Shenzhen's Cloud Brain III opens 10,000 petaflops in trial run

On the stage of the World Computing Network Conference, the announcement was short: 10,000 petaflops of domestically built AI computing are now in trial operation at Peng Cheng Cloud Brain III, the large-scale science facility of Peng Cheng Laboratory in Shenzhen. The machines are not finished. The lab describes the phase as 'building while operating', with the rest of the capacity still being installed.

The capacity that is already running is not idle. According to the lab, it supports domestic large-model training and a dedicated research network for AI for science. The completed system is planned to reach 16,000 petaflops, or 16 EFLOPS, a figure reported by IT Home, which also describes it as the largest single AI computing facility planned in China. The lab's release says only that the system uses a 'pioneering domestic intelligent-computing supernode architecture' designed to ease data-transfer bottlenecks, which means the way the machines are wired together to move data quickly. IT Home, citing public reports, says the system is built on Huawei's Ascend 384 SuperPoD architecture; the laboratory itself does not name a chip supplier.

The ambition goes beyond one building. Lu Yutong, director of the National Supercomputing Center in Shenzhen, told local media that the facility gives the city a base for linking computing across its districts and for making computing services widely accessible, and that high-speed links should extend to Hong Kong and Macao. The site is designed as a core node of China Computing NET (C2NET), a national pilot program that Gao Wen, the laboratory's director, says has now run for four years. In his keynote, Gao outlined a three-step path: pooling computing resources, scheduling tasks across regions, and eventually making computing available on demand, 'like electricity'. His longer-term target, which he called ZPG, combines zettaflop-scale computing, petabit-scale communication rates and gigawatt-scale green power supply.

What changes, and when? For researchers and companies that need large-scale AI training, the immediate gain is access to capacity that is already online, even as the rest is installed; the sources do not say how that access will be priced or allocated. The full 16,000 petaflops depends on completing the installation, and no completion date is given in the sources. The longer horizon is set out in a roadmap released at the conference, produced with Huairou Laboratory: 2026 to 2030 are years of planning and building the framework, and 2031 to 2035 are meant for coordinated operation and system-wide optimization of computing and power systems. Whether the network moves from pooling to on-demand computing within that timeline is the bet the lab is now making, in public.