DeepSeek-R1 AI Model Said to Be Powered by Huawei Chipsets

DeepSeek-R1 artificial intelligence (AI) model is being hosted on China-based Huawei's ModelArts Studio platform. While the Chipmaker has not detailed the chipsets being used to power the AI model, a tipster claimed that it is being run using the Ascend 910C GPUs. The leak has led to speculations on whether the Chinese AI firm also trained its models on the same hardware infrastructure, although there is no conclusive evidence available for this. Notably, OpenAI alleged that it has evidence of DeepSeek using its proprietary models to train the DeepSeek AI models.

Huawei Chipsets Powering DeepSeek-R1

In a post on X (formerly known as Twitter), tipster Alexander Doria (@Dorialexander) shared a promotional image from Huawei announcing that the distilled version of the DeepSeek-R1 model will now be hosted on its ModelArts Studio platform. Notably, the chipmaker calls the platform “Ascend-adapted,” highlighting that the data centres are being powered by its Ascend series chipsets. However, the tech giant has not revealed which particular GPUs are being used.

The tipster claimed that the inference of the DeepSeek large language model (LLM) is being powered by the Huawei Ascend 910C chipsets, which is considered to be an alternative to the Nvidia H800, although there are some performance trade-offs.

While the legitimacy of the claim cannot be verified, there is an important case to be made here. The DeepSeek-R1 AI model was released just a week ago. Usually, AI models are optimised to run on the same chipset it has trained on.

While other GPUs can also be optimised to run the model, it usually is a time-consuming process. If Huawei can run the model's inference on its Ascend-adapted platform, there could be a possibility that the same infrastructure was used to train the model. However, it is noted that there is no conclusive evidence available for this correlation.

DeepSeek-R1 AI model's “black box” release is shrouded in several mysteries which has led to these speculations. Despite the open-source release, the AI firm only released the model weights, not disclosing the datasets or the training process. In addition, the tall claims made by the firm such as the total cost taking just $6 million (roughly Rs. 51.9 crores) have resulted in several industry veterans raising concerns over the methodologies used by the firm.

Notably, last year, the US Government put restrictions on US-based GPU manufacturers to not sell flagship-tier AI chipsets to China, to solidify its position as AI leader and to slow down China's AI development.