Summary of Key Points
On August 13th, the AI computing power sector saw a significant surge, with concepts such as computing power leasing and exchange machines leading the gains. Ruijie Networks hit a record high, and Gongjin Shares experienced a limit-up. The direct cause was two catalytic events: Alibaba Cloud launched the first domestic super-node computing power service capable of running large models with over 2 trillion parameters, and the official release of DeepSeek Model V4Pro (which significantly enhanced its Agent capabilities). Super nodes are a key technology to overcome the limitations of single-chip computing power and high latency, and they are now entering a phase of scaled deployment. This will drive demand for related industries such as switches, chips, and liquid cooling solutions. The iteration of domestic large models and the implementation of super nodes have created a synergistic effect, attracting investment to these areas. The approaches to developing super nodes differ between domestic and international markets, with multiple domestic industry segments expected to benefit from this development.
Detailed Analysis
1. The "Trigger" for the Sector's Surge: Two Critical News Events Boost Market Confidence
The surge in the computing power sector over the past couple of days was directly driven by two major developments:
- Alibaba Cloud's Super Node Launch: It's like opening a "supercomputing power supermarket" where companies can rent high-speed computing power consisting of 64 connected cards without having to build their own data centers, enabling them to run large models with tens of trillion parameters (such as KimiK3 and Qwen3.8Max). Previously, companies had to purchase multiple servers and connect them themselves; now, they can simply use these services in the cloud, which is more convenient and cost-effective.
- DeepSeek Model Upgrade: This AI model has moved from a "test version" to a fully functional one, with significantly enhanced Agent capabilities (e.g., automatically completing tasks and writing code). The TerminalBench test score increased from 12.8 to 62.7, approaching international top levels. The more powerful the large models, the greater the demand for computing power.
These news events demonstrated that domestic computing power solutions are not only feasible but also practical, and with the rapid advancement of large models, investment has flowed into the computing power sector.
2. What Exactly Are Super Nodes? They Solve the "Bottleneck" Issues in AI Computing Power
Super nodes can be described as "power packaging solutions":
- Previous Challenges: Single AI chips (e.g., GPUs/NPUs) have limited computing power, and running large models required connecting multiple chips, which led to high latency due to slow data transfer, affecting model performance.
- Super Nodes' Solution: Multiple chips ( dozens or even hundreds) are connected using high-speed interconnection technologies to form a "supercomputing unit"—similar to combining several motorcycles into a high-speed train. For example, Alibaba Cloud's super nodes support 64 cards, while Huawei's can support 384 cards, significantly improving efficiency.
Super nodes have evolved from laboratory prototypes to commercial products. Companies like Lenovo, Inspur, and Huawei have launched related solutions, and China Mobile has purchased over 6,000 AI acceleration cards. The industry's focus has shifted from whether super nodes are possible to whether they can be delivered stably and at reduced costs.
3. Which Industry Segments Can Profit from the Implementation of Super Nodes?
Super nodes represent a comprehensive "superproject" that offers opportunities for various segments:
- Complete System Manufacturers: Companies that provide complete super node solutions, such as Huaqin Technology (which shipped super nodes in the second quarter with annual revenue expected to exceed 10 billion yuan and higher gross margins than ordinary AI servers) and Unisplend (whose super nodes support 128 cards and also sell 800G switches).
- Core Components:
- Switches/Exchange Chips: Super nodes require high-bandwidth, low-latency switches. Ruijie Networks is delivering 400G switches in bulk, while Shengke Communications is mass-producing 2.4T chips and promoting 12.8T chips. Analysts suggest this segment has the greatest potential for growth due to increased demand as super nodes upgrade internal server networks.
- Liquid Cooling/Power Supplies: Super nodes consume a lot of power (a single cabinet may require several times the energy of a regular one), so liquid cooling and high-power supplies are essential. Companies like Oulutong supply these components to leading manufacturers like Inspur and Lenovo.
- ODM Manufacturers: While traditional ODMs earned low profits from individual servers, super nodes involve entire cabinets, making them more complex and valuable. Manufacturers verified by top cloud providers (like Huaqin) can see increased profits.
- Connectors: More high-speed connections are needed between chips and boards in super nodes, driving demand for high-speed copper cables and optical interconnection products.
4. Differences Between Domestic and International Approaches to Super Nodes: Where Are the Opportunities for Domestic Companies?
International companies (e.g., NVIDIA) have a "full closed-loop" approach, developing everything from chips (Blackwell) to interconnection technologies (NVLink) to software (CUDA), making it difficult for others to enter the market. In China, the approach is a combination of an independent ecosystem and system integration:
- Huawei/Alibaba Cloud: Develop their own chips (Ascend/NPUs), interconnection architectures, and software to create end-to-end solutions (e.g., Huawei's 384-card super nodes).
- Lenovo/Inspur: Focus on adapting to various domestic GPUs and providing complete system deliveries and industry implementations.
This difference creates more opportunities for domestic industries, such as domestic computing power chips, switch chips, and connectors, allowing them to avoid being monopolized by international players and gain their own market space.
5. Future Trends: Super Nodes and Large Models Create a "Positive Cycle"
Super nodes and domestic large models reinforce each other:
- As large models evolve (with increasing parameters and capabilities), they require more computing power from super nodes.
- As super nodes mature (with reduced costs and stable deliveries), the operating costs of large models decrease, expanding their application scenarios.
The industry will move in two directions: larger-scale deployment of super nodes (from 64 cards to 384 cards or more) and lower per-unit computing power costs, making AI more accessible to more companies. Investment will continue to flow towards these segments with proven commercial potential.
Conclusion
The surge in the AI computing power sector is not incidental; it signals the transition of domestic computing power from technological breakthroughs to large-scale commercialization. As a key technology, super nodes are driving demand across the entire industry chain. Investors can focus on companies with actual product deliveries and close collaborations with leading manufacturers, but they should also be aware of sector volatility, as hot sectors can experience rapid rises and falls. In the long term, AI computing power is the "infrastructure" of the AI industry, and demand will only continue to grow.