DeepSeek — AI Training Cluster

DeepSeek’s AI training cluster in Hangzhou, operated by parent company High-Flyer Quant (幻方科技), gained global attention in early 2025 when its DeepSeek-R1 and V3 models matched GPT-4-class performance at a fraction of the training cost — reportedly $6 million for the V3 training run vs $100M+ for comparable US models. The cluster uses a mix of older NVIDIA H800 chips (permitted for China export, lower memory bandwidth) and domestically-produced hardware, demonstrating that aggressive algorithmic efficiency can compensate for compute restrictions. DeepSeek’s multi-head latent attention (MLA) and mixture-of-experts architecture innovations represent a significant advancement in compute-efficient training. The facility also hosts inference infrastructure serving DeepSeek’s public API, which saw viral adoption worldwide in January 2025.