DeepSeek’s AI model built for China’s next-generation chips
Chinese artificial intelligence startup DeepSeek has signalled that China may soon unveil its own “next-generation” chips designed to support the country’s fast-growing AI ecosystem.
In a comment under a post on its official WeChat account, the company stated that its newly released V3.1 model uses a data format known as UE8M0 FP8, tailored for upcoming domestically built processors. FP8, or 8-bit floating point, improves computational efficiency for training and running deep learning models.
The statement suggests that DeepSeek could be preparing to align more closely with China’s emerging AI chip industry, as Beijing urges developers to reduce reliance on U.S. hardware.
Analysts say the move highlights China’s determination to achieve chip self-sufficiency amid Washington’s tightening semiconductor export restrictions.
DeepSeek did not specify which local chips would be compatible with its new model, nor which processors were used to train V3.1.
However, its comments come just weeks after reports that Beijing had instructed AI developers to adopt domestic alternatives to Nvidia’s GPUs, the industry standard for AI training.
Although Chinese chipmakers remain behind Nvidia in both scale and technological advancement, companies such as Huawei have been making steady progress.
The absence of Nvidia’s H20 processors in the market, following a U.S. export ban earlier this year and a pending Chinese security review, could open new opportunities for local manufacturers.
DeepSeek has drawn global attention since the release of its R1 reasoning model earlier this year, which demonstrated capabilities rivalling those of Western competitors such as OpenAI.
The achievement was notable given that U.S. sanctions have limited Chinese firms’ access to Nvidia’s most advanced AI training chips.
The company first released its V3 model in December, trained on approximately 2,000 of Nvidia’s less powerful chips. With V3.1, DeepSeek says it has introduced “major changes”, including faster response times and a hybrid reasoning architecture that can switch between reasoning and non-reasoning modes for complex tasks.
DeepSeek also announced that, starting 6 September, it will adjust its API pricing, allowing developers to integrate the upgraded model into their apps and platforms.
