WAIC 2026: Enhanced Certainty for China's AI Supply Chain

Overview

This is a summary of UBS's key takeaways from the 2026 World Artificial Intelligence Conference (WAIC) Tech Tour. Through site visits to 12 industry chain companies (covering LLMs, GPUs, networking, analog chips, packaging, ODM, and semiconductor equipment), as well as observations at various booths and keynote speeches, the report summarizes the latest advancements in China's AI technology and infrastructure. The core view is that the visibility for rapid growth in China's AI technology supply chain in 2026 and beyond has significantly improved, with positive signals across domestic computing power, large models, network interconnects, and advanced packaging.

Core views

UBS observed at WAIC 2026 that China's AI ecosystem is exhibiting clearer characteristics of internationalization and industrialization. China announced the establishment of the World Artificial Intelligence Cooperation Organization (WAICO), involving 29 countries, and signed AI governance initiatives; furthermore, new contracts worth approximately $3 billion were signed with overseas companies. In terms of scale, over 1,100 exhibitors and more than 400,000 attendees attended, with unprecedented institutional interest, which UBS believes reflects the momentum and confidence within China's AI industry chain. On the model front, domestic LLM iterations are accelerating, with new models such as Kimi K3 and Qwen 3.8-Max making their debut. Specifically for companies: Zhipu's ARR (Annual Recurring Revenue) reached $1 billion as of July this year, ahead of its 2026 internal target. Management believes the recent HK$31 billion H-share placement will support further computing capacity expansion, while API price hikes and inference optimization are expected to support margin improvement; however, uncertainty in computing leasing costs remains a key variable for margin trajectory. MiniMax's ARR also accelerated growth, doubling from approximately $100 million at the end of 2025 and $150 million disclosed in the Q1 earnings call, before the M3 launch. Management remains confident in reaching the $1 billion ARR target by the end of 2026, primarily driven by rapid growth in API/Token Plans, which accounted for nearly half of total revenue in May (approximately 30% in 2025). Regarding computing and infrastructure, domestic AI accelerator companies generally reported strong demand from Chinese customers (including CSPs), attributed to domestic supply tightness and increased willingness among domestic CSPs to adopt domestic cards. Huawei showcased the Atlas 950 SuperPoD (1,024 cards), based on Ascend 950PR chips and the UnifiedBus interconnect architecture, delivering 1 EFLOPS FP8 / 2 EFLOPS FP4 compute power and 256TB unified memory addressing, emphasizing ultra-low latency of 3us and TB-level NPU interconnects, targeting Agentic AI, MoE models, and long-context inference scenarios. Huawei stated it has deployed over 750 sets of Ascend 384 SuperPoD systems across industries such as internet, telecommunications, finance, and more. Feedback from various links in the supply chain further validates the prosperity of the demand side. On the networking front, Montage Technology noted that DDR5 Gen3/4 memory interface chips will continue to contribute positively in 2026, with Gen5 chips expected to start shipping in the second half of 2026. CXL/MRDIMM demand is rising, and the company participated in CXMT's IPO strategic allocation to strengthen long-term cooperation. Shenchip and Shengke Networks jointly released a 51.2T CPO switch prototype based on a domestic supply chain. For packaging, JCET maintained a utilization rate of 80% in Q2, with AI-related businesses, wafer-level packaging, and bumping fully loaded. A newly announced $10.7 billion RMB investment in a Shanghai advanced packaging plant (2.5D packaging) is scheduled for completion in late 2028 and mass production starting in 2029. Management remains optimistic about the TAM growth for AI power-related packaging in 2025-2027. Huaqin expects data center revenue of RMB 70 billion in 2026 (up 50% YoY), driven by switches, AI servers, and SuperPoD deployments, and guides for recurring profit growth of over 20% in the second half of 2026, accelerating in 2027, due to product mix improvements from increased proportion of SuperPoDs and switches. In the analog and power segment, while consumer electronics demand remains weak for Southchip Semiconductor, it is building advanced analog solutions for AI power and memory power. DrMOS and V-core are targeted to ramp up starting in 2027, and the company has signed a long-term agreement with SMIC to lock in capacity. Dongxin Semiconductor believes the upturn in SLC NAND is driven by the exit of global top manufacturers from this market (supply contraction rather than demand pull). With limited new capacity before the first half of 2027 and high technical barriers at the 1Xnm node, the company stands to gain market share and maintain strong profitability, while investing in downstream GPU, Wi-Fi 7, and MLC NAND to transition. Overall, UBS believes the WAIC 2026 tour reinforces its confidence and visibility in the rapid growth of major Chinese AI technology supply chain companies in 2026. It prioritizes Zhipu in China's AI LLM space (positive fundamentals, attractive valuation) and gives a Buy rating to MiniMax. In the AI-related semiconductor supply chain, it prefers NAURA (WFE localization), JCET (advanced packaging), and Han's Laser (PCB equipment). In the hardware supply chain, it favors Apple-chain names: Luxshare, Lens Technology, and AAC Technologies.

Analysis framework

This is a typical industry chain field research study: UBS does not limit itself to published financial data but utilizes WAIC as an industry platform to conduct intensive C-level/IR interviews with representative companies across the AI value chain (12 companies), combined with booth product observations and keynote information to cross-validate industry prosperity. The institution's logic flows sequentially from "Conference Momentum -> Corporate Operating Data -> Supply Chain Order Feedback -> Product Roadmaps": First, confirm AI industry momentum through macro signals like exhibition size and contract values; second, verify demand authenticity with quantifiable operating metrics such as ARR, utilization rates, and revenue guidance; finally, judge growth sustainability using product iteration plans (e.g., GPU generations, SuperPoDs, DDR5 Gen5, 2.5D packaging). This bottom-up, multi-point cross-validation approach helps the institution conclude that supply chain demand is solid and visibility is enhanced, rather than relying on a single company or indicator. Notably, the institution applied layered valuation methods: PE for mature hardware companies like semiconductor equipment/Apple chains; PB for JCET (asset-heavy OSAT); Price/ARR for MiniMax (software/SaaS-like revenue); P/S with SOTP cross-validation for Zhipu. This reflects industry conventions where AI companies at different life cycle stages require matching valuation anchors.

Methodology notes

• Industry / Industrial Analysis FrameworkUpstream-Midstream-Downstream Transmission

  Examine prosperity at each link along the AI value chain (Large Models -> GPU/Accelerators -> Networking/Optical Interconnects -> Advanced Packaging -> ODM/Servers -> Analog/Memory Chips) and judge overall supply chain demand through multi-point cross-validation.

  The report does not look at isolated links but strings together feedback from companies across all segments, from large models to storage/analog. When multiple upstream and downstream links provide positive signals, the judgment on overall AI demand becomes more credible.

• Cycle and Prosperity FrameworkInventory cycle (Kitchin)

  When analyzing Dongxin Semiconductor, the report adopts an inventory cycle perspective: the upturn in SLC NAND is driven by supply-side contraction rather than demand surge.

  Dongxin management pointed out that the exit of global top manufacturers from the SLC NAND market led to insufficient supply, and downstream clients failed to build inventories due to skyrocketing memory prices, leaving current inventory low. The report leverages this supply-driven upcycle to assess the sustainability of the company's profitability, focusing on the key timeframe of lacking new capacity until the first half of 2027.

• Valuation MethodologyPE/PEG valuation

  Apply PE valuation to relatively mature hardware companies such as NAURA, Luxshare, Lens Technology, AAC Technologies, and Han's Laser.

  These companies have mature business models and high growth visibility; PE (Price-to-Earnings) is the most intuitive and common valuation method in the market, suitable for comparing relative valuations of similar hardware companies.

• Valuation MethodologyPB valuation

  Apply PB valuation to Longjietech/JCET.

  OSAT is an asset-heavy industry where profits are significantly affected by capacity cycles and depreciation; PB (Price-to-Book) better reflects the true value of such asset-heavy companies. When the industry is in an upswing and capacity utilization rises, PB valuation aligns better with the logic of book value repair.

• Valuation MethodologySOTP (Sum-of-the-Parts)

  Use P/S valuation for Zhipu, cross-validated with SOTP.

  SOTP values each business segment separately and sums them up, suitable for companies with diverse businesses or those in early stages where overall profitability is not yet realized but segment values differ significantly. Zhipu has both ARR-driven commercial lines and strategic assets like cutting-edge model capabilities; SOTP allows for a more comprehensive valuation.

• Company Fundamentals and Financial FrameworkProfit Quality Analysis

  The report measures the quality and sustainability of AI company growth through indicators such as ARR, API revenue share, gross margin trajectory, utilization rates, and recurring profit growth.

  For LLM companies (Zhipu, MiniMax), the report focuses on the absolute level of ARR and the proportion of API revenue, as these reflect the recurrence of income and business model health. For manufacturing companies (JCET, Huaqin), it looks at utilization rates and recurring profit growth to assess the quality of profit improvement.

Asset mapping & comparison

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

• Zhipu/Knowledge Atlas Technology (2513.HK)

  Top choice for China AI LLM, ARR met ahead of schedule, positive fundamentals, attractive valuation

  Strengths

  ARR reached $1 billion and annual target achieved ahead of time; HK$31 billion H-share placement supports computing expansion; API price hikes and inference optimization support margin improvement

  Weaknesses

  Computing remains a key constraint; Margin improvement pace affected by uncertainty in computing leasing costs

  Comparison

  UBS's top choice in China AI LLM, also gives Buy rating to MiniMax

  Risks

  Domestic computing supply tightness and rising leasing prices may suppress margin elasticity

• MiniMax (0100.HK)

  Buy target in China AI LLM, ARR growing rapidly, API/Token Plan revenue share increasing quickly

  Strengths

  ARR climbing rapidly from $100 million towards $1 billion goal; M3 flagship model parameters close to twice those of M2.7; Early computing procurement brings cost advantages

  Weaknesses

  2C subscription revenue is stable but growth mainly relies on API; Computing costs remain a balancing item for gross margins

  Comparison

  Both are core LLM targets; UBS holds positive views on both, but prioritizes Zhipu

  Risks

  Fluctuating computing costs may affect gross margin health

• Longjietech/JCET (600584.SS)

  Core beneficiary of AI advanced packaging, UBS rated Buy

  Strengths

  Q2 utilization rate 80%, AI packaging fully loaded; Planning 7.8 billion RMB Shanghai 2.5D packaging plant; AI power-related packaging TAM expected to grow several times between 2025-2027

  Weaknesses

  New capacity expected only in 2029 for mass production, limited short-term contribution; Tight supply of upstream materials like BT substrates

  Comparison

  Among preferred targets in AI-related semiconductor supply chain alongside NAURA and Han's Laser; JCET primarily positioned in advanced packaging

  Risks

  Tight supply of upstream packaging materials (BT substrates, electronic glass cloth) may limit near-term upside

• NAURA (002371.SZ)

  Beneficiary of semiconductor equipment localization (WFE localisation), UBS APAC Key Call, Buy

  Strengths

  Benefits from trends in semiconductor equipment localization

  Comparison

  Alongside JCET and Han's Laser, listed as preferred target in AI semiconductor supply chain

• Huaqin (3296.HK/603296.SS)

  Beneficiary of AI data center ODM, strong data center revenue guidance

  Strengths

  Data center revenue in 2026 expected to reach 70 billion RMB (+50%), driven by switches and AI servers; Product mix improvements drive profit growth exceeding 20%

  Weaknesses

  Smartphone shipment volumes declining; While maintaining stable revenue through customer structure improvement, traditional business growth is limited

  Comparison

  Positioned to capture data center opportunities in ODM segment, hedging against consumer electronics ODM business

• Montage Technology (6809.HK, Buy)

  Beneficiary of AI server memory interface and networking chips, Buy

  Strengths

  DDR5 Gen3/4 contribution continues to rise; Gen5 expected to ship in H2 2026; PCle retimer benefits from 8-card AI server demand; Participated in CXMT IPO strategic allocation to strengthen cooperation

  Weaknesses

  MRCD/MDB volume requires next-gen X86 platforms (late 2026 to early 2027); CXL commercialization still needs 2-3 years

  Comparison

  Positioned in server upgrade and CXL trends within memory interface chip sector

Key data

• WAIC 2026 Attendance ScaleOver 1,100 exhibitors, over 400,000 attendeesReflects soaring heat in China's AI industry and unprecedented institutional interest
• China-Overseas Contracts SignedApproximately $3 billionDuring WAIC, China signed new contracts worth approximately $3 billion with overseas companies, reflecting global collaboration
• Zhipu ARR$1 billion (as of July 2026)Reached 2026 internal target ahead of schedule, supported by increased token usage and API price hikes
• MiniMax ARRFrom ~$100 million (End of 2025) -> ~$150 million (Q1 2026) -> Doubled before M3 launchManagement remains confident in achieving $1 billion ARR by end of 2026; API/Token Plan revenue share approached 50% in May, ~30% in 2025
• Huawei Atlas 950 SuperPoD1,024 cards, 1 EFLOPS FP8 / 2 EFLOPS FP4, 256TB Unified MemoryTargeted at trillion-parameter model training and inference, emphasizes ultra-low latency of 3us
• Huawei Ascend 384 SuperPoD DeploymentsOver 750 setsCovers internet, telecommunications, finance, education, healthcare, transportation, manufacturing, and other industries
• JCET Q2 Utilization Rate80%AI-related businesses, wafer-level packaging, and bumping fully loaded; New Shanghai advanced packaging plant investment of 7.8 billion RMB, construction completed in H2 2028, mass production in 2029
• Huaqin 2026 Data Center Revenue Guidance70 billion RMB (YoY +50%)Driven by switches, AI servers, and SuperPoD deployments; Recurring profit growth >20% in H2 2026
• Dongxin SLC NAND Prosperity LogicSupply-side Contraction DrivenGlobal top manufacturers exited the market; Lack of new capacity before H1 2027; High 1Xnm technical barriers

Impact & implications

The report argues that information conveyed by WAIC 2026 reinforces the long-term growth logic of China's AI technology supply chain: Demand has expanded from model training to inference, edge AI, autonomous driving, and other implementation scenarios, while the urgency of localization (due to domestic supply tightness) provides structural opportunities for local computing, networking, packaging, and analog sectors. Specifically, the computing sector (GPU/Accelerators, SuperPoDs) benefits from increased willingness among domestic CSPs to switch to domestic cards. Networking and optical interconnect segments benefit from demands for large cluster interconnects and rack-level expansion. Advanced packaging benefits from the ramp-up of domestic AI accelerators and spillover effects into overseas markets. ODM and server manufacturers benefit from structural shifts in data center capex. Analog and memory chip manufacturers open a second growth curve by leveraging supply contraction and new AI power/memory demands. In terms of stock selection, the report prioritizes Zhipu in the LLM sector, prefers NAURA, JCET, and Han's Laser in the semiconductor supply chain, and favors Luxshare, Lens Technology, and AAC Technologies in the hardware supply chain (Apple chain). These preferences reflect the institution's multiple judgments on localization, advanced packaging, equipment localization, and the cyclical innovation of consumer electronics.

Risks

• Tight domestic AI computing supply and rising leasing costs may suppress the pace of LLM company margin improvement
• Tight supply of upstream packaging materials (BT substrates, electronic glass cloth) may limit near-term upside for advanced packaging
• Rapid technological changes, intensifying competition, and macroeconomic cycle impacts in the tech sector make financial forecasting difficult and valuation challenging for tech companies

What to watch

• Sustainability of computing expansion and ARR milestone achievement for Zhipu and MiniMax, particularly the realization of the $1 billion ARR goal by end of 2026
• New product launches and mass production ramp-up schedules for domestic GPU/Accelerator companies (e.g., Biren BR20X)
• Shipment of Montage DDR5 Gen5, maturity of CXL ecosystem, and progress in MRDIMM volume ramp-up
• Construction progress of JCET's Shanghai advanced packaging plant and increase in 2.5D packaging share
• Supply landscape of Dongxin SLC NAND and timing of new capacity investments (before H1 2027)
• Increasing localization rate of domestic AI demand and subsequent landing of overseas contract collaborations