Summary of Key Points
This news article focuses on the latest developments in domestic computing power at the 2026 World Artificial Intelligence Conference (WAIC). The main takeaway is that the explosive growth of computing power in the AI era does not rely on a single chip, but rather on a “full-stack strategy for computing power” (a comprehensive solution ranging from chips to software to clusters). Domestic computing power is transitioning from a focus on individual chip performance to a competition based on system-level efficiency. This shift is evident in the accelerated open-source evolution of AI software stacks (such as Pingtouge’s SAIL) and the widespread introduction of super-node products, indicating that domestic computing infrastructure is moving towards greater openness, coordination, and efficiency.
Detailed Analysis
1. Intelligent Agents Bring New Challenges to Computing Power
What are intelligent agents? Simply put, they are AI systems capable of “autonomous thinking and multi-round interaction” (for example, robots that can engage in continuous conversations for hundreds of sentences or industrial AI that can collaborate on complex tasks). They have two distinctive characteristics:
- Long Context Memory: They need to retain all previous conversations to understand user requests.
- Multi-Agent Collaboration: Multiple AI systems work together (e.g., one analyzing data and another generating reports).
This leads to three issues:
- Idle Chips: Chips are underutilized while waiting for data.
- Waste of Computing Power: Chips do not operate at their full capacity.
- Increased Energy Consumption: Frequent task switching results in excessive energy consumption.
Therefore, a single chip’s brute force is insufficient; a systematic approach is necessary to address these challenges.
2. The AI Software Stack: The “Intelligent Butler” That Maximizes Chip Performance
Think of a chip as an automobile engine, and the software stack as the “transmission, driving system, and navigation.” Without it, even the strongest engine cannot perform efficiently.
Pingtouge’s open-source SAIL software stack serves as the underlying software for its Zhenwu chips, with three primary functions:
- Unleashing Chip Performance: It enables chips to operate at 100% of their capacity (for example, increasing from 60% to 90%).
- Connecting with Higher-Level Applications: It accurately communicates the requirements of AI models (such as large-scale model training) to the chips.
- Compatibility with Popular ecosystems: It supports over 260 commonly used AI frameworks, eliminating the need for re-development.
Open-source development allows global developers to contribute to optimizing this software stack, making it more efficient.
3. Super Nodes: Combining Chips into “Supercomputers”
Super nodes integrate multiple chips and enable them to work together through high-speed connections. Examples include:
- Pingtouge Super Node: A single cabinet contains 128 chips, with data transfer speeds in the “nanosecond range” (faster than the blink of an eye).
- Huawei Super Node: It can be expanded to accommodate up to 500,000 chips, providing sufficient computing power for training large models.
- Alibaba Cloud: Super nodes are offered as cloud services, making them accessible to small companies that need to train their own AI models.
These super nodes solve the problem of large models requiring massive amounts of computing power and enable domestic computing power to support more complex AI applications.
4. The “Second Half” of Domestic Computing Power Competitiveness: Focus on Overall Coordination, Not Individual Components
In the past, domestic computing power competition focused on the peak performance of individual chips (e.g., how many calculations a chip could perform per second). Now, the focus has shifted to full-stack efficiency:
- Software Ecosystem: The ability to support a variety of AI applications.
- Cluster Engineering: The capability to efficiently combine multiple chips.
Guosheng Securities has labeled this year as the “year of mass production of super nodes,” indicating a change in the competitive landscape. The key barrier is no longer the performance of individual chips, but the coordination of the entire system. This marks a shift from localized breakthroughs to systemic advantages, bringing us one step closer to an AI era of widespread innovation.
Conclusion
Domestic computing power is evolving from a focus on individual chips to a emphasis on system efficiency. The software stack and super nodes play a crucial role in this transition. This not only addresses the challenges posed by intelligent agents but also makes domestic AI infrastructure more open and efficient, laying the foundation for a true AI revolution. For users, this means faster responses and lower costs when utilizing AI technologies. Even small companies will have access to advanced computing resources.