摩根士丹利：2026年中国新兴前沿领域：人工智能路径-以更低算力成本实现更高智能回报报告（英文版）

China's AI path – High intelligence density per token cost: Beyond smaller model sizes (parameters), we highlight engineering-efficiency gains in: (1) architecture (dense vs. MoE, attention mechanisms), (2) post-training (reinforcement learning, model distillation), and (3) inferencing infrastructure (hardware optimization, KV cache efficiency). The recent launch of DeepSeek-V4 further proved Chinese models' efforts in efficiency. As a result, China AI labs deliver similar model intel