China’s national data authorities say the country had reached 2.45 million PFLOPS of intelligent-computing capacity, measured at FP16 precision, by the end of July. The figure, announced at the China International Big Data Industry Expo in Guiyang, provides a new official measure of the infrastructure being assembled to train and serve artificial-intelligence systems at scale.
Securities Times, citing a National Data Administration official, reported the number on August 29. A PFLOPS is one quadrillion floating-point operations per second; FP16 refers to the half-precision calculation format commonly used for a portion of modern AI workloads. The announcement is not a measure of a single cloud provider’s deployed GPUs or a direct comparison with another country. It is a government-reported national aggregate intended to show the scale of China’s intelligent-computing network.
The timing is significant. China’s AI discussion has often focused on individual models, chip launches or flagship data centers. The 2.45 million PFLOPS figure shifts attention to the more difficult question of how capacity is organized across regions, energy systems and operators. Compute is useful only if developers can reach it, workloads can be scheduled efficiently and local infrastructure does not remain stranded far from the companies that need it.
FP16 Capacity Puts a Number on China’s AI Infrastructure
The National Data Administration said the 2.45 million PFLOPS total was current as of July 31. According to the official material, more than 85 percent of China’s intelligent-computing capacity is located in eight national computing hubs and three computing-power-and-electricity synergy zones. That concentration reflects a strategy of combining large clusters with locations chosen for power availability, cooling conditions and network links.
The same disclosure says approximately 1.45 million PFLOPS, or about 60 percent of the reported national total, has been integrated into a national monitoring and scheduling platform. The distinction between installed capacity and capacity connected to a common platform matters. A rack of accelerators in a data center is not automatically useful to an AI company elsewhere. Scheduling and monitoring are meant to make supply visible, coordinate workloads and reduce the frictions created by fragmented operators.
These figures should be read as official infrastructure indicators, not as independently audited commercial utilization. They do not reveal how much capacity is available for each type of customer, how much is occupied at peak periods or which portion uses domestic versus imported accelerators. Still, the numbers offer more substance than broad statements that China is expanding compute. They identify the scale the government says has been built and the share it says is connected to the national platform.
China’s drive to integrate compute has a regulatory as well as commercial dimension. EastFrontier previously reported that state-funded data centers were ordered to use domestic AI chips. That policy increases the importance of a network that can accommodate different local hardware and operators. A national capacity figure alone does not prove that every workload can move freely across systems, but it indicates why interoperability and scheduling have become policy priorities.
National Hubs Link Power, Networks and AI Workloads
The Guiyang announcement places the capacity update within China’s larger integrated-computing strategy. The eight national hubs and three energy-computing coordination zones are designed to distribute demand rather than concentrate every large workload in the eastern technology centers. Such a system needs network links that can move data and tasks, electricity supplies that can support power-hungry clusters, and service layers that let customers locate and reserve capacity.
FP16 is central to how this is presented. Many AI inference and training operations can use lower-precision arithmetic to improve speed and reduce memory demand, although actual requirements vary by model and workload. By presenting the national total in FP16 PFLOPS, the government is using a format closer to AI computing than to conventional general-purpose server capacity.
The network dimension is also why AI infrastructure has become a question for companies beyond cloud platforms. Alibaba’s claim that China AI data centers can be built in 100 days focused on construction speed and deployment methods. The National Data Administration’s latest figure highlights what has to happen after construction: capacity needs to be measured, connected and dispatched across a larger system.
The announcement also aligns compute planning with the 15th Five-Year Plan period from 2026 through 2030. The National Data Administration’s emphasis on an integrated network signals that it sees compute as a strategic resource comparable to transport or energy infrastructure. The policy question is not only how many accelerators exist. It is whether they can support model developers, enterprise users and public-sector services without leaving regional clusters underused.
The Next Test Is Useful Capacity, Not Installed Capacity
The 2.45 million PFLOPS figure will be most meaningful if later updates show how the system is used. Capacity can grow quickly in headline terms while applications, data access, software compatibility and power constraints determine whether it creates economic value. China’s 1.45 million PFLOPS scheduling-platform figure suggests that integration is already a stated objective, but it does not settle how effectively workloads are balanced or how easily smaller companies can obtain compute.
This is particularly relevant as Chinese AI firms pursue lower-cost models and more specialized applications. A company training a foundation model needs sustained access to large clusters. A manufacturer using vision AI or a logistics firm running an agent may need smaller but reliable capacity closer to operations. A national network has to serve both, often on equipment that differs by provider and location.
The Big Data Expo announcement therefore offers a snapshot rather than a finish line. It gives investors, developers and policy analysts a reported total, a geographic concentration and a platform-integration ratio. The next evidence to watch will be whether new hubs improve access and reduce idle resources, whether scheduling translates into lower effective costs, and whether the network supports a broad range of Chinese AI workloads instead of a limited set of flagship projects.
For now, the official message is straightforward. China says it has moved beyond scattered AI clusters toward a national compute system measured in millions of FP16 PFLOPS. The value of that system will depend not only on the amount of hardware deployed, but also on the software, networks, electricity and governance that allow those operations to become usable AI capacity.