Huawei chips are powering DeepSeek AI

The AI wars are heating up with DeepSeek, a Chinese AI model that claims to surpass US competitors significantly when it comes to cost efficiency. Its open-source chatbot has propelled the app to the top position in the App Store in 51 countries, and it's now revealed that it operates on a Huawei AI chip.

The DeepSeek R1 LLM (large-language model) was trained on Nvidia H100 but uses an Ascend 910C chip for inference, which is the action of using the trained model to generate responses.

I feel this should be a much bigger story: DeepSeek has trained on Nvidia H800 but is running inference on the new home Chinese chips made by Huawei, the 910C. pic.twitter.com/6IAgQlQ3ou

— Alexander Doria (@Dorialexander) January 28, 2025

The information comes from @Dorialexander, who points out that Ascend chips are not dealing with training, so the GPU power requirements are not that high.

However, the Ascend 910C's relatively lower performance limits its suitability for training. Huawei plans to address this issue with the upcoming 920C chip aims to compete with Blackwell B200, the leading Nvidia chipset for AI operations.

The DeepSeek app on iOS