WAIC shows China's AI path remains determined but more pragmatic, with semiconductors and data centers becoming the core beneficiaries

Overview

This report is Jefferies' post-WAIC view on China's technology and AI industry chain in 2026. The core conclusion is that China's AI policy direction remains firm, but implementation is becoming more pragmatic: serving the emerging-market ecosystem through open-source models, exports of AI/compute power/green energy, and international standards cooperation; meanwhile, compute power, semiconductor manufacturing, memory, optical interconnects, data resources, and agent productization will become the key drivers of industry progress.

Core views

First, Chinese AI players may continue to maintain an open-source strategy, which helps lower adoption barriers for emerging-market customers and may also create pricing pressure on U.S. closed-source models, but is unfavorable for direct model monetization. Second, compute power remains the strategic core, with optical interconnect supernodes, NPO/LPO optical modules, 3D DRAM stacking, and domestic wafer foundry capacity as key routes. Third, China is building out capabilities in data generation, labeling, trading, and agent orchestration. Fourth, AI smartphones still lack sufficient differentiation in user experience and face ecosystem constraints; robotics remains a major theme, but is shifting from humanoids to more easily commercialized wheeled robots.

Analysis framework

The report conducts thematic research by combining on-site WAIC observations, policy signals, industry-chain technology roadmaps, company valuations, and risk disclosures. The analysis focuses on China's AI policy orientation, its competitive relationship with U.S. model and chip ecosystems, supernode and memory-stacking pathways, data-factor policy, agent productization capabilities, and the commercialization challenges of AI smartphones and robotics.

Methodology notes

• Industrial policy analysisAI strategy and emerging-market ecosystem framework

  Open-source models and AI/compute power exports

  The report interprets China's AI path as building a friendly ecosystem for emerging markets. Open source helps expand usage and reduce costs, but weakens model monetization capability.

• Technology roadmap analysisCompute bottleneck and localization substitution framework

  Optical-interconnect supernodes and 3D DRAM stacking

  Against the backdrop of constrained advanced GPUs and HBM, Chinese vendors are increasing memory bandwidth through supernode optical connections and DDR5 stacking, driving demand for wafer foundries, OSAT, memory, and IDC.

• Valuation methodologyP/E, P/B, SOTP, and target price framework

  Company target price methodology

  AMEC target price is based on 2026E/2027E P/E; NVIDIA target price is based on a CY28E EPS multiple; SMIC target price is based on 2026E P/B; VNET target price is based on SOTP.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Advanced Micro-Fabrication Eqp Inc China (688012 CH)

  Beneficiary in Chinese semiconductor equipment; Top pick; BUY

  Strengths

  China's AI strategy requires more logic and memory capacity, and fab capex expectations are strong, benefiting demand for domestic equipment.

  Weaknesses

  Affected by customers' expansion pace, new customer expansion, and progress in new tool development.

  Comparison

  Compared with the model layer, the equipment segment benefits more directly from compute power and localized capacity construction.

  Risks

  Further U.S. Department of Commerce sanctions, customer expansion slower than expected, slower-than-expected acquisition of new customers outside China, and slower-than-expected development of new tools.

• Semiconductor Manufacturing International Corporation (981 HK)

  Beneficiary in Chinese wafer foundry; Top pick; BUY

  Strengths

  The report believes China's AI path requires substantial domestic logic and memory foundry capacity, placing SMIC in a core position.

  Weaknesses

  Changes in mature-node equipment supply and demand may affect supply-demand balance and ASP.

  Comparison

  Compared with advanced overseas process nodes, China's domestic solutions place more emphasis on availability and memory-bandwidth orientation.

  Risks

  Continued U.S. approval of mature-node supply licenses to U.S. semiconductor equipment companies may ease supply chain tightness; stronger-than-expected demand for mature nodes such as 5G, IoT, and CIS may continue to drive ASP upward.

• VNET Group (VNET)

  Beneficiary in Chinese IDC; Top pick; BUY

  Strengths

  Supernode solutions have high power consumption, and AI compute expansion increases demand for data centers, power, and server rooms.

  Weaknesses

  Dependent on demand growth, access to power quotas, and the market pricing environment.

  Comparison

  Under high-power domestic AI compute solutions, the infrastructure value of the IDC segment is more prominent.

  Risks

  Demand growth slower than expected, inability to obtain sufficient government power quotas, and intensified competition leading to lower-than-expected pricing.

• NVIDIA Corporation (NVDA)

  Global AI GPU leader; listed as BUY in the report

  Strengths

  Vera Rubin NVL72 is used as the benchmark for performance and energy-efficiency comparison against China's supernode solutions as a globally advanced AI compute platform.

  Weaknesses

  China's open-source models and domestic compute substitution may alter parts of the competitive landscape.

  Comparison

  Under the report's metrics, China's Atlas 950 SuperPoD has higher memory bandwidth and compute indicators, but significantly higher power consumption.

  Risks

  Competition and ASP pressure from INTC, AMD, and hyperscaler in-house ASICs; slower data center capex from enterprises and hyperscalers; and slower-than-expected ramp-up of the automotive platform.

Key data

• Huawei Atlas 950 SuperPoD scale1,024 Ascend950 GPUs, with the next generation at 8,192Supernodes are connected through 2,048 low-power optical modules.
• Atlas 950 SuperPoD performance107.52 TB/s memory bandwidth; 2 EFLOPs compute powerThe report compares this with NVIDIA Vera Rubin NVL72's 20.7 TB/s and 1.4 EFLOPs.
• Atlas 950 SuperPoD power consumption1.7MW, about 7.4x that of Rubin NVL72High power consumption underpins the report's bullish view on Chinese IDC players.
• 3D DRAM roadmap of Chinese GPU vendors4-layer DDR5 stacking, targeting 20+ TB/s bandwidthThe report says the bandwidth is close to HBM4, but the GPU process node is still 14nm.
• Share of robot exhibition area at WAICAbout 25%Robotics remains an important theme, but humanoid robot displays declined while wheeled robots received more attention.
• AMEC688012 CH；CNY350.69；BUY；PT Rmb415.00Target price is based on 80x/58x 2026E/27E P/E.
• NVIDIA CorporationNVDA；$202.81；BUY；PT $300Target price corresponds to 21x CY28E EPS of $14.14.
• SMIC981 HK；HK$67.70；BUYThe Hong Kong target price is based on 4.0x 2026E P/B, and the A-share target price is based on 7.0x 2026E P/B.
• VNET GroupVNET；$7.39；BUYTarget price is based on SOTP.

Impact & implications

In investment terms, the report is more tilted toward segments in the industry chain that provide compute supply and localized capacity buildout, including semiconductor equipment, wafer foundries, memory-related packaging, optical interconnects, and IDC. Open-source models may compress the charging power of the model layer, but will expand AI application adoption and penetration in emerging markets. Potential U.S. responses include tighter export controls and requiring advanced models to have anti-distillation capabilities, which will continue to affect China's compute supply chain and the pace of localization substitution.

Risks

• Further tightening of U.S. export controls on AI chips and semiconductor equipment could affect the pace of China's compute and capacity buildout.
• The open-source model path may compress pricing and commercialization potential at the model layer.
• The high power consumption of domestic supernode solutions puts pressure on power supply, IDC capacity, and operating costs.
• Although 3D DRAM stacking improves bandwidth, the underlying GPU is still on 14nm, so computing power may remain constrained.
• AI smartphones lack sufficient differentiation and may face challenges such as internet platform API blockades, weak bargaining power in the hardware supply chain, and high memory prices.
• Humanoid robot commercialization still faces constraints in compute power, memory, thermal management, dexterous hands, motors, and reducers.

What to watch

• Whether China establishes a central data SOE to promote the collection, labeling, and trading of high-quality datasets.
• U.S. policy responses regarding AI chip transshipment, model anti-distillation, and control of advanced models.
• The pace of Chinese fab capex and expansion of logic and memory capacity.
• The subsequent scale, power consumption, deployment progress, and IDC demand pull from Huawei Atlas supernodes.
• Mass production, bandwidth performance, and OSAT/CXMT ecosystem progress for Chinese GPU vendors' 3D DRAM stacking solutions.
• The actual post-WAIC commercialization progress of wheeled robots, humanoid robots, and AI smartphones.