DeepSeek, a Chinese startup that has attracted global attention, claims to have developed an AI system on par with OpenAI for just $5 million. However, a report from financial advisory firm Bernstein questions the accuracy of this claim.
According to Bernstein, the $5 million cost only reflects computing resources, not other investments such as research, testing, and development. “DeepSeek is no miracle, and the market panic is somewhat overblown,” the report stressed.
DeepSeek has two main AI models: V3 and R1. V3 uses the MOE architecture, which optimizes performance by combining multiple smaller models. It has 671 billion parameters but only activates 37 billion parameters at a time, reducing computational resources. Training V3 required 2,048 Nvidia H800 GPUs for two months, equivalent to 5.5 million GPU hours.
R1, meanwhile, is based on the V3 platform and uses reinforcement learning (RL) to improve its inference capabilities. While it can compete with OpenAI’s AI on some tasks, Bernstein notes that R1’s development required much more resources than DeepSeek claims.
Still, Bernstein acknowledged that DeepSeek has made significant progress, with pretraining V3 consuming just 2.7 million GPU hours, or about 9% of some leading models. But the idea of building an OpenAI competitor for just $5 million is still considered unrealistic.
