The DeepSeek-R1 artificial intelligence (AI) model is currently hosted on ModelArts Studio, a platform owned by the China-based technology company Huawei. Although the chipmaker has not specified which chipsets support this AI model, a source suggests that it operates using Ascend 910C GPUs. This revelation has sparked discussions about whether DeepSeek also utilized similar hardware for training its models, although firm evidence remains elusive. Furthermore, OpenAI has claimed to possess information that indicates DeepSeek employed its proprietary models during the training of the DeepSeek AI models.
Huawei Chipsets Powering DeepSeek-R1
In a recent post on X, formerly known as Twitter, tipster Alexander Doria (@Dorialexander) disseminated a promotional image from Huawei announcing that the distilled DeepSeek-R1 model would now be hosted on the ModelArts Studio platform. The company has labeled the platform “Ascend-adapted,” indicating that its data centers operate on its Ascend series chipsets. However, specific details regarding the GPUs in use have not been disclosed.
The anonymous source has alleged that the inference capabilities of the DeepSeek large language model (LLM) utilize Huawei Ascend 910C chipsets, which are viewed as alternatives to the Nvidia H800, albeit with some performance trade-offs involved.
While the authenticity of this claim remains unverified, it raises significant questions. The DeepSeek-R1 AI model was launched only a week ago, and typically, AI models are optimized to function on the same chipsets on which they were trained.
Although other GPUs might also be adjusted to run the model, doing so is often a labor-intensive process. Should Huawei be able to execute the model’s inference on its Ascend-adapted platform, it may indicate that the same infrastructure was indeed used during the model’s training. Nonetheless, it is important to emphasize that there is no conclusive evidence at this stage to support this potential link.
The release of the DeepSeek-R1 AI model, described as a “black box,” has stirred curiosity and speculation. Despite being released as open-source, the firm has only provided the model weights without revealing details about the datasets or the training methodologies. Additionally, the company’s bold assertion that the entire development process cost just $6 million (approximately Rs. 51.9 crores) has raised eyebrows among industry experts, who express concerns regarding the methodologies employed by the firm.
It’s worth noting that last year, the U.S. government implemented restrictions preventing American GPU manufacturers from selling high-end AI chipsets to China, a move designed to enhance its leadership in AI technology while hindering China’s advancements in the sector.
