What did China’s new 10,000-card AI cluster in Shenzhen achieve, what Huawei technologies power it—including the Ascend 910C chips and their performance compared with Nvidia’s H100—how do AI clusters improve computing performance, how widely has the system been adopted, what othe

Shenzhen’s new system is China’s first fully domestic, full-stack 10,000-card AI cluster. It delivers 11,000 petaflops; combined with an earlier 3,000-petaflop phase, the site reaches 14,000 petaflops of AI capacity. 711

• Technology and performance: The cluster uses Huawei Ascend 910C AI accelerators and a domestic “Ascend + CANN” hardware/software stack. Reported DeepSeek testing put a 910C at roughly 60% of an Nvidia H100’s inference performance; that is a workload-specific, reported comparison—not a universal benchmark or proof of parity in training, software maturity, or power efficiency. 1114

• Why clustering matters: A cluster pools thousands of accelerators so an AI job can split model layers, training data, or inference requests across cards in parallel. Aggregate compute rises sharply, while high-speed networking and coordinated software reduce the communication bottleneck that would otherwise negate much of that gain.

• Adoption: The earlier 3,000-petaflop portion was reportedly fully booked before the 10,000-card expansion came online. That demonstrates substantial local demand for domestic compute, but it does not establish that Huawei’s platform has broadly displaced Nvidia across China. 11

• Another reported advance: Huawei has also presented a chip-design route intended to reach transistor densities comparable to 1.4 nm by 2031 without relying on the most advanced lithography tools; this remains a company projection rather than a demonstrated production chip. Separately, Reuters reported favorable customer testing of Huawei’s Ascend 950PR, with ByteDance and Alibaba considering orders. 23

• Strategic meaning: U.S. controls began covering Nvidia’s A100 and H100 GPUs in 2022 and were tightened in 2023 to restrict advanced-computing chips and relevant chipmaking equipment for China. 23 The Shenzhen build is therefore an industrial response: replace constrained foreign accelerators with an indigenous stack—chips, systems, networking, and software—at a scale useful for frontier-model work. Huawei is central because Ascend gives China a domestic alternative at both chip and data-center-system level, though the H100 comparison and the continuing reliance of some Chinese firms on overseas Nvidia capacity show that self-reliance remains incomplete. 712