The Zhitong Finance App learned that CICC released a research report saying that against the backdrop of large model parameter sizes, concurrent reasoning, video memory usage, and continuous increase in power consumption density, AI servers have switched from single-card performance to a cabinet form. Looking ahead to 2027, as conditions in terms of computing power card supply/inference demands/model parameters/solution technology gradually mature, supernodes will have a quantitative foundation. With regard to the evolution of cabinet shapes, the bank believes that it will achieve a shift in value in various aspects such as chip/interconnection/ODM.
CICC's main views are as follows:
What is the value of a supernode?
As the development of AI enters the era of inference, 2027 will enter the first year of supernode deployment. 1) On the demand side: The iteration of large domestic models has accelerated, and the open source model represented by K3 has ushered in qualitative changes; ARR improvement/token growth driven by increased inference demand, and CSP capital expenditure supports continuous upgrading. 2) On the architecture side, supernodes became an inevitable choice after the MoE architecture model parameters were large (parameter storage required). On the other hand, there is a nonlinear relationship between model performance and cluster size, and supernodes are also optimized in terms of interconnection and storage. 3) On the supply side, supernodes use cluster computing power to make up for a single card board.
How big is the supernode market, and which aspects will benefit?
The bank's comprehensive top-down (capex perspective) and bottom-up (supply bottleneck perspective) estimates showed that the market size of domestic supernode cabinets in 2026/2027 was about 759/22.56 million yuan, and the corresponding shipment volume of 64 card cabinets was about 0.69/23,800 cabinets. Judging from the changing ideas of architecture design, the value of connectivity and ODM has clearly jumped; in addition, bottlenecks in actual production need to be considered, including computing power chips/HBM/switching chips/PCB/DrMOS, etc.
What are the types of supernode solutions?
The choice of the supernode solution takes into account cost and actual deployment efficiency. Currently, the overall domestic supernode route shows the characteristics of “multiple parallel specifications, adaptation of domestic chips, advanced reasoning, and follow-up training”. Unlike Nvidia's closed-loop promotion through the NVLink ecosystem, in addition to Shengteng, domestic solutions place more emphasis on open agreements, heterogeneous chip compatibility, full cabinet delivery, TCO optimization, and binding to its own cloud business and model system.
Profit forecasting and valuation
Keep the original profit forecast, target price, and valuation of the covered company unchanged, and are optimistic that supernodes will bring ODM, PCB/connector interconnection, etc. Attention: Huaqin Technology, Lixun Precision, Shennan Circuit, etc.
risk
CSP capital expenditure fell short of expectations, domestic AI chip mass production and certification progress fell short of expectations, new technology yield in the industrial chain fell short of expectations, and competition in some areas suppressed gross profit margins.