Nomura: DeepSeek Expert Call → DeepSeek V5: Next-generation flagship model reportedly targeting 1.5–2× the parameters of V4 Pro, with improvements in complex reasoning and long-horizon agent tasks. → Dual-model strategy: V4 Pro pushes frontier intelligence, while V4.1 Flash drives the majority of API revenue through lower-cost inference. → Pricing: DeepSeek offers a 50% discount during off-peak hours to improve compute utilisation, maintain competitive pricing and support margins. → NVIDIA dominance: DeepSeek primarily relies on NVIDIA GPUs for large-scale pre-training. Domestic accelerators are mainly used for adaptation testing and limited-scale training. → Chinese AI chips: Domestic accelerators face significant challenges in interconnect bandwidth, software compatibility and cluster stability. Substitution is expected to happen faster in inference than training. → Inference optimisation: DeepSeek's MLA architecture reduces KV-cache HBM requirements to approximately one-quarter, lowering inference costs. Infrastructure engineering and deployment optimisation are becoming as important as model architecture. → Competitive landscape: Expert favours DeepSeek (cost efficiency), Moonshot AI/Kimi (complex reasoning) and ByteDance (distribution and ecosystem). Future differentiation increasingly depends on inference economics and application integration.