DeepSeek: Beyond the Hype - Bernstein Report Sheds Light

A Bernstein report addresses the buzz around AI company DeepSeek. While the models are impressive, claims of building OpenAI equivalents for USD 5 million are misleading. DeepSeek's models use advanced techniques but require extensive resources, challenging exaggerated claims circulating on social media.

In the midst of growing excitement on social media and stock markets about the AI company DeepSeek, a Bernstein report has offered a reality check. Contrary to popular belief, DeepSeek's success did not come at a mere USD 5 million price tag.

The Bernstein report highlights that the claims of creating an OpenAI-like model for such a minimal sum are deceptive. DeepSeek has created two main AI model families: 'DeepSeek-V3' and 'DeepSeek R1', both exhibiting remarkable capabilities but not without significant investment.

DeepSeek-V3 employs a Mixture-of-Experts architecture to optimize performance while conserving resources, operating with 671 billion parameters. Training required 2,048 NVIDIA H800 GPUs over two months. Meanwhile, the R1 model, built on V3, shows advanced reasoning skills, likened to OpenAI's models, demanding yet undisclosed resources for development.

(With inputs from agencies.)