DeepSeek says China set to unveil next-gen AI chips

Chinese AI startup DeepSeek said China will soon roll out domestically made next-generation AI chips, sparking speculation of a breakthrough in the local chip sector.

The Hangzhou-based company revealed the plan on its WeChat account while introducing its new V3.1 model, which uses the UE8M0 FP8 scale and is built to run on upcoming local chips.

DeepSeek did not name the chip supplier or clarify whether the chips are meant for AI training or inference.

The V3.1 model combines reasoning and non-reasoning modes and was trained on the UE8M0 FP8 data format to ensure compatibility with microscaling.

The announcement comes as Chinese firms such as Huawei Technologies and Moore Threads have scrambled to adapt their self-developed chips to run Chinese open-source AI models, including those from DeepSeek, amid US export controls.

DeepSeek did not disclose the chips used to train its most recent models, but its earlier V3 model was trained on 2,048 Nvidia H800 chips.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ Hardware-software integration marks new phase in China’s AI development

DeepSeek’s announcement about “home-grown chips” specifically designed for their AI models signals a shift toward integrated development rather than adapting foreign hardware.

The company’s use of UE8M0 FP8 data format could reduce memory usage by up to 75%, making AI training significantly more efficient on domestic processors compared to standard approaches 1.

This coordination between DeepSeek’s software optimizations and upcoming Chinese chips represents a different strategy than simply creating chip-for-chip replacements of Nvidia products.

Joint research published in June by SiliconFlow and Huawei claimed that Huawei’s Cloud Matrix architecture combined with Ascend chips could run DeepSeek’s R1 models more efficiently than systems using Nvidia’s H800 chips 1.

This suggests Chinese firms are moving beyond performance parity toward specialized hardware-software combinations that could offer distinct advantages for specific AI workloads.

2️⃣ Export controls accelerate China’s chip ecosystem development despite performance gaps

China’s domestic AI chip production is projected to reach 55% by 2027, up dramatically from current levels, driven largely by U.S. export restrictions 2.