The DeepSeek-R1 artificial intelligence (AI) model is being hosted on the ModelArts Studio platform headquartered in Huawei, China. While the chipmaker has not yet detailed the chipset used to power the AI model, prompters claim it is running using the Ascend 910C GPU. The leak has raised speculation about whether Chinese AI companies are still training their models on the same hardware infrastructure, although there is no definite evidence. Notably, Openai claims that it has evidence that DeepSeek uses its proprietary model to train the DeepSeek AI model.
Huawei chipset powers DeepSeek-R1
In an article on X (formerly known as Twitter), Tipster Alexander Doria (@Dorialexander) shared Huawei’s promotional image, announcing that the distilled version of the DeepSeek-R1 model will now be hosted on its ModelArts Studio Platform. It is worth noting that chipmakers call the platform a “adaptive upgrade,” which emphasizes that data centers are powered by their rising series of chipsets. However, the tech giant has not yet disclosed which specific GPUs are being used.
Promptenders claim that the inference of the DeepSeek Big Language Model (LLM) was powered by the Huawei Ascend 910c chipset, which is considered an alternative to the NVIDIA H800 despite some performance trade-offs.
Although the legality of the claim cannot be verified, here is an important case. The DeepSeek-R1 AI model was released only a week ago. Typically, AI models are optimized to run on the same chipset being trained.
Although other GPUs can also be optimized to run the model, it is often a time-consuming process. If Huawei can run inferences of models on its rising adaptation platform, it may be possible to use the same infrastructure to train the models. However, it was noted that there is no concluding evidence about this correlation.
The “black box” release of the DeepSeek-R1 AI model is shrouded in several mysteries, which leads to these speculations. Despite the open source release, AI companies only released model weights without revealing data sets or training processes. In addition, the company’s taller claims, such as a total cost of just $6 million (about Rs 51.9 crore), have led to concerns among several industry veterans about the methods the company uses.
It is worth noting that last year, the U.S. government restricted U.S.-based GPU manufacturers from selling flagship AI chipsets to China to consolidate its position as an AI leader and slow down China’s AI development.
