2026世界人工智能大会（WAIC）：坚定但更务实，数据与半导体成焦点

China (PRC) | Technology WAIC KTA: Determined but More Pragmatic;Data/Semi in Focus China's AI policies focus on open-source models, export of AI/compute toEMs, and int'l collaboration on standards/governance. Semis set to be thekey driver. We see super nodes connected by optical modules/NPO, and3D DRAM stacking, as emerging solutions. China is aggressively catchingupon data generation/labelling and agent harnessing/productization.Robotics still a big theme, but with a pragmatically lower focus onhumanoids. Top picks: AMEC, SMIC, VNET. China's AI strategies focus on developing a China-friendly EM ecosystem.President Xi'saddress at the 2026 WAIC (his first attendance) indicates China's key strategy is to export AI/compute/green energy to EMs, which complements its One-Belt, One-Road program. Thus,Chinese AI players will likely have to maintain an open-source approach, which is less positivefor monetization. Compute and, therefore, semi remain the key success factors in this strategy,implying a high likelihood of big foundry capex ahead. Given that China's Kimi 3 (by MoonShot)is now only 5% behind Anthropic's Fable 5 in terms of intelligence (see here), China's open-source approach is very likey to put pricing pressure on US closed-source models, but willalso make China's models affordable for other EMs. In our view, this strategy could "kill twobirds with one stone." We believe the most likely US response is to 1) tighten export controlsto prevent AI chips from being diverted to China, and 2) require all US AI players to put anti-distillation features in their advanced models. Supernode connected by optics and stacking DRAM on GPUs seem a popular roadmap forChina.Huawei launched Atlas 950 SuperPoD, a supernode design connecting 1,024 Ascend950 GPUs (next gen will be 8,192) using LPO (2,048 lower-powered optical modules). It offersmemory bandwidth of 107.52 TB/s and 2 EFLOPs of compute, vs 20.7 TB/s and 1.4 EFLOPs(FP4) for NVDIA's Vera Rubin NVL72. But it consumes 7.4x more power (1.7MW) than RubinNVL72. Hence, we like China's IDC players. Moreover, we have seen two Chinese GPU playersadopt 3D DRAM tech, stacking 4 layers of DDR5 on top of their GPUs using hybrid bonding inresponse to the HBM constraints. They believe this will deliver 20+ TB/s of bandwidth, similarto HBM4. Stacking will be done by Chinese OSAT players, and the DDR5 by CXMT. But theseplayers' GPUs are based on only 14nm. Therefore, such 100%-localized solutions focus moreon memory bandwidth than compute, likely driven by agentic AI inference demand. We seefoundry capacity (both logic and memory) as the biggest investment areas to support China'sAI strategies. Hence, we like AMEC and SMIC. China catching up on data generation and agent harness capabilities.The US has led indata collection for AI. However, data availability is now a big challenge as text data has beenexhausted, and the remaining data is mostly proprietary at the enterprise level. For the worldmodel, real-life 3D data is hard to collect for individual AI players. Following the "data asa factor of production" policy in 2022, the Bureau of Data was created in Oct 2023. Thisyear, it will likely establish a central SOE to facilitate the collection, labeling, and trading ofhigh-quality and reliable datasets across industries. China is also in a strong position, as thecost of hiring local professionals/experts to verify/contribute datasets remains competitive.Moreover, Chinese AI players have offered agentic products with strong harnessing in contextmgmt, sub-agent orchestration, tool utilization, guardrails, managing feedback loops, etc.Harness is as important as model intelligence for delivering strong agent performance. Edison Lee, CFA * | Equity Analyst852 3743 8009 | edison.lee@jefferies.com Matt Ma * | Equity Analyst852 3767 1109 | matt.ma@jefferies.com Nick Cheng * | Equity Analyst+852 3743 8750 | nick.cheng@jefferies.com Jacky He * | Equity Analyst+852 3743 8084 | jacky.he@jefferies.com Pls see P2 for AI smartphone and robotics. Annie Ping, CFA, FRM * | Equity Associate+852 3767 1273 | annie.ping@jefferies.com AI smartphone remains a big challenge.A Chinese AI player demostrated a self-designed AIsmartphone powered by the smartphone versions of its LLM (1bn-10bn data parameters) insteadof a traditional OS (but the kernel is still Android) and made by an ODM. We believe OpenAI may wantto do something similar. But we do not think the experience is differentiated enough (vs cloud-basedservices). The ecosystem could also be limited as major Internet players may block its APIs forsecurity purposes, or to defend their own ecosystems. A lack of bargaining power in the hardwaresupply chain and high memory prices are additional challenges. Robotics still a big theme with many players, but de-emphasizing humanoids is a realisticstrategy.Robotics occupies ~25% of the WAIC's exhibition space, bigger than last year and thusstill a very important AI theme for China. However, we