At the AI Investment Summit hosted by Sina Finance in Beijing on September 16, with the theme "Certain Opportunities in the AI Infrastructure Era," GF Fund's Equity Investment Director and General Manager of the International Business Department, Li Yaozhu, joined a panel discussion. He stated that supernodes are an inevitable development trend accompanying the evolution of large models from hundreds of billions to trillions of parameters, with the primary driving force coming from changes in the model architecture itself.
Regarding the impact of supernodes on AI infrastructure and who might emerge as winners in this cycle, Li Yaozhu explained that the core reason for discussing supernodes is the shift in model architecture. Currently, the mixed expert architecture is used, with N experts deployed on different chips. This involves communication between experts, especially during architecture advancement. In China, both overseas and domestic chips are used, each with different HBM and computing capabilities. With the MOE architecture, while single-card computing power is strong, the bottleneck lies in communication and storage. Supernodes address how to systematically solve the efficiency problems of these cards during training or inference, functioning more like a system integration role.
He noted that this process affects not only optical communication but also liquid cooling and power dispatch. For example, from a communication perspective, within a 51.2T total bandwidth architecture, configurations can support either 64 ports at 800G or 128 ports at 400G. The chosen method depends heavily on the MOE architecture. Whether to use more channels for chip-to-chip communication or to use higher-bandwidth chips with fewer channels depends on the final form of model deployment, making communication protocols and software scheduling critical.
Li Yaozhu believes that across the value chain, optical communication, software scheduling, and architecture design, especially in topology design, offer significant value appreciation. In China, despite constraints on hardware and advanced manufacturing processes, leveraging software engineering strengths can maximize value. Beyond overall value enhancement, even minor optimizations in system integration could substantially impact the entire system. Particularly in the agent era, which demands higher scheduling frequencies, subtle system refinements could lead to greater efficiency gains. From his perspective, these represent major future investment opportunities and key focus areas.
While domestic supernode construction advances rapidly, overseas hyperscale cloud providers and new cloud companies are also increasing AI infrastructure capital expenditures. Li Yaozhu pointed out differences in construction models, financing methods, resource scheduling, and return-on-investment requirements. Overseas new cloud clients, primarily two major AI companies focused on centralized delivery, adopt various financing methods including bond and equity financing. This differs from typical CSPs, which weigh internal returns more heavily and support capital expenditure through cash flow along with new financing tools like convertible bonds. Additionally, unlike new cloud providers, CSPs design their own chips alongside purchasing computing power in bulk.
He highlighted that open-source model adoption is rising, with CSPs increasingly using them. Since overseas computing costs are lower, open-source model token costs are also cheaper. This leads to greater computing power deployment during the agent era, benefiting server companies, which contrasts with the concentrated delivery model overseas. Li Yaozhu observed that domestic cards and overseas cards inherently have different requirements, and domestic new cloud providers form their own ecosystems based on overseas card availability, delivering to clients in a centralized manner. However, system architecture integration differs significantly from overseas approaches, and due to restrictions on importing overseas cards, design cannot follow the same cost-unconscious methods. Domestic compatibility is essential.
Given the rapid iteration of domestic cards, future system-level solutions require consideration, such as redundancy margins in power and liquid cooling. Since domestic cards have lower single-card computing power but need greater interconnectivity, systems require advance planning. The ratio of overseas to domestic cards may also change, necessitating long-term thinking about solutions, particularly liquid cooling solutions that accommodate both card types. This offers important lessons for China—the complexity of implementing domestic systems exceeds that of overseas systems, potentially generating investment opportunities, especially in areas like liquid cooling and power system interconnectivity.
Looking ahead three years, Li Yaozhu identified the most certain growth directions in AI infrastructure and warned about potential bubbles. He noted that opportunities are continuous, spanning storage, optics, and liquid cooling. This year, markets have given high valuation premiums to supply chain segments experiencing short-term shortages. Caution is warranted for areas with short-term scarcity but no long-term gap—where short-term means roughly one year and long-term means two to three years. He advised careful evaluation of areas with significant capacity release by 2028 that might face insufficient demand. From his perspective, sectors with mid-to-long-term scarcity have higher upside probabilities, while those with only short-term tightness risk valuation declines as the balance shifts.