瑞银中国AI产业链调研：光互连和SiC升温，胜宏押注Rubin PCB
TL;DR
- After UBS surveyed five Chinese tech hardware companies, AI infrastructure demand remains a common theme.
- Victory Giant Technology plans fixed asset investment of no more than 18 billion yuan in 2026, with the Rubin platform being the focal point of PCB growth.
- Several key figures come from management discussions; actual realization depends on mass production, certification, and external constraints.
In research materials released on June 26, UBS provided more specific details on the diffusion of demand within China's AI hardware chain: AI infrastructure demand remains strong, with its pull extending beyond just GPUs to include PCBs, optical interconnects, liquid cooling, SiC substrates, and optical communication materials.
The news value of this material lies not in the order changes of a single chip company, but rather in several segments simultaneously providing more optimistic timelines and capacity targets. Biren Technology has scheduled its GPU roadmap out to 2028. Victory Giant Technology plans fixed asset investment of no more than 18 billion yuan in 2026. SICC management states that SiC orders are saturated and has selectively raised prices for urgent orders from some small clients.
This content primarily comes from management communications and company guidance within UBS's research and does not constitute formal earnings forecasts. They serve more as a temperature gauge for the AI hardware chain, reflecting industry heat but cannot be directly equated to realized revenue and profit.
Domestic GPU Roadmap Extends to 2028, Victory Giant Bets on Rubin PCB
The timeline provided by Biren Technology is one of the most direct signals regarding domestic AI computing power in the research material.
The company stated to Caijing that the BR20X is expected to be commercially available in 2026, while the BR30X/BR31X are expected to launch commercially in 2028. Management communications within the UBS material further mentioned that the BR20X series GPUs are planned for release in the second half of 2026, featuring upgrades in computing power, memory capacity, and interconnect bandwidth compared to the previous generation.
The catch-up for domestic GPUs doesn't just depend on single-chip computing power, but also on whether memory, interconnects, packaging, and the software ecosystem can keep pace simultaneously. For AI training and inference clusters, bottlenecks often occur between chips, between servers, and at the system scheduling level, not just in theoretical computing power.
The limitations are also clear. Advanced process nodes, advanced packaging, customer mass production timelines, and software ecosystem adaptation will all affect the subsequent deployment of BR20X and BR30X. The external export control environment will continue to impact the pace of the local AI hardware supply chain.
Closer to near-term orders and capital expenditure is Victory Giant Technology's PCB guidance.
Management states that the raw material costs for AI PCBs are generally stable, and the cost pass-through for new-generation products is relatively smooth. They expect the gross margin for the fiscal year 2026 to be roughly flat compared to 2025. The proportion of AI revenue is expected to rise from below 50% to 60%-70%.
The Rubin platform represents the most anticipated growth driver. According to the research material, Victory Giant maintains a majority share in Nvidia's Rubin compute tray HDI, consistent with the situation for the GB300 platform, with an overall customer share target of approximately 50%. As this data point hasn't been verified through public announcements, it's better understood as a management target and sell-side research information. If realized later, Nvidia's new platform iterations will continue to drive growth in China's high-end PCB supply chain.
The expansion figures also require more precise interpretation. Public information shows Victory Giant Technology's total investment in 2026 will not exceed 20 billion yuan, with fixed asset investment not exceeding 18 billion yuan, primarily directed towards Huizhou Factory 10-13. This scale indicates that AI server PCBs have been designated by management as a core production capacity direction for the next few years.
The material solution for the orthogonal backplane required by the Rubin Ultra has not been fully finalized. The company states there are no delays in related adoption, but it is still evaluating multiple material and supplier combinations, including Q-glass and PTFE. While the order direction is relatively clear, the process and material routes are still being screened.
Optical Interconnects and Liquid Cooling Emerge as Supporting Demands for AI Clusters
As AI clusters continue to scale up, the importance of interconnects and cooling increases. Lightelligence and Lens Technology provide two pieces of supporting evidence.
The UBS research material states that Lightelligence holds approximately an 88% market share in independent scale-up optical interconnect solutions, focusing on LPO, NPO, and other solutions serving large-scale GPU and ASIC clusters. Its optical computing processor utilizes electronic ICs and silicon photonics stacked vertically via 3D TSV, offloading some computing tasks to the silicon photonic layer, thereby reducing latency and decreasing reliance on advanced process nodes.
The key point of such solutions isn't that "optical computing replaces GPUs," but rather that the larger the AI cluster, the more prominent the interconnect bottleneck becomes. If LPO, NPO, and silicon photonics packaging can achieve stable mass production, they may provide incremental performance gains in bandwidth, power consumption, and latency for large-scale clusters.
Lens Technology's AI infrastructure layout stems from optical communication and liquid cooling. Public reports indicate that Lens Optical Electronics strategically acquired a controlling stake in Shenzhen Tongsheng Optoelectronics Co., Ltd. in June 2026, entering the field of hollow-core fiber optic cables. Management targets within the UBS research show that the optical communication business is expected to make a more meaningful contribution to the income statement starting in 2027, with a revenue potential of approximately 10 billion RMB.
In the liquid cooling sector, management's goal is to become the largest shareholder of Yuans Tech, with revenue for the liquid cooling business in fiscal year 2026 potentially reaching billions of yuan. The per-unit value for cold plates, manifolds, and UOD is approximately $40,000 to $50,000. These figures lack public announcement verification and should be understood within the context of targets and guidance.
Lens Technology's stock price and target price historical trend chart alongside historical rating and target price adjustment records.
SiC Orders Saturated, 8-Inch Shipment Target Raised to 50%
In the power device chain, the signals from SICC (Shandong Tianyue Advanced Materials Co., Ltd.) are closer to "supply-demand tightness has already started to impact some prices."
A company press release previously cited data from Fuji Keizai, stating that in 2025, SICC held the global top market share for SiC substrates, including 6-inch and 8-inch products, with an 8-inch market share of 51.3%. Management communications within the UBS research material further state that current demand is strong and orders are saturated. Prices in the first quarter of 2026 were generally stable, but selective price increases were implemented for urgent orders from small clients.
Management's goal is to have 50% of shipments come from higher-margin 8-inch SiC substrates in 2026 to drive gross margin expansion. If this target is achieved, it will directly impact the company's product mix and profitability.
Historically, SiC demand has been primarily driven by electric vehicles, photovoltaics, and energy storage. Now, power systems for high-power AI servers and data centers are also becoming a new source of demand. HVDC, SST, AI glasses, and advanced packaging are also listed as potential drivers.
The key point here is not that SiC demand has suddenly shifted from automotive to AI, but rather that AI data centers are expanding the demand boundary for power semiconductors. Saturated orders and selective price increases indicate that, in certain customer segments and urgent order scenarios, supply tightness is beginning to affect prices.
Whether prices can continue to rise depends on the ramp-up of 8-inch capacity, customer qualification, and the pace of downstream demand. If capacity release outpaces order growth, the potential for price increases may be weakened. If data center demand materializes faster than expected, the scope for SICC's gross margin improvement will become clearer.
Order Direction Clear, Challenges Remain in Mass Production and Constraints
This set of research information points towards one common change: AI hardware demand is diffusing outwards along the GPU, PCB, optical interconnect, liquid cooling, and SiC supply chains. Chinese suppliers are presenting more aggressive capacity and product roadmaps across multiple segments.
However, it is not appropriate to portray this as a "comprehensive breakthrough in domestic AI hardware." Biren Technology's GPU roadmap remains constrained by process nodes and ecosystem limitations. Victory Giant's share related to the Rubin platform still depends on platform mass production and finalization of material solutions. Lightelligence's optical interconnects and optical computing require large-scale customer validation. Lens Technology's contributions from optical communication and liquid cooling will primarily be seen in 2026-2027. Whether SICC's selective price increases can be sustained depends on 8-inch capacity and data center demand realization.
The larger external constraints remain in areas related to advanced process nodes, advanced packaging, high-end GPUs, and capital expenditure by cloud providers. The closer a segment in the AI hardware chain is to these positions, the more susceptible it is to policy and supply chain constraints.
The most valuable aspect of these figures is they indicate that orders for AI infrastructure are spreading to more hardware segments. The boundaries that need to be maintained are equally clear: most data points are still management targets and research information, and actual realization depends on customer mass production, finalization of materials, and the external environment.
