Shenzhen-based Pengcheng Laboratory has put domestically developed computing capacity rated at 10 EFLOPS at FP16 precision into trial operation at its Pengcheng Cloud Brain III facility, even as construction of the project continues.
The development was announced at the World Computing-power Network Conference 2026, held in Shenzhen from Oct. 15 to Oct. 16. The newly available capacity is equivalent to 10,000 PFLOPS, according to information displayed at the launch ceremony.
Pengcheng Cloud Brain III uses a first-of-its-kind, domestically developed intelligent-computing "supernode" architecture, which the laboratory said could overcome bottlenecks in data transmission. The system will support the training of home-grown large AI models and research on dedicated networks for scientific AI.
Once completed, the facility is expected to become one of the core nodes of China Computing NET (C²NET). The pilot programme is designed to connect computing centres in different regions and allow computing tasks to be scheduled across them.
The longer-term aim is to make computing power available much like electricity, allowing users to access it when needed without having to know where the resources are located.
Gao Wen, director of Pengcheng Laboratory and a member of the Chinese Academy of Engineering
Gao Wen, director of Pengcheng Laboratory and a member of the Chinese Academy of Engineering, said the network would develop in three stages: bringing computing resources together, enabling tasks to be scheduled across regions and, eventually, providing computing power on demand.
He said major obstacles remained, including computing supply, network connectivity, task scheduling, data security and coordination between computing facilities and power systems. Researchers and businesses would need to work together in an open-source ecosystem to accelerate technological advances, he added.
Jack Dongarra, a Turing Award laureate, attends the conference.
Jack Dongarra, a Turing Award laureate, said AI was reshaping computer hardware. Next-generation high-performance computing would require mixed-precision computing, less movement of data, more efficient algorithms and closer coordination between hardware and software design, he said.
Lu Yutong, chair of the conference's technical programme committee and a professor at Sun Yat-sen University, said connecting different types of computing resources was only the first step. Software, applications and service platforms were also needed to turn those resources into services for businesses and research institutions.
About 1,000 people attended the conference, roughly three times the number at the previous gathering. The number of parallel forums increased to eight from three, and a dedicated academic-paper programme was introduced.
Two white papers were also released. One proposed a more open, inclusive, secure and sustainable global computing-power network, while the other outlined plans to improve coordination between computing facilities and power systems through 2035.
Reporter: Hu Nan
Photos: Pengcheng Laboratory
