China's Cloud Data Centers: A New Geography of Electricity and Compute

The short answer: China no longer treats data centers as an extension of the digital sector. By organising where compute physically runs — latency-sensitive workloads in the east, deferrable workloads in the west — it has made cloud infrastructure part of industrial, energy and territorial policy. That relieves some bottlenecks. It does not make Chinese cloud automatically low-carbon.

In a previous article we analysed the still limited, but fast-growing, place of AI and data centers in Chinese electricity demand. The subject now deserves its own examination. Behind the words “data center” sits an extremely heterogeneous infrastructure: a public cloud campus, a colocation site, a corporate server room and an AI training cluster do not share the same use, the same load profile or the same relationship with the power grid.

So the right question is not only how much electricity these sites consume. It is also where they are located, what computation they run, and which electricity actually serves their load. The carbon trajectory of Chinese cloud is decided at that intersection between digital services, territory and the power system.

Compute has become systemic

China is already one of the world's major data center poles. According to the International Energy Agency (IEA), it accounted for roughly 25% of global data center electricity consumption in 2024, in the order of 100 TWh, behind the United States. In the same year, total national electricity consumption reached 9,852.1 TWh. The order of magnitude is therefore around 1% of national demand: still a minority weight in macroeconomic terms, but already comparable to a large industrial segment in some territories.

The physical size of the estate confirms this rise. At the end of 2023, China had more than 8.1 million racks in service and ranked second worldwide for the scale of its computing power. In February 2022, the national planning authority still referred to 5 million standardised racks. The progression reflects an accelerating deployment, even if capacity comparisons should always be read with caution.

Cloud is the first explanation for this movement. According to the China Academy of Information and Communications Technology (CAICT), the Chinese cloud market reached CNY 616.5 billion in 2023, up 35.5% year on year. The diffusion of public cloud, the growth of digital platforms and, now, the integration of generative AI are changing demand: fewer isolated servers, more pooled resources, high-capacity networks and specialised compute clusters.

“A methodological point: IT equipment consumption and whole-site consumption are not the same thing. The second also includes cooling, electrical losses, power distribution and backup systems. That is why two serious estimates can differ without contradicting each other.”

From server room to industrial cloud: three layers, not a replacement

China's evolution is not a simple replacement of old data centers by new buildings. It results from the superposition of three layers.

The first is the legacy estate of corporate, telecom-operator and government server rooms. These assets remain essential for proximity functions, business continuity and sovereignty. The second is public cloud and colocation: pooled capacity operated by cloud providers or specialised operators, allowing companies to rent compute, storage and network on demand. The third is intelligent computing — infrastructure optimised for training and inference of AI models, with accelerators, very fast interconnect and higher power densities.

This classification by operating model must be crossed with Chinese industrial-policy terminology. The 2023 national document calls for coordinating general-purpose computing, intelligent computing and supercomputing. That second grid is not an administrative subtlety: it is a reminder that a general cloud site, a scientific supercomputer and an AI cluster do not call on the grid in the same way.

The IEA, for its part, distinguishes enterprise, colocation/server-provider and hyperscale centers, and stresses that their consumption structures differ, notably because cooling use varies strongly with site density and efficiency. To read the Chinese landscape you therefore need to hold three variables together: type of center, type of workload, and location.

East Data, West Computing: moving the right workloads, not the users

That is precisely the objective of the strategy known as East Data, West Computing (Dongshu Xisuan). Launched nationally in 2022, it organises eight national computing hubs and ten data center clusters: Beijing-Tianjin-Hebei, the Yangtze River Delta, the Guangdong-Hong Kong-Macao Greater Bay Area, Chengdu-Chongqing, Inner Mongolia, Guizhou, Gansu and Ningxia.

The reasoning is simple. The large pools of digital demand sit in the eastern coastal regions, where land, electricity and sometimes water are more constrained. Western provinces have, to varying degrees, more favourable land resources, climatic conditions or renewable potential. The programme therefore seeks to pool computing power at national scale rather than endlessly reproducing large capacity in coastal metropolises.

But “move the data west” is a reductive shorthand. The NDRC explicitly sets a hierarchy of uses. Background processing, offline analytics and backup can be transferred. By contrast, highly latency-sensitive activities — industrial internet, finance, disaster warning, telemedicine, video calls and AI inference — must stay close to demand, in the eastern hubs. The programme does not make urban data centers disappear; it allocates workloads according to their tolerance for transmission delay.

This nuance is fundamental for energy. Model training, rendering or backup can sometimes be scheduled where electrical capacity is available. Real-time services stay anchored where users and economic activity are concentrated. The future geography of Chinese cloud will therefore be less an exodus from east to west than a functional division of compute.

A minority share of national demand, a major local challenge

The IEA projects a sharp rise in Chinese data center electricity consumption: around +175 TWh between 2024 and 2030, a 170% increase in its central scenario. From a level close to 100 TWh in 2024, that leads to an order of magnitude of 275 to 280 TWh in 2030. These figures are large enough to justify grid planning; they do not mean data centers will become the primary driver of Chinese electricity demand.

An independent synthesis of IEA work indicates that data centers accounted for only about 3% of the increase in Chinese electricity demand since 2022, and could represent around 6% through 2027. Industry, electrification of end uses, electric vehicles and air conditioning remain far more massive growth factors. The analytical risk is to replace one excess with another: it would be as wrong to ignore digital infrastructure as to make AI the single explanation for rising Chinese demand.

The decisive point is concentration. An intelligent computing cluster can call several hundred megawatts within a limited area, whereas an equivalent rise in residential consumption spreads across space. Grid connections, transformers, transmission lines and network construction schedules then become as strategic as processors. The IEA insists on this local character: data centers are a modest share of total demand, but their concentration can create electrical bottlenecks that are far more visible at regional scale.

Efficiency, green electricity, grid: three levers that should not be confused

The Chinese response combines several instruments of differing reach. The first is operational efficiency. The 2024 national plan targets an average PUE below 1.5 for data centers by 2025, and a 10% annual increase in the sector's renewable utilisation rate. A low PUE means a higher share of total site electricity serves IT equipment rather than cooling and auxiliaries. That is necessary; it is not sufficient to cut emissions.

The second lever is siting and contractual access to low-carbon electricity. The 2023 national notice sets the objective that, by end-2025, more than 60% of new computing resources are installed in the national hubs, and that the green electricity share of new data centers in those hubs exceeds 80%. These targets signal an intention to bring compute planning and power-system planning closer together.

The third lever is flexibility. Policy encourages integration between compute, grid, renewable generation and storage, notably through source-grid-load-storage models. Non-urgent tasks can in principle be scheduled at times and in places more favourable to the system. The prospect is promising, but it depends on the quality of price signals, on interprovincial interconnection, and on the real ability to defer workloads without degrading service.

Why “green cloud in the west” must not become a shortcut

Western China is not an energy blank page. Renewable resources are abundant in several provinces, but their output is variable, transmission needs are considerable and local mixes remain contrasted. Moving a computing task to a western province can improve the electrical profile of a project; it is not, on its own, a demonstration of hour-by-hour low-carbon supply.

A Carbon Brief analysis recalls that interprovincial green electricity trading, transmission costs and integration of variable resources remain challenges. It also stresses that the build-up of centers in the north and west must be examined against water stress, because cooling can create an additional constraint. Caution is all the more necessary as sector consumption estimates vary strongly with scope and assumptions.

Serious reporting should therefore ask for three elements. First, an efficiency indicator such as PUE. Second, evidence of power sourcing — purchase contract, certificate traceability and, where possible, the time profile of generation. Third, the local grid context: available capacity, marginal emissions, reinforcement needs and water constraint. Without that triple reading, the word “green” risks describing an intention rather than a demonstrated performance.

Conclusion: digital infrastructure has become energy policy

Chinese cloud data centers are no longer the invisible background of the digital economy. They are becoming industrial-policy infrastructure, on the same footing as transmission lines, renewable parks or telecom networks. East Data, West Computing is ambitious because it does not only seek to build more servers: it attempts to orchestrate the placement of compute loads alongside the geography of electricity.

Three lessons emerge. First, the sector's national electricity footprint remains limited in the short term, but its growth and concentration make its regional impact significant. Second, the distinction between low-latency compute and movable workloads is the key to understanding the new Chinese cloud map. Third, the energy efficiency of a site, its green electricity purchasing and the real decarbonisation of the grid are three linked but non-interchangeable subjects.

For companies that use cloud, train AI models or buy digital services in China, good due diligence should therefore no longer stop at requesting a PUE or a certificate. It should cover workload location, latency tolerance, power sourcing model and the mix of the grid that serves the load. It is at that level of precision that cloud can become a credible decarbonisation tool — and not merely a new and growing electricity use.

Editorial note: all 2030 estimates are presented as scenarios. Data center consumption figures vary depending on whether IT energy alone, the whole facility, networks and small enterprise sites are included.

If cloud and AI services sit inside your Scope 3 inventory, a Value Chain Readiness Check can identify which digital suppliers require evidence rather than a certificate, and how to write that evidence into your next contract.