This is not investment advice. The author has no position in any of the stocks mentioned. Wccftech.com has a disclosure and ethics policy.
US broker research is aggregating on a few consensus points as they relate to China's AI GPU sphere, especially as the sector is becoming increasingly competitive, with domestic players proliferating while Western giants such as NVIDIA face guard rails.
Specifically, US broker research is now honing in on four key elements of China's AI GPU supply chain:
- SMIC's 7nm node process yield and capacity remain a significant source of consternation. Do note that the vast majority of Huawei's Ascend 910C GPUs reportedly use TSMC 7nm dies, which were purchased in bulk by Huawei by reportedly routing its orders via third parties.
- The varying strategies deployed by China's cloud service providers (CSPs) to gain access to in-demand AI GPUs. To curb the access of Chinese companies to cutting-edge Western AI GPUs via the cloud, the US House of Representatives is currently considering the Remote Access Security Act.
- NVIDIA's upcoming China-focused B40 AI GPU is another fundamental aspect of the Asian giant's ever-evolving AI landscape. While the Trump administration has allowed NVIDIA to resume the shipments of the older H20 GPUs to China for a 15 percent cut, that chip faces an increasing pushback from China's policymaking circles.
- AI CapEx plans form the final rung of this labyrinthine landscape. China is currently pushing for a 100 percent self-reliance in AI compute, which entails a significant cash outlay on the part of major players.
Do note that DeepSeek's DeepGEMM AI model - released only in February 2025 - is written in NVIDIA's CUDA and was trained on NVIDIA's GPUs. However, many domestic Chinese AI GPU players can also support the model using the UEBMO FP8 memory calculation format.
Huawei's CloudMatrix 384, which bundles up to 384 Ascend chips, lacks direct support for memory-optimal calculation formats such as the FP8. When AI models are trained on the FP8 format, they typically consume a lot less memory. Of course, Huawei has created a translation tool to induce artificial compatibility with FP8, but this solution remains suboptimal.
Alibaba is developing its own AI GPU. While China's NVIDIA, Cambricon, is currently experiencing a stock mania courtesy of the soaring sales for its Siyuan 590 GPU.
However, most broker research still finds outsized preference for NVIDIA GPUs.
This is largely due to the fact that NVIDIA's GPUs offer superior software support, especially via the CUDA ecosystem. These GPUs also perform better in a cluster, thanks to the company's NVLink interconnect.
Finally, do note that NVIDIA's RTX Pro 6000D systems that are based on the B40 chip do not require a separate license to go on sale in China, as these systems do not use high-bandwidth memory (HBM) and their main use is for inference rather than training foundational AI models. It is, therefore, quite likely that these chips would sell like hot cake once they do become available for Chinese companies.
Follow Wccftech on Google to get more of our news coverage in your feeds.
