With the rapid development of artificial intelligence technologies worldwide, the Chinese artificial intelligence model called DeepSeek has begun to gain an important place. DeepSeek claims that it has managed to significantly outperform its US competitors in terms of both cost-effectiveness and performance. In particular, thanks to the open-source chatbot they developed, this application has topped the App Store in 51 different countries. An important factor in DeepSeek’s success is the artificial intelligence hardware it uses, namely the Ascend chips developed by Huawei. The use of these chips has enabled DeepSeek’s application to attract a great deal of attention with the low costs it provides.

DeepSeek will be powered by Huawei’s chips

Although the DeepSeek R1 big language model (LLM) is trained primarily with Nvidia H100 processors, Huawei’s Ascend 910C chips are used for inference, which enables the model to be used in the real world. The training process, i.e. the ‘learning’ of the model with the data, requires very powerful GPUs. At this point, it is stated that Ascend 910C chips will not be a very powerful alternative for training operations. However, the fact that the chips have sufficient capacity for ‘inference’, which is its main function, ensures that the model can run efficiently and at low cost in practice.

This information was shared by @Dorialexander, one of the important names in the artificial intelligence industry. It is stated that Ascend chips manage to keep costs lower thanks to not requiring high GPU power for training. However, the Ascend 910C’s performance makes it unsuitable for training AI models on a large scale. At this point, Huawei is preparing for the introduction of the Ascend 920C chip, which aims to exceed the performance limits of the Ascend 910C. The new chip is aimed to compete with Nvidia’s Blackwell B200 chipset, the leader in AI operations. This breakthrough may allow Huawei to have a stronger presence in AI hardware in the future and increase its share in the industry.