China has put a concrete number behind its next phase of AI infrastructure growth. A new Ministry of Industry and Information Technology plan sets 2030 capacity at 9,800 EFLOPS, a level that would require a sharp expansion from the 2,185 EFLOPS recorded at the end of June. The target appears in the 2026–2030 information and communications industry plan, which was detailed by SCMP on September 8.
EFLOPS, or exa floating-point operations per second, has become one of the clearest ways to describe the scale of computing available for training and operating modern models. The target does not mean China has committed to one single national machine. It describes the aggregate intelligent-computing capacity the country wants to build across clusters, regional facilities, cloud infrastructure, and application-specific inference systems by the end of the next five-year period.
The new figure makes a sprawling policy priority easier to measure. It also makes clear that Beijing sees computing capacity, electricity, data-center construction, and domestic accelerators as connected parts of the same industrial program.
A fourfold target from a fast-rising base
MIIT said intelligent computing capacity had reached 2,185 EFLOPS by the end of June, up 177% from a year earlier. SCMP also reported that the National Data Administration put the figure at about 2,450 EFLOPS by the end of July. Even from that higher point, the 2030 objective requires a large multiplication of capacity rather than a marginal increase.
The plan calls for orderly deployment of intelligent-computing clusters equipped with 10,000 accelerator cards, as well as larger systems using 100,000 cards or more. It also asks for inference facilities designed around particular uses. That distinction matters. Training frontier models often requires bursts of tightly connected compute, while serving models at scale requires infrastructure that can manage latency, cost, reliability, and different industry workloads.
China has already been experimenting with infrastructure at that scale. The country’s previous push toward 100,000-card systems, including the Zhengzhou super-cluster project, showed how local governments and industrial partners have treated very large compute installations as a strategic asset. The new national document places those projects inside a larger capacity target rather than leaving them as isolated regional announcements.
The scale of the goal should not be mistaken for an assurance that every proposed project will be built on schedule or fully utilized. Data-center economics depend on power access, financing, network links, software tools, and paying demand. Still, specifying a national capacity objective creates a common reference point for provinces, operators, cloud providers, and chip suppliers that want to align their own plans with policy priorities.
From data centers to domestic accelerators
The plan does more than call for more machines. It emphasizes adapting infrastructure to home-grown computing chips. That reflects the practical reality of China’s AI buildout: capacity is not simply a matter of importing the most advanced accelerators and adding more racks. Hardware availability, interconnection, compilers, frameworks, and application software all shape how much useful work a cluster can deliver.
China’s efforts to develop local chips have been visible from smartphones to cloud systems. It also extends a longer infrastructure buildout, including the earlier national data-center investment plan, into a more explicit target for intelligent computing. Xiaomi’s effort to develop additional in-house processors illustrates the broader domestic push, even though a handset chip and a data-center accelerator solve different technical problems. At the infrastructure level, adapting facilities to Chinese chips involves ensuring that hardware, networking, power design, and software stacks can work together at volume.
That is why the new plan’s emphasis on inference capacity is important. A domestic stack does not need to replace every imported component at once to affect the market. It can gain relevance when an enterprise or local government can run a model, deploy an agent, or operate a specific industry application with hardware and software suited to that task.
The plan also puts information infrastructure spending in a wider context. It calls for 3.8 trillion yuan in cumulative investment over 2026–2030. That sum encompasses information infrastructure broadly, not just AI facilities. Yet the inclusion of intelligent-computing targets means AI infrastructure is one of the clearest places where that spending will be tested.
The next stage of East Data, West Computing
China’s AI infrastructure expansion builds on a geographic rebalancing program that began in 2022. That initiative routes power-hungry digital workloads away from crowded eastern regions toward areas with more available land and energy. China has since organized eight national computing hubs, 10 data-center clusters, and three regions dedicated to coordinating computing facilities with power supplies.
The new plan suggests that this geography will remain central as AI demand grows. Large clusters need reliable power and cooling, while inference systems may need to sit closer to users, industries, or data sources. Building more capacity is therefore also a coordination problem: how to connect distant computing regions, match workloads to available energy, and prevent local investment plans from producing expensive underused facilities.
AI adds a particular urgency to that balancing act. A general data center can host many kinds of workloads, but model training and high-volume inference increase demand for specialized accelerators, denser power systems, and software that can use those resources efficiently. The policy goal is significant because it combines the physical side of computing with the requirements of the AI application layer.
By 2030, the headline number will offer one way to judge whether that strategy has delivered. The more important question may be how much of the added capacity is usable, economical, connected to domestic hardware, and matched to real applications. China has now set the destination. The harder work is turning clusters, chips, power systems, and workloads into an AI infrastructure network that performs as a system.