While Chinese artificial intelligence models rival US counterparts in capability, their reliance on domestic silicon remains limited, particularly during the crucial pre-training phase. Recent projects show progress as several local AI models begin to leverage China’s chips, signaling long-term ambitions for self-sufficiency.
- Domestic chips widely used for AI inference, less for pre-training
- Zhipu AI and Meituan train major models on Huawei silicon
- Startups show feasibility of efficient, lightweight on-device models
What happened
China’s AI community has made notable advancements in training artificial intelligence models using domestically produced chips, particularly from Huawei’s Ascend series. Projects such as Zhipu AI’s GLM-Image model and Meituan’s LongCat-2.0-Preview have been trained on local hardware clusters, highlighting a move beyond inference tasks. However, none have shifted the massively resource-intensive pre-training of large language models fully onto Chinese silicon yet.
Other efforts include ModelBest’s development of compact AI models optimized for on-device use, leveraging Huawei’s Ascend hardware to maximize computational efficiency. A collaborative team from Huawei and Shenzhen focused on post-training processes, which require substantially less computing power than pre-training. These developments together denote an ecosystem striving for greater technological independence in AI.
Why it matters
China’s push to utilize domestic chips across AI development stages is driven by growing export controls affecting access to US technology and a broader ambition for technological self-reliance. Although current Chinese AI silicon lags behind global leaders like Nvidia in raw performance, expanding capabilities in model training represent critical progress toward reducing dependency on foreign suppliers.
Building a full domestic AI hardware and software supply chain is rare globally and carries significant long-term strategic value for China. Success in this arena could allow Chinese AI developers to innovate and produce at scale without external bottlenecks or restrictions, a critical advantage amid intensifying geopolitical competition over semiconductor technologies.
What to watch next
Attention will focus on whether Chinese firms can scale up to pre-train large language models like those from leading global AI developers entirely on native silicon. Achieving this milestone would demonstrate parity in hardware capability and reinforce China’s AI sovereignty goals. Collaboration between academic institutions, startups, and tech giants such as Huawei will be crucial in overcoming current performance and efficiency gaps.
Additionally, monitoring how quickly these domestic chips penetrate broader commercial AI applications and evolve to support increasingly complex models will indicate the maturity of China’s AI ecosystem. Developments in software tools analogous to Nvidia’s CUDA, such as Huawei’s CANN framework, will also play a key role in enhancing hardware utilization and attracting developer engagement.