Summary of Key Points
Meituan has released the LongCat-2.0 large model, and what's most impressive is not its 1.6 trillion parameters, but the fact that it completed the entire process from pre-training to inference using 50,000 domestically produced computing cards. This marks a significant shift in domestic computing power from a stage where the focus was on whether chips could be manufactured, to a new phase where the ability to deliver stable systems in ultra-large-scale commercial scenarios is being tested. As a result, previously underestimated components in the supply chain, such as connectors, power supplies, and cooling systems, are now being re-evaluated. Companies like Huafeng Technology, Yihua Co., Ltd., Aerospace Electrical Appliance, and Jiehuate have opportunities, but their success will depend on different validation criteria. 2026 will be a critical turning point for domestic computing power, as it moves from focusing on expectations to demonstrating actual financial performance.
I. Meituan's Large Model: A Real-World Test for Domestic Computing Power
Meituan is not just playing with concepts; they have invested heavily in domestic computing power, using 50,000 domestically produced cards for pre-training and a complete commercial inference process involving 30 trillion tokens. This is equivalent to subjecting domestic computing power to an extreme test using the real business of an internet giant that processes tens of millions of orders daily:
- The focus is not on peak chip performance, but on system stability: The cluster of 50,000 cards must work simultaneously without power outages, data transmission delays, or overheating issues—these are all engineering challenges that cannot be tested in laboratories.
- More convincing than any marketing claims: Previously, there were doubts about the usability of domestically produced cards. Now that Meituan has dared to rely on them for their core AI services, it shows that domestic computing power can handle commercial demands. This will encourage more companies (such as other internet giants and operators) to use domestic computing power, providing a foundation for upstream orders.
II. AI Servers: More Than Just Chipping Up
Many people think that the stronger the chips in an AI server, the better, but this is not entirely true. The real challenge is ensuring that the system runs stably over the long term:
- Fast connections: High-speed data transfer between numerous chips and servers is essential, similar to the need for smooth highways, which relies on high-speed connectors (such as Huafeng's backplane connectors and Yihua's external I/O connectors).
- Stable power supply: AI chips consume a lot of power, so power management chips (like those from Jiehuate) must act as intelligent schedulers to ensure efficient operation under various loads without malfunctions or overheating.
- Effective cooling: As computing density increases, cooling becomes more critical. Liquid cooling interconnects (provided by Aerospace Electrical Appliance) are like air conditioners for servers, preventing them from crashing due to overheating.
- Increasing value of system components: The value of these components will rise; for example, as power consumption increases, the value of power management chips also goes up, and as bandwidth increases, so do connector prices. Therefore, upstream companies have the opportunity to not only sell more but also at higher prices.
III. Connectors: The Underestimated “Vessels” of Computing Power
Connectors are like the “veins” of a computing system, transporting data, power, and coolant. They were once seen as simple plugs and sockets, but in the era of large-scale card clusters, they have become crucial for system stability. The three companies mentioned each have different roles:
- Huafeng Technology: Focuses on high-speed backplane connectors inside servers, deeply integrated into system design, making them difficult to replace (similar to motherboard interfaces in smartphones). Key indicators include the proportion of revenue from high-speed connectors and whether their gross margins remain stable (avoiding pressure from major customers).
- Yihua Co., Ltd.: Develops high-speed I/O connectors and copper cables for connections between servers and cabinets. The focus is on whether the proportion of AI-related business can continue to grow (since they also have traditional businesses, growth in AI could change their valuation).
- Aerospace Electrical Appliance: Uses military-grade reliability technology for liquid cooling interconnects and connectors. The key is whether they can secure independent orders for their products, proving that their success is not just a extension of their military business but a standalone growth area for AI.
In the short term, these companies will collaborate, but in the long run, there may be competition. However, each has its own market opportunities.
IV. Power Management Chips: A Slowly Changing Factor in Domestic Replacement
As AI chip power consumption rises, power management chips play a vital role. The high-end market was previously dominated by overseas manufacturers, but with the rise of domestic computing power, local companies (like Jiehuate) now have opportunities:
- Jiehuate’s opportunity: Partnerships with Huawei's Ascend platform to develop high-end products like DrMOS and multi-phase controllers. However, the chip industry requires significant upfront investment and long validation periods, so success cannot be measured solely by technological breakthroughs.
- Key indicators for success: Whether high-end products are included in major customers’ reference designs, whether gross margins improve (due to scale effects), and whether losses narrow (as R&D costs decrease). Only when all three indicators improve simultaneously will there be a real profit turnaround.
V. 2026: From “Talking About Possibilities” to “Looking at Financial Reports”
Domestic computing power has moved beyond the stage of questioning its existence; now, the focus is on whether it is practical and profitable. Investors should focus on three key areas:
1. Actual shipment trends: Are more large companies using domestic computing power, such as operators and government/enterprise customers for intelligent computing clusters?
2. Chip delivery timelines: The launch event is not the ultimate test; attention should be paid to prototype validation, small-scale deliveries, and mass production (order confirmation depends on the final product).
3. Company-specific metrics:
- Huafeng Technology: Proportion of revenue from high-speed connectors and gross margins.
- Yihua Co., Ltd.: Growth and proportion of AI-related business.
- Aerospace Electrical Appliance: Independent orders for liquid cooling products and the breakdown of AI-related revenues.
- Jiehuate: Proportion of high-end products, improvement in gross margins, and reduction in losses.
In summary, the industry trend is about direction, commercial success is reflected in financial reports, and company value depends on their competitive advantages. The story of domestic computing power is finally moving from the laboratory to the profit statements.
Conclusion
The significance of Meituan's LongCat-2.0 lies not just in the model itself but in it being a milestone for domestic computing power to evolve from a laboratory prototype to a commercial tool. In the coming years, we no longer need to debate whether domestic chips are feasible; instead, we should focus on which companies can turn these opportunities into real profits. For investors, identifying those that can solve real problems, secure orders, and generate revenue is crucial. After all, no matter how promising the industry trend, it is ultimately the financial performance that matters.