Nomura Global AI Trend Tracker: Will DeepSeek V4 or R2 Disrupt the Global AI Value Chain Again?

Nomura expects DeepSeek's forthcoming V4 large language model, anticipated in mid-February 2026, to achieve breakthroughs in programming, ultra-long code processing, and reasoning reliability. While these advances should accelerate China's AI innovation cycle and narrow the gap with global peers, Nomura argues V4 is unlikely to disrupt global AI infrastructure markets as severely as V3/R1 did a year ago. Major cloud service providers continue pursuing AGI and sustaining compute investments.

Two recent DeepSeek papers may underpin V4's architecture. The mHC (Manifold-Constrained Hyper-Connections) mechanism creates multiple internal information flows with double stochastic matrix constraints, resolving deep network bottlenecks to improve training efficiency. The Engram module decouples static knowledge retrieval from dynamic reasoning using sparse memory tables in system DRAM, potentially lowering both training and inference costs.

These innovations should benefit Chinese AI hardware and application companies. Nomura highlights ZTE (neutral), whose Doubao AI phone demonstrates multi-step task execution, and Alibaba (buy), whose Qwen APP performs similarly. In the software sector, Kingsoft Office and Kingdee International are top picks. Historically, V3/R1 demonstrated that low-cost, high-performance models did not reduce actual compute demand but rather stimulated accelerated global CSP infrastructure investment. However, DeepSeek's token consumption share has declined as the market becomes more fragmented.