Artificial intelligence startup DeepSeek has subtly confirmed that its updated V3.1 model is optimized for “next-generation” China-made AI chips, hinting at deeper integration with the country’s advancing semiconductor ecosystem. This announcement immediately ignited a rally in local chipmaker stocks, signaling renewed confidence in China’s push for technological independence.
What’s New with V3.1?
DeepSeek’s new V3.1 release employs a data processing format called UE8M0 FP8, a refined version of the FP8 format that’s becoming a standard for efficient AI operations. The startup notes that this format is “designed for the next generation of domestically produced chips to be released soon,” an indicator of its strategy to align with local silicon innovation.
In addition to hardware compatibility, V3.1 features a hybrid reasoning architecture, enabling it to switch between reasoning and non-reasoning modes. DeepSeek claims the update delivers faster response times while improving computational efficiency.
Market Reaction & Industry Implications
The announcement sparked a surge in share prices for Chinese chipmakers:
- Cambricon Technologies rose 20% to a record valuation exceeding RMB 500 billion (~$70 billion).
- SMIC and Hua Hong Semiconductor also gained significantly—rising 10% and 18%, respectively.
Industry analysts view DeepSeek’s endorsement of domestic chips as a potential tipping point. “If DeepSeek can use China-made chips, then the rest of the semiconductor universe will fly,” said BNY strategist Wee Khoon Chong.Financial Times
These developments align closely with Beijing’s broader self-reliance strategy, especially amid tightened U.S. export restrictions on advanced AI processors.
Who Could Be Supplying the Chips?
While DeepSeek didn’t disclose specific chipmakers, speculation is growing around several Chinese firms including Huawei, Cambricon, Moore Threads, Hygon, and MetaX. Industry insiders suggest that V3.1 could support multiple vendors rather than just one.
Earlier in 2025, major domestic players—including Huawei’s Ascend, Moore Threads, and Hygon—announced that they support DeepSeek’s R1 and V3 models. Alibaba, Baidu, and Tencent also made DeepSeek accessible through their cloud platforms.
Challenges and Prior Issues
Compatibility hasn’t been seamless. DeepSeek’s previous model, R2, faced setbacks in training on Huawei’s Ascend chips due to hardware instability, prompting a fallback to Nvidia GPUs while still using Ascend for inference.
What to Watch Next
- Model deployment: Will V3.1 run both training and inference natively on domestic chips, or remain limited to inference?
- Partner confirmations: Which domestic chipmakers will officially support DeepSeek’s V3.1?
- Market expansion: Broader adoption among cloud service providers and automakers could further accelerate China’s AI chip ecosystem.
