Disjointed AI environments and restricted inventory: Understanding China's struggle to replace Nvidia hardware for AI development

In the rapidly evolving world of artificial intelligence (AI), China is making strides to establish its own semiconductor industry, particularly in the realm of supercomputers and fabrication tools. This push has been ongoing since the mid-2010s, as the country seeks self-sufficiency in this critical technology sector.

One of the key players in this narrative is Huawei, the multinational technology company based in Shenzhen. In 2025, Huawei unveiled the CloudMatrix 384, one of several domestic AI accelerators it has developed. This year, the company took another significant step by open-sourcing its CANN (Compute Architecture for Neural Networks) software stack, specifically optimised for AI and its Ascend hardware.

However, the maturity of Huawei's CANN lags behind Nvidia's CUDA due to a lack of a broad, stable installed base of Ascend processors outside Huawei's own projects. This has led to unstable performance, slower chip-to-chip connectivity, and limitations in Huawei's CANN software toolkit, as experienced by DeepSeek, a Chinese AI company that abandoned plans to train its next-generation R2 model on Huawei's Ascend platforms.

Analyst Lennart Heim believes that Huawei illegally obtained around three million Ascend 910B dies from TSMC in 2024, comparable to what Nvidia supplied to China in the same period. This alleged deception highlights the challenges China faces in its quest to compete with global leaders in semiconductor manufacturing, such as TSMC and AMD.

China's efforts to switch AI companies to using domestic hardware have been influenced by the U.S.'s imposed sanctions against China's high-tech sectors and the subsequent cancellation of the AI Diffusion Rule. President Trump's announcement of a 15% sales tax on hardware sold to China from AMD and Nvidia has further fuelled China's push for self-sufficiency.

In response, the Model-Chip Ecosystem Innovation Alliance was formed this summer, aiming to build a fully localised AI stack, linking hardware, models, and infrastructure. The alliance is attempting to address the issue by setting common mid-level standards, such as shared model formats, operator definitions, and framework APIs.

Despite these efforts, China's inability to produce hardware that is on par with AMD or Nvidia in volume domestically remains a significant hurdle. SMIC, a China-based foundry, cannot match the process technologies offered by TSMC. SMIC is expected to start building chips on its 6nm-class process technology and even 5nm-class production node, but it remains unclear if volumes will meet the demands of AI training and inference.

Meanwhile, competition on various fronts, the low volume of China-developed AI accelerators, and a lack of common standards will make it hard for Chinese companies to challenge Nvidia's already dominant ecosystem. However, the open-sourcing of Huawei's CANN software stack could provide a platform for Chinese companies such as Baidu, iFlytek, and SenseTime to compete with Nvidia's CUDA software stack.

As the race for AI dominance continues, it is clear that China's push for self-sufficiency in semiconductors will be a key factor in shaping the future of this technology sector.