According to @deepseek_ai, DeepSeek has disclosed detailed strategies for training its advanced mixture-of-experts models, DeepSeek-R1 and DeepSeek-V3, by leveraging 2,048 Nvidia H800 GPUs and innovative memory-efficient methods such as FP8 precision. These approaches enabled DeepSeek to achieve significant computational savings, drastically reducing training expenses compared to standard large language model training costs (source: @deepseek_ai, 2024-06-21). This development demonstrates practical opportunities for AI startups and enterprises to scale state-of-the-art models with lower infrastructure investments, accelerating AI adoption in cost-sensitive markets and enhancing accessibility for AI-driven business applications.