AI Pioneer's DeepSeek Unveiled: Unveiling True Development Costs

DeepSeek's surprisingly inexpensive AI model challenges industry norms. The company claims to have trained its powerful DeepSeek V3 neural network for a mere $6 million, using only 2048 GPUs, significantly undercutting competitors. However, this figure is misleading.

Image: ensigame.com

DeepSeek V3 leverages innovative technologies: Multi-token Prediction (MTP) for enhanced accuracy and efficiency; Mixture of Experts (MoE), utilizing 256 neural networks, to accelerate training and improve performance; and Multi-head Latent Attention (MLA) to focus on crucial sentence elements, minimizing information loss.

Image: ensigame.com

Contrary to their initial claim, SemiAnalysis revealed DeepSeek's actual infrastructure involves approximately 50,000 Nvidia Hopper GPUs, representing a total investment of roughly $1.6 billion and operational costs of $944 million. This massive investment, coupled with high salaries for its researchers (exceeding $1.3 million annually), contradicts the low training cost narrative.

Image: ensigame.com

DeepSeek's unique structure, as a subsidiary of the High-Flyer hedge fund, allows for direct ownership of data centers and self-funding, fostering agility and rapid innovation. This contrasts with competitors reliant on cloud computing. The $6 million figure only reflects pre-training GPU costs, excluding research, refinement, data processing, and infrastructure. DeepSeek's total investment in AI development surpasses $500 million.

Image: ensigame.com

While DeepSeek's success showcases the potential of a well-funded independent AI company, the "revolutionary budget" claim is an oversimplification. Their competitive edge stems from substantial investment, technological breakthroughs, and a highly skilled team. However, even with these significant expenses, DeepSeek's costs still remain considerably lower than those of its competitors, with previous model training costs at $5 million (R1) compared to ChatGPT's $100 million (ChatGPT4o).