Huawei’s ModelArts Studio Hosts DeepSeek-R1, Fueling Speculation Over Its Training Hardware
Huawei has announced that the distilled version of the DeepSeek-R1 artificial intelligence (AI) model is now available on its ModelArts Studio platform, which runs on its in-house Ascend GPUs. While the Chinese tech giant has not explicitly revealed the exact chipsets powering the model, a tipster claims it is utilizing the Huawei Ascend 910C—considered a potential alternative to Nvidia’s H800, though with some trade-offs in performance.
Is DeepSeek-R1 Trained on Huawei’s Hardware?
The claim, made by tipster Alexander Doria (@Dorialexander) on X (formerly Twitter), has led to growing speculation within the AI community. Typically, AI models are optimized to run inference on the same hardware they were trained on, as adapting them to new GPUs is often a complex and time-consuming process. If DeepSeek-R1 is natively running on Huawei’s Ascend-adapted infrastructure, it raises the possibility that the model was also trained on the same hardware. However, no conclusive evidence supports this theory.
Adding to the mystery, DeepSeek AI has kept many aspects of its development under wraps. Despite an open-source release, the company has only shared model weights without disclosing the training datasets or methodology. Furthermore, DeepSeek’s claim that the entire model was trained for just $6 million has sparked skepticism among industry experts, who question whether such a budget could produce a cutting-edge AI model.
Amidst US Restrictions, China Pushes Forward
The potential use of Huawei’s Ascend chipsets also underscores China’s growing push for AI independence, particularly after the U.S. government imposed restrictions on American chipmakers, preventing them from selling high-end AI GPUs to China. These export bans aimed to curb China’s rapid AI development and maintain the U.S.’s leadership in the field.
For now, DeepSeek-R1’s development and infrastructure remain largely a “black box,” leaving the AI industry to piece together the puzzle. Whether Huawei’s Ascend 910C GPUs played a crucial role in training the model or were merely used for inference is still up for debate.
