DeepSeek CEO Liang Wenfeng believes Nvidia is gradually losing its advantage as China’s domestic AI systems draw closer in compute capability, despite a significant remaining gap.
In an exchange with investors, he stated that to train an AI model at a scale equivalent to the largest models today, DeepSeek would need about 50,000 Nvidia GB300 GPUs or around 200,000 Huawei Ascend 950 GPUs.
According to him, the biggest gap between DeepSeek and US AI companies today lies not in algorithms, but in compute resources.
Liang revealed that DeepSeek’s current compute capacity is only equivalent to about 20,000 Nvidia H100 GPUs, with most of these resources added within the “past month or two.” The company is expanding its infrastructure rapidly.
He suggested that Huawei’s Atlas 950 SuperPoD system can functionally and deployment-wise replace Nvidia GB200 and GB300 clusters at the system level, even though it requires using more GPUs.
However, Liang also admitted that to achieve equivalent performance, DeepSeek needs about 4 new Huawei Ascend 950 GPUs to equal 1 Nvidia GB300 GPU. He estimated that Huawei is currently about 2 years behind Nvidia in GPU technology.
According to DeepSeek’s CEO, even if deploying Huawei hardware might cost 50%, 100%, or even 200% more, it remains acceptable if China can secure its own supply chain and reduce dependency on US technology.
Liang reaffirmed that an open-source AI strategy is a crucial direction for DeepSeek, while arguing that AI labs pursuing a “money-burning to dominate” model will find it hard to survive in the long run.
DeepSeek is also investing heavily in high-quality data annotation. According to Liang, this is an area where OpenAI and Anthropic are years ahead, and it has now become a critical bottleneck in developing advanced AI models.
The above statements are reported to originate from an investor exchange leaked on social media. At the time of publication, DeepSeek had not officially released the full transcript or fully confirmed all quoted statements.
📌 Conclusion: DeepSeek admits that Nvidia still leads in GPU performance, with 1 GB300 GPU said to equal roughly 4 Huawei Ascend 950 GPUs, and Huawei lagging by about 2 years in technology. However, CEO Liang Wenfeng argues that Nvidia’s advantage could erode if China successfully builds a sufficiently large domestic AI ecosystem, accepting higher hardware costs in exchange for autonomy. These statements were cited in the article from an investor exchange and have not been officially confirmed by DeepSeek.
