China Internet & New Media: Takeaways from an LLM expert call China’s LLM evolution: scaling, agent integration and cost optimization

Research evidence excerpt

China Internet & New Media: Takeaways from an LLM expert call China’s LLM evolution: scaling, agent integration and cost optimization

Global Markets Research

27 July 2026China Internet & New Media

EQUITY: INTERNET & NEW MEDIA

Takeaways from an LLM expert call Research Analysts

China Internet & New Media

China’s LLM evolution: scaling, agent integration and cost Rachel Guo - NIHK

rachel.guo@nomura.comoptimization

+852 2252 1400

Key highlights Jialong Shi - NIHK

The Nomura China internet team held a group call with an expert from a leading Chinese Jialong.shi@nomura.com

large language model (LLM) player. According to the expert, the next phase of +852 2252 1409

competition among Chinese LLM developers will be shaped by three interrelated trends: China Technology

continued model scaling, deeper model-harness integration and greater emphasis on Bing Duan - NIHK

system-level token economics. DeepSeek (unlisted) appears positioned at the bing.duan1@nomura.com

intersection of these trends, according to the expert, seeking to narrow the capability gap +852 2252 2141

with overseas frontier models while maintaining a structurally lower cost base and

substantially lower token prices.

Model scaling continues, but efficiency also matters

Chinese LLM are likely to move toward multi-trillion-parameter architectures, supported by

larger compute clusters, supernodes, high-speed interconnects and optical networking.

The expert views approximately 3tn parameters as the next practical milestone for

Chinese LLMs.

However, pure parameter scaling is approaching diminishing returns. Beyond parameter

scale, model competitiveness will increasingly depend on architectural and post-training

efficiency, memory management, reliable tool use and long-horizon task performance.