What have experts said about the performance of the Huawei AI chips (Ascend 950, Atlas 950 superpod, etc.)?

Huawei's Ascend AI chips have emerged as a formidable competitor within the Chinese market, driven by strong domestic demand and significant performance gains. Major Chinese technology firms, including ByteDance, Alibaba, and Tencent, are reportedly rushing to place orders for Huawei's Ascend 950PR chips [1, 7, 19]. This demand is bolstered by substantial production capabilities, with Huawei shipping millions of AI chips in a single year and stating an annual production capacity of **750,000** units [2, 25]. NVIDIA CEO Jensen Huang has acknowledged that Huawei is "very, very strong" in the Chinese market, with local chip companies performing well in the vacuum left by NVIDIA's partial withdrawal due to export controls . This market traction is enabling the development of a domestic alternative to the NVIDIA-centric AI hardware ecosystem .

Direct performance comparisons between Huawei and NVIDIA chips present a complex picture with some conflicting assessments. While one expert claims Huawei has developed AI chips with performance comparable to NVIDIA's H100 GPUs , others maintain that its semiconductor technology remains significantly behind NVIDIA's top-tier offerings like the H200 and the upcoming Rubin platform [11, 21]. A more precise benchmark indicates that within the constraints of U.S. export controls, Huawei's Ascend 950PR inference chip performs **2.87 times better** on inference tasks than NVIDIA's H20, the most advanced chip currently available for sale to China . This suggests Huawei has successfully optimized its hardware to outperform the specific export-grade chips it competes with domestically, even if it lags the global state-of-the-art. The company's innovation also extends to components like custom memory, which surprised some analysts .

To compensate for any single-chip performance deficit against top Western models, Huawei's strategy emphasizes system-level integration and co-design . Experts note that by clustering a larger number of its domestic chips, Huawei can match the performance of systems using fewer, more powerful NVIDIA chips [6, 14]. This approach is exemplified by the Cloud Matrix system, which networks 384 Ascend chips to achieve competitive performance at the rack level . This full-stack integration allows Huawei to maximize the effective performance of its hardware, a critical strategy when access to leading-edge fabrication from foundries like TSMC is restricted . The Western semiconductor supply chain is estimated to have a manufacturing capacity lead of **10x to 15x** over Huawei's current domestic capabilities, making efficient system design paramount .

The success of this strategy is evident in the real-world adoption by major Chinese AI labs like DeepSeek. DeepSeek is explicitly building its software stack on Huawei's hardware, rather than NVIDIA's , with its models being adapted for and powered by Ascend chips [9, 20]. Critically, the DeepSeek V4 model was co-optimized for both NVIDIA's next-generation Blackwell architecture and Huawei's AI chips, indicating that Huawei is considered a primary development target, not a secondary alternative . This deep integration of hardware and software challenges the long-term dominance of NVIDIA's CUDA ecosystem in China and validates the performance of Huawei's full stack in powering cutting-edge AI models .