Key Takeaways:
Chinese AI developers unveiled models with up to 3 trillion parameters at WAIC, matching frontier US capabilities at lower cost.
Chinese AI developers unveiled a wave of large language models with up to 3 trillion parameters at the World Artificial Intelligence Conference in Shanghai, matching the capability of frontier US systems while pricing API access at a fraction of the cost.
"The rapid iteration of domestic models has accelerated impressively, with leading models reaching 2-3T parameter scales comparable to Claude Opus 4.8," China Merchants Securities International said in a report, "while maintaining cost efficiency and open-source advantages."
Moonshot AI launched Kimi K3 with 2.8 trillion parameters and a 1-million-token context window. Alibaba released the preview version of Qwen3.8Max at 2.4 trillion parameters, while DeepSeek is preparing to launch its official V4. Zhipu AI, Minimax and Tencent also debuted new models — GLM-5.2, M3 and HY3 respectively — at the conference, which ran from July 17 to July 20 under the theme "Intelligent Partners, Co-create the Future."
The accelerated iteration cycle threatens to compress margins for US AI leaders as Chinese developers undercut on price. CMSI maintained its "Strongly Recommend" rating on the sector, naming Alibaba, Tencent, Kuaishou, Zhipu AI and Minimax as key investment themes.
Upstream, the competitive picture is shifting from individual chip specs to system-level architecture. Huawei and Alibaba's T-Head both showcased supernodes that link hundreds to thousands of chips into unified computing pools through high-speed interconnection, using architectural innovation to offset the performance gap of individual processors.
The mass production of Huawei's Ascend 950 series in the second half of 2026 will be a significant catalyst for large-model efficiency, CMSI said, driving down inference costs and improving gross margins for model companies. The Ascend 950 competes directly with Nvidia's H100 and B200 in the Chinese market, where US export controls have restricted access to advanced AI chips.
Alibaba, trading on the Hong Kong exchange, has the most to gain from the dual tailwind of its own Qwen model family and T-Head's hardware capabilities. Tencent's consumer ecosystem advantages and Kuaishou's Kling AI video generation model provide differentiated exposure, while Zhipu AI's SOTA coding capabilities and Minimax's native full-modality approach target specific high-value segments.
The lower API pricing relative to US models may pressure near-term revenue but could drive faster adoption across China's enterprise and developer ecosystem. CMSI expects the combination of model iteration and domestic chip output to create a self-reinforcing cycle: cheaper inference enables broader deployment, which generates more data for model improvement.
This article is for informational purposes only and does not constitute investment advice.
