs had generally met a 20-millisecond target, while more than 1.46 million standard server racks had been installed across the 10 national data center clusters as of the end of March 2024.  The implication is significant. The country is not merely relocating machines. It is attempting to construct a national computing fabric that allows workloads to move according to the relative economics of power, capacity and latency.

China may be redrawing the map of computing

The broader implication reaches beyond China’s regional development strategy. AI is making computing infrastructure more physical at exactly the moment the technology is often described as becoming more abstract and ubiquitous. Models may live in the cloud, but the computing required to train and serve them still depends on substations, transmission lines, cooling systems, fiber routes and buildings capable of carrying enormous electrical loads. That makes geography an increasingly important part of AI economics.

China’s northwestern shift should therefore not be read simply as a search for cheaper land. It is an experiment in matching computational demand with the physical conditions that computing increasingly requires. The country’s coastal technology centers will remain critical for AI companies, research, talent and applications. But the machines supporting that ecosystem do not necessarily need to sit beside the companies building the software. That could prove to be the more consequential development.

If AI continues to increase the total electricity, cooling and physical infrastructure required to deliver computing capacity at scale, the winning locations may not be the places with the largest technology industries. They may be the places where power, climate, land and networks intersect at the lowest sustainable cost. China’s AI race is therefore becoming a test of something broader than chip access or model performance. It is becoming a test of whether a country can reorganize its physical economy around computation. And that raises a question with implications far beyond China: as AI becomes more power-intensive, will the future of computing be determined as much by geography as by technology?