Now think about the possibility of packing 100,000 AI chips into an area the size of a large supermarket. Now visualize 100,000 AI chips in an area about the size of a large supermarket. It’s the power of a new report on AI computing infrastructure in China, and the amount of computing power that’s being crammed into ever more compact facilities.
These chips are no regular hardware. As a team, they can process information to train AI models, serve millions of user requests, and assist with research that may not be feasible on traditional computers.
China’s compact AI supercomputer packs an enormous amount of computing power
It is named Sugon 8000 (Dengfeng) in Zhengzhou, Henan province, in the middle of China. The system came into place officially on 10th July 2026 and was linked to the central server of China’s National Supercomputing Internet, according to a report published in People’s Daily.
For a relatively small site of 2,000 square meters, it is believed that the facility contains more than 3 billion electronic components, cables over 1,600 km and about 1,500 tons of hardware. Its chief designer, Li Bin, explained the engineering feat of working out so many components in such a small space.
It becomes easier to understand the scale if you were to look at this as an average computer room. It is a SUPERCOMPUTING CLUSTER, a cluster of computers that are connected together and operate as a single very powerful computer.
Why 100,000 AI chips need more than just space
The sheer number of AI accelerators in a single facility presents a number of engineering challenges. These chips need to communicate quickly, receive a good electrical supply, and not overheat.
Sugon 8000 is not only embracing conventional room air conditioning, but also phase-change immersion cooling technology. The reported power usage effectiveness (PUE) is less than 1.04, indicating that relatively little extra electrical power is needed for cooling and other functions of the facilities compared to the electrical power consumed by the computing equipment.
The efficiency is key, as the systems powered by artificial intelligence require a lot of power. Improved cooling and reduced space and power requirements can help operators to optimize the use of the available space and energy, but this will also depend on workload, hardware usage, and the electric power supply.
So what is this much AI computing power capable of?
The power of a supercomputer is more than capable of handling the training of chatbots. The report said that Sugon 8000 is meant to be capable of running AI applications and scientific computing, such as climate modeling, biomedical studies, and industrial simulations.
AI and high-performance computing, for example, can be used to analyse complex scientific information, create physical simulations, and even develop medical advancements. Other applications in which AI companies can leverage the large computing clusters would be inference, where the trained models answer queries.
If all the system’s resources were used for inference, it could serve the current demand for tokens in China by 5% to 10%, the chief designer of the system said. Tokens are very small pieces of text that are used by AI models. This is an estimated figure, not necessarily reflecting the amount of demand on the system in normal use.
A bigger signal about China’s AI ambitions
Sugon 8000 embodies a trend of developing interconnected computing systems throughout China, instead of stand-alone systems. The objective is to provide computing resources for a broader spectrum of AI applications, scientific research, and industry uses.
But a system’s chip count is not a good determinant of the performance of competing supercomputers. The chip’s capabilities, memory, networking, software, and what’s being tested also influence the performance. The numbers in the report should thus be viewed as an indicator of the size of the facility, rather than as a measure of its superiority over other facilities.
The indirect impact for everyday users might be greater: as AI services and research advance, and as business applications increase, there could be more capacity available for computing. The real challenge is to see how well the system functions and the amount of useful work it produces.
