DeepSeek Announces Next-Generation AI Model Tailored for China’s Emerging Homegrown Chip Ecosystem

Chinese AI startup DeepSeek has suggested that China is poised to introduce domestically produced “next-generation” chips to support its artificial intelligence (AI) models, coinciding with an update to one of its large language models.

In a post on its official WeChat account, DeepSeek revealed that the precision format UE8M0 FP8 of its newly launched V3.1 model is optimized for upcoming domestically developed chips. The 8-bit floating point (FP8) data processing format enhances computational efficiency for training and inference of large deep learning models.

DeepSeek’s reference to China’s forthcoming next-generation chips may indicate a shift towards closer collaboration with China’s burgeoning AI chip industry amid Washington’s stringent semiconductor export regulations and Beijing’s push for self-reliance in chip production.

The comments were made approximately two weeks after Beijing reportedly urged Chinese AI developers to consider domestic alternatives to Nvidia’s graphics processing units (GPUs) used in AI training. Although China’s domestic AI chip manufacturers are deemed to trail behind Nvidia in technological advancement and scale, companies like Huawei have been making strides.

In its recent post, DeepSeek declined to disclose the chips employed for training the V3.1 or which local chips the UE8M0 FP8 might be compatible with.

Earlier this year, DeepSeek made waves in the tech industry by releasing its R1 reasoning model, which showcased performance akin to that of Western competitors such as OpenAI, despite U.S. export controls preventing it from utilizing Nvidia’s most advanced AI training chips. Prior to this, in December, the company unveiled its V3 model, which it claimed had been trained on around 2,000 of Nvidia’s less advanced chips.

Following DeepSeek’s model innovations, the U.S. implemented tighter export restrictions in April, effectively prohibiting Nvidia’s H20 chips, specifically designed to meet prior export restrictions on China. Although officials from the Trump administration announced plans to allow Nvidia to resume shipping the chips to China last month, the H20s are now under review in China, with regulators reportedly instructing companies against purchasing the chips until a national security assessment is completed.

Chip analysts have informed CNBC that companies like Huawei, which have been endeavoring to establish an alternative AI chip ecosystem in China, could potentially benefit from a scarcity of Nvidia’s H20s in the market.

DeepSeek announced on Thursday that its V3.1 features “significant changes,” including improved response times, and a hybrid reasoning architecture that enables the model to support both reasoning and non-reasoning modes. Reasoning models are capable of executing more complex tasks through a step-by-step logical thought process.

Starting on September 6, DeepSeek will also modify the pricing for using the model’s API, which allows developers of other apps and web products to integrate DeepSeek on their platforms.