In a remarkable twist, DeepSeek is leveraging Huawei’s latest Ascend 910C chips for its AI model’s inference tasks, signaling significant progress in China’s AI landscape.
Recently, DeepSeek has been in the spotlight, rattling global AI markets with the launch of its R1 LLM model, which led to a $2 trillion dent in the US stock market. This wave of investor anxiety stems from the uncertainties surrounding DeepSeek’s computing capabilities. An intriguing piece of this puzzle involves Huawei, as insider @dorialexander has revealed that DeepSeek’s R1 model employs the Ascend 910C chips for inference, adding an unexpected layer to the story.
For some context, Huawei’s Ascend 910C AI chip stands as a formidable contender to NVIDIA’s Hopper H100 AI accelerators. Even though specific details about Huawei’s chip are still under wraps, the tech powerhouse plans to enter mass production in the first quarter of 2025, drawing considerable attention from major Chinese AI players like ByteDance and Tencent. Touted as a premium alternative to NVIDIA’s offerings within China, Huawei’s chips are steadily capturing a larger share of the market, indicating a promising uptick in popularity.
There’s still a fog of confusion surrounding the computational resources behind DeepSeek’s R1 AI model. It’s crucial to note that the much-discussed “$5 million” is not the entire training budget but rather the operating cost for the finalized model. Moreover, DeepSeek reportedly has access to over 50,000 of NVIDIA’s H100 GPUs, suggesting a reliance on resources comparable to other cutting-edge AI models in its class.
DeepSeek’s choice to use Huawei’s chips for inference is particularly striking. These chips are not only readily available to domestic companies, but they also offer competitive pricing against NVIDIA’s limited editions or even unauthorized accelerators. Huawei’s development of next-gen Ascend AI chips aims to compete with NVIDIA’s upcoming Blackwell products, setting the stage for an intense global tech showdown.