UBS Research on China's AI Industry Chain: Large Models Accelerate Iteration, Funds Begin to Flow to Semiconductors

TL;DR
· UBS is increasingly optimistic about China's AI industry chain after researching 12 tech companies during WAIC, with the core basis not being conceptual enthusiasm, but rather the simultaneous enhancement of model iteration, domestic computing power adoption, and infrastructure demand.
· Semiconductors and equipment have become the fastest-growing sectors in institutional research in July, reflecting that funds are beginning to spread towards domestic chips, equipment, and advanced packaging along with AI demand.
· Domestic AI infrastructure is shifting from single-card competition to SuperPods, network interconnection, and system optimization, essentially compensating for the performance gap of single chips through cluster capabilities.
· Both Zhipu and MiniMax have set their annual recurring revenue targets at $1 billion, but whether revenue growth can translate into profits still depends on inference optimization and computing power leasing costs.
· Changdian Technology maintains high utilization rates for AI-related production capacity and is ramping up 2.5D packaging, indicating that the volume of domestic AI chips is beginning to transmit to the packaging segment.
· UBS believes that the software and services sector remains in a low-crowding zone, which may retain more expectation differences compared to the already heated semiconductor sector.
· The demand for AI data centers is also beginning to spill over into energy storage and power equipment, but order fulfillment cycles, rising lithium prices, and low-price market competition will still suppress short-term profit margins.

On August 14, UBS released "On-Site Observations in China - July 2026." The report combines data from WAIC 2026 technology research, corporate management communications, and institutional research to provide a clearer judgment: investment clues in China's AI industry are spreading from large models themselves to computing power chips, servers, network interconnections, advanced packaging, and power infrastructure.
During WAIC, UBS organized about 40 investors to research 12 tech companies, covering large models, GPUs, networks, analog chips, packaging testing, server ODM, and semiconductor equipment. The core judgment after the research is that China's large models are still rapidly iterating, and the product forms of domestic AI infrastructure are also improving, with downstream demand for AI computing power showing no significant decline.
Domestic AI Competition Shifts from Single Cards to Systems ===============
UBS observed on-site at WAIC that domestic manufacturers are now focusing not just on single chips but on system-level solutions composed of chips, servers, interconnection networks, and cluster management.
Huawei showcased the Atlas 950 SuperPod composed of 1,024 chips; Sugon displayed the Sugon 8000 solution supporting interconnections at the scale of 100,000 cards; MuXi, ZTE, and Alibaba also launched their own super-node products. On the network side, relevant manufacturers began to showcase prototypes of 51.2T CPO switches and complete optical interconnection solutions.
This change means that domestic computing power is reducing the impact of single-chip performance gaps by expanding cluster sizes, improving interconnection efficiency, and optimizing system architecture. For investors, the beneficiary scope has further expanded from GPUs to switching chips, optical interconnections, server ODM, storage interfaces, and advanced packaging.
Domestic AI accelerator manufacturers have reported to UBS that demand from clients such as cloud service providers continues to strengthen. On one hand, overseas high-end computing power remains tight in the Chinese market; on the other hand, domestic cloud manufacturers' willingness to adopt domestic computing cards is increasing. Some second-tier manufacturers are developing products to compete with H100 and H200 while iterating SuperPod technology.
Model Revenue Growth, Computing Power Costs Remain Key Variables =================
The commercialization progress of large model companies is also accelerating, but the quality of growth still depends on computing power costs.
According to Zhipu's management, the company's annual recurring revenue reached $1 billion by July, exceeding the original target set for the end of 2026, mainly driven by the increase in token usage and API price hikes. The company recently completed an H-share placement of approximately HKD 31 billion, which will support computing power expansion while also considering obtaining larger-scale computing resources through leasing.
UBS believes that API price increases and inference optimization may drive improvements in Zhipu's profit margins, but the tight supply of domestic computing power and rising leasing prices remain important variables for profitability. In other words, revenue coming in first does not mean that the unit economics of model companies have stabilized.
MiniMax's ARR has risen from about $100 million at the end of 2025 to approximately $150 million disclosed in the first quarter earnings call, and it is expected to grow more than double before the release of the M3. Management remains confident in achieving the $1 billion ARR target by the end of 2026. The revenue share of API and Token Plan business has increased from about 30% in 2025 to nearly half in May 2026, indicating that growth momentum is shifting from consumer subscriptions to developer and enterprise calls.
However, the parameter scale of MiniMax's new generation M3 model is nearly twice that of M2.7. Larger models typically mean higher inference costs, and the company needs to rely on inference engineering optimization, cluster operations, and advance procurement of computing power to maintain gross margins.
Semiconductor Research Heat Intensifies, Advanced Packaging Benefits First ==================
Institutional funds' attention to China's AI is also beginning to focus on the hardware side.
According to UBS's quantitative team statistics, the three sectors with the fastest increase in institutional research share in July were semiconductors and semiconductor equipment, technology hardware and equipment, and banking; the research share of capital goods, energy, and food and beverage sectors has decreased significantly.
Among them, the research heat of the software and services sector has increased, but the crowding index remains in negative territory; the research heat of materials, food and beverage, and capital goods has decreased, yet positions are relatively crowded. This indicates that semiconductors have become the most obvious direction for the heating up of AI trading, while the software sector may still have some expectation differences.
Specifically in the industry chain, Lanqi Technology stated that DDR5 third and fourth generation memory interface chips are rapidly ramping up production, with fifth-generation products expected to start shipping in the second half of 2026. MRDIMM-related products have already become an important source of new business revenue, and PCIe 6.0 Retimer validation is also underway.
The prosperity of the advanced packaging segment is even more direct. Changdian Technology's capacity utilization rate in the second quarter was about 80%, with AI-related businesses, wafer-level packaging, and bump businesses all operating at full capacity. The company plans to invest 7.8 billion yuan to build a new factory focusing on 2.5D packaging, expected to be completed in the second half of 2028 and put into production in 2029.
UBS believes that this round of packaging cycle is different from previous consumer electronics-driven cycles, as the structural demand brought by AI may support continued improvements in utilization and gross margins. However, the new factory is still a long way from production, and the final returns depend on the volume of domestic AI chips, customer onboarding, and the pace of industry expansion.
AI Demand Begins to Spill Over into Power and Energy Storage ===============
The report also extends the observation scope of AI infrastructure to energy storage and power equipment.
Sungrow Power Supply stated that since the first quarter of 2026, the company has been selling AI data center energy storage systems to developers and integrators, who then provide the products to U.S. cloud service providers. Since energy storage is typically equipment purchased in the later stages of data center construction, the order confirmation speed is relatively slow, and the likelihood of cloud service providers directly becoming their customers is also low.
The company is also collaborating with U.S. cloud service providers to promote the research and validation of solid-state transformers, with products expected to be launched in the coming months. Solid-state transformers and energy storage systems use modular designs, which can shorten on-site deployment times and reduce installation costs.
However, AI demand cannot completely offset the cyclical pressures of the energy storage business. Sungrow Power Supply expects that rising lithium prices may continue to suppress gross margins in the second quarter of 2026, and the company's expansion into low-margin markets will also lower overall profitability. Therefore, UBS believes that AI data center energy storage and solid-state transformers provide long-term growth space, but short-term performance still depends on order rhythms, raw material prices, and overseas supply chain arrangements.
Overall, the signals released by this report indicate that China's AI industry has not yet escaped all constraints, but the growth path is becoming more complete: model call volumes drive computing power demand, domestic computing power adoption promotes server and interconnection upgrades, and chip volume further transmits to packaging, equipment, and power infrastructure.
The next stage that truly needs verification is no longer just whether domestic models can continue to refresh capabilities, but whether this gradually forming AI technology stack can convert demand into stable revenue, controllable costs, and sustained profits.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.