China Plans National Cloud Network to Put Idle Data Centers to Work

A National Network for Computing Power Going Unused

China is developing a national cloud service to sell spare computing power from data centers whose utilization rates have fallen to between 20% and 30%, seeking to recover value from an infrastructure boom that has left expensive facilities underused across the country.

As of August 4, 2025, the Ministry of Industry and Information Technology (MIIT) is working with China Mobile, China Telecom and China Unicom, the three state telecom giants, on a service that would pool surplus computing resources. The goal is standardized interconnection of public computing power nationwide by 2028.

The financial pressure is already substantial. More than 100 data center projects have been canceled over the past 18 months, compared with 11 cancellations in 2023. Government procurement nevertheless reached 24.7 billion yuan, about $3.4 billion, in 2024, with another 12.4 billion yuan allocated for 2025.

China poured billions of dollars into data centers to support artificial intelligence (AI), cloud computing and advanced digital services. Government and private investors expected those facilities to help make the country a global technology leader. Instead, speculative investment, changing demand and technical problems have left much of the new infrastructure struggling to attract paying customers.

The proposed network would let companies, researchers and government agencies request computing power through a shared service. Its success depends on more than connecting buildings: the system must match customers with suitable machines, move data quickly enough and make different hardware platforms work together.

For operators, the immediate aim is to turn idle equipment into revenue. For Beijing, the larger challenge is to coordinate infrastructure investment with actual demand while preserving the computing resources needed for its AI ambitions.

How the Western Building Boom Took Shape

The Eastern Data, Western Computing initiative, launched in 2022, encouraged local governments and companies to build data centers in less populated western regions with abundant energy. Those facilities were intended to serve the economically active eastern regions, where demand for digital services was growing.

Electricity costs made the idea attractive. Moving work that requires large amounts of computing power to places with cheaper energy could reduce costs while supporting businesses in the east. Data centers need electricity both to run their machines and to cool them.

Local governments, eager to bring in investment and strengthen regional economies, rushed to approve and finance projects. Many did so without a clear understanding of customer demand or the technical requirements of AI and cloud services. Hundreds of large data centers appeared, including facilities far from major customers.

Generative AI added another burst of investment. Excitement about large language models such as ChatGPT pushed companies and government agencies to acquire advanced graphics processing units (GPUs), including Nvidia products and domestic alternatives such as Huawei Ascend chips. These processors can perform many calculations at once, making them useful for demanding AI tasks.

Investors hoped to profit from the AI boom. As the market cooled and attention shifted from building massive models to using them efficiently, demand for raw computing power fell sharply. The equipment that investors had bought was not always matched to the work customers wanted to pay for.

Industry analysts estimate that up to 80% of newly built computing resources remain unused. Many facilities were constructed to obtain subsidies, loans or tax incentives rather than serve a clear business need. Some operators pursued those benefits without intending to run large AI workloads, leaving expensive projects as financial liabilities.

The Cost of Keeping Empty Capacity Running

Data centers require continuing spending long after construction ends. Land, buildings and hardware account for the initial investment, while electricity, cooling and maintenance create recurring bills. Idle high performance central processing units (CPUs) and GPUs still consume power and require upkeep, reducing the return an operator can earn.

Hardware also has a limited useful life. New generations of equipment can make existing machines less competitive before their owners recover the purchase cost. A facility waiting for customers therefore faces both operating losses and the risk that its equipment will become outdated.

The cancellation figures show how sharply investment plans have changed. More than 100 projects canceled over 18 months is a much larger number than the 11 recorded in 2023, although the periods cover different lengths of time. State procurement has continued alongside those cancellations, leaving China to manage existing surplus while still funding computing infrastructure.

The main dates show how the building campaign, spending commitments and national ambitions fit together:

• 2022: China launched Eastern Data, Western Computing to connect demand in the east with computing capacity in the west.
• 2023: Data center project cancellations totaled 11.
• 2024: Government procurement for data centers reached 24.7 billion yuan, about $3.4 billion.
• 2025: Another 12.4 billion yuan had been allocated as of August 4.
• 2028: China aims to achieve standardized interconnection of public computing power nationwide.
• 2030: China aims to become the global AI leader, with an AI industry worth $100 billion and $1 trillion in added economic value.

The money allocated for 2025 is already about half the amount procured in all of 2024. Because one figure is an allocation and the other is procurement, they describe different stages of government spending. Both show that substantial public funding remains involved despite the surplus.

What the National Cloud Would Do

MIIT and the three telecom companies are developing a national service run by the state to organize and sell spare capacity. Pooling resources would give underused data centers a route to customers beyond their immediate region.

Companies, researchers and government agencies would specify how much processing power and network capacity they need. The proposed service would then organize those resources without requiring users to manage the location or underlying hardware themselves.

Chen Yili, deputy chief engineer at the China Academy of Information and Communications Technology, described that ambition at an industry conference:

Everything will be handed over to the cloud for unified organization, orchestration, and scheduling capabilities.

Orchestration means coordinating the machines and software needed to complete a task. Scheduling means deciding where and when that task runs. A national service would need both capabilities to direct work toward available equipment while meeting customer requirements.

Officials hope the approach will improve the economics of existing data centers, reduce waste and lower costs. Better use of infrastructure would also support AI and cloud services and help China maintain its position in global technology competition.

The 2028 goal concerns standardized interconnection across the country. No separate launch date for the national cloud service has been given. Connecting resources and making them straightforward for customers to use remain substantial parts of the work.

Why Cheap Electricity Cannot Solve Distance

Western data centers benefit from cheaper electricity, but many are far from the eastern businesses they were intended to serve. That distance creates network latency, the time data takes to travel between a user and a computing facility.

China originally targeted a maximum latency of 20 milliseconds. Many facilities fail to meet that standard. For services that need an immediate response, even a short delay can make a remote facility unsuitable.

Financial trading and interactive AI services are among the applications affected. A customer using an interactive service needs information to travel to the computing equipment and a response to come back quickly. Cheaper processing becomes less attractive when the delay prevents the service from working as required.

The location problem helps explain why having spare machines does not automatically bring customers. A national system must find capacity that meets both the computing requirement and the network requirement. An available processor in a distant province may be unable to serve a particular task at an acceptable speed.

Central scheduling could help direct suitable work to western facilities, but it must account for those limits. The economic promise of cheap energy and the practical need for fast communication have to be addressed together.

Different Chips Make Integration Harder

Nvidia and Huawei equipment create another obstacle to a unified service. Some data centers use Nvidia GPUs and its CUDA software stack, while others use Huawei Ascend accelerators and the CANN stack. Other facilities contain a mixture of domestic and foreign components.

A software stack is the collection of programs and tools that allows applications to use the hardware. Equipment built around different stacks needs different management tools, making it difficult to move work smoothly between facilities.

The proposed cloud must therefore handle differences in both machines and software. A customer may need a particular platform, rather than any available processor. Giving the user a simple ordering process requires complex coordination behind it.

Analysts suggest the government may need to let customers choose their hardware platform. That could increase use of clusters powered by Nvidia equipment, while doing less for facilities built around Huawei chips. The outcome would depend on which platforms customers need for their applications.

The lack of standardization also complicates the promise that users will be able to request resources without thinking about the equipment underneath. Operators must make the service convenient while preserving the technical choices needed for the work to run.

Officials remain optimistic that centralized orchestration and scheduling can support AI and cloud development. Integrating varied hardware will take time, however, and the service may initially deliver less than its full ambition.

Beijing Tightens Controls on New Projects

The National Development and Reform Commission (NDRC) has introduced tighter requirements to prevent another round of excessive construction. New data center projects must meet specified utilization thresholds and secure purchase agreements before approval.

Those requirements tie construction more closely to customer demand. Purchase agreements give authorities a basis for assessing whether a project has buyers, while utilization requirements focus attention on whether computing resources are being used.

Local governments are also barred from starting small computing infrastructure projects without clear economic justification. The NDRC is conducting a nationwide assessment intended to remove unqualified projects and ensure facilities meet basic efficiency criteria.

The policy combines two tasks: putting existing resources to better use and limiting new construction that could deepen the surplus. For local authorities, it places greater weight on the business case for a data center rather than the investment associated with building one.

President Xi Jinping has questioned whether every province needs to pursue the same industries. His concerns include AI, computing power and new energy vehicles. He has described provinces as "chasing the same trends," reflecting a broader effort to align local development with national priorities and reduce fragmented investment.

The regulatory changes aim to direct money toward projects that can deliver economic value. They also address an incentive problem exposed by the building boom: local gains from securing investment can encourage construction even when the eventual demand for services is weak.

AI Ambitions Meet Limits on Advanced Chips

China views the national cloud effort as part of its wider industrial policy for AI and advanced technology. Its 2030 ambition combines global leadership with a $100 billion AI industry and $1 trillion in added value to the economy.

Beijing is using investment funds led by the state, support for domestic chip development and the creation of a National Integrated Computing Network to pursue those goals. Alibaba, Tencent and Baidu are also major participants, investing heavily in AI research and infrastructure.

Export controls led by the United States restrict Chinese access to advanced AI chips and manufacturing equipment. Those restrictions have pushed China toward domestic alternatives. Huawei is making progress, but its chips still lag behind global leaders in performance and adoption.

The national network must operate within that technology environment. China is trying to make more efficient use of what it has while also developing the equipment it needs. Domestic platforms are part of that effort, even as differences in performance and software adoption affect customer choices.

State support can help address shortages in chip development, skilled workers and the energy needed by data centers. The same spending carries risks when resources are allocated inefficiently, customers face pressure to use less advanced domestic platforms or projects lack a match with market demand.

The surplus therefore reaches beyond empty server capacity. It concerns how public investment, domestic technology development and private demand can work together without generating another wave of facilities that struggle to earn revenue.

Operators Adjust as AI Work Changes

Some data center operators are closing underused facilities to avoid continuing losses. Others are offering free computing vouchers to local technology businesses in an effort to attract customers and increase use of their machines.

Alibaba and ByteDance are planning large new investments, betting that demand for AI and cloud services will eventually catch up with supply. Their plans sit alongside closures and cancellations, showing that businesses are making different judgments about where future demand will appear.

AI workloads are changing as well. Training a massive model requires large clusters of advanced GPUs to learn patterns from data. Inference means running a trained model to produce an answer or perform a task. That work can use more specialized accelerators at lower cost.

The shift from training toward inference changes the type of computing capacity customers need. It may help absorb some spare resources over time, while also requiring operators to match their equipment more closely to applications rather than relying on broad enthusiasm for AI.

Industry expectations are that the surplus will be temporary as AI applications become more widely used and the ecosystem supporting domestic chips matures. Consolidation and standardization could improve efficiency and support more sustainable growth. Those expectations depend on customers finding useful applications and suitable computing services.

China's experience shows the tension between rapid infrastructure planning and a technology market that changes quickly. Public support can accelerate construction and supply essential resources, but projects can become wasteful when their design and location are poorly matched to real demand.

Private firms remain the main drivers of AI and cloud innovation. The national cloud's success will depend on central coordination responding to market needs, allowing technical flexibility and continuing investment in research and talent. For operators carrying the cost of idle machines, the practical test is whether that coordination brings paying work to equipment already installed.

Key Points

• China's data center utilization rates are between 20% and 30%; industry analysts estimate that up to 80% of newly built computing resources are unused.
• MIIT and China Mobile, China Telecom and China Unicom are developing a national cloud service to pool and sell surplus capacity.
• China targets standardized nationwide interconnection of public computing power by 2028.
• More than 100 projects were canceled over 18 months, while substantial government spending continued.
• Distance from customers and differences between Nvidia and Huawei platforms complicate the proposed network.
• The NDRC has tightened approval requirements and is assessing projects nationwide.
• Changing AI workloads, private investment and customer demand will shape how much spare capacity finds a use.