WAIC Observation: The larger the model, the greater the computational demands; domestic AI companies are collectively struggling with “utilization rates.”

Summary of Key Points

At the 2026 World Artificial Intelligence Conference (WAIC), domestic AI models made breakthroughs in terms of large parameter sizes and performance (such as the 2.8 trillion parameters of the domestically developed model Kimi K3). However, the resulting surge in computational power requirements has become a bottleneck for the industry. Leading companies (like Yuezhiyanmen and Zhipu) have increased their investments in computational infrastructure, shifting the competition from a simple "parameter size race" to a more complex realm that focuses on "computational infrastructure, engineering efficiency, and ecosystem collaboration." The construction of super-node clusters has become a focal point, with improvements in computational power utilization efficiency and software/hardware compatibility becoming new core competitive areas.

I. The Parameter Race for Large Models Has Intensified, but Computational Power Has Become a Bottleneck

In simple terms, model parameters are akin to the number of neurons in the brain; more parameters mean stronger capabilities in handling complex tasks, such as understanding long texts and performing multi-task reasoning. At WAIC, Kimi K3 from Yuezhiyanmen set a new record with 2.8 trillion parameters, attracting global attention. However, larger parameter sizes also lead to a corresponding increase in computational power demands:

• On the day K3 was launched, users reported task failures due to excessive demand and insufficient computational resources;
• Yuezhiyanmen had to temporarily suspend new user subscriptions and allocate all available power to existing users while urgently expanding its infrastructure;
• Zhipu also acknowledged that to achieve performance on par with K3, they need larger parameter sizes, but this is delayed due to limited computational power.

This indicates that the issue is not that domestic models are unwilling to grow in scale; rather, they face limitations in their ability to process data effectively due to a lack of sufficient computational power.

II. How Leading Companies Are Competing for Computational Power and Improving Efficiency

Faced with this shortage, top companies are adopting various strategies:

1. Yuezhiyanmen: Investing heavily in GPUs for aggressive expansion

Founder Yang Zhilin stated that over half of the company's funding is being used to purchase high-end GPUs, aiming to match OpenAI's computational resources. K3 requires tens of thousands of high-end GPUs for continuous training, and currently, the demand exceeds the cluster's capacity, so the company is prioritizing maintaining user experience.

2. Zhipu: Optimizing engineering processes and integrating the computational ecosystem

• Zhipu's annual recurring revenue (ARR) has increased by 15 times to $1 billion, primarily due to improved inference efficiency—more tasks can be processed with the same amount of computational power;
• In the first half of this year, Zhipu acquired AI infrastructure company Zhongke Jiahe, which complements its software capabilities, such as compilers and inference engines, reducing the need for customizations between different domestic chips (e.g., Loongson and Cambricon);
• Zhipu is also developing an intermediate layer ecosystem similar to NVIDIA's CUDA to make domestic computational resources more compatible with a wider range of models.

In summary, Yuezhiyanmen is focusing on hardware acquisition, while Zhipu is optimizing software and integrating ecosystems to overcome computational power constraints.

III. Super-Node Clusters: A New Solution to the Computational Power Challenge

Previously, AI computations relied on individual chips, but models with tens of billions of parameters require thousands of chips working together. This has led to the development of super-node clusters, which connect multiple chips to form a supercomputing unit that can perform efficiently without the need for switches, resulting in lower latency and energy consumption:

• Qingwei Intelligence's 4096-chip super-node: Capable of supporting over 200 large models out of the box for direct use in complex AI tasks;
• Industry consensus: It is difficult for a single company to build a cluster with thousands of GPUs; collaboration among chip manufacturers, software providers, and network vendors is essential. Domestic computational power solutions are moving from a fragmented approach to ecosystem-based collaboration.

Why are super-node clusters important? Models with billions of parameters cannot be trained or inferred efficiently with a single chip; clusters are necessary for optimal performance. Companies are developing various super-node technologies, with differences in the specific implementation details.

IV. The Core Conflict in the Computational Power Industry Has Shifted

The focus has shifted from whether there is enough computational power to how well it is utilized:

• Inference power becomes the mainstream: In 2026, the domestic computational power market saw a shift where inference (model execution) power surpassed training (model learning) as the primary growth driver;
• Utilization rate is more critical than peak capacity: The focus is now on actual usable power and cost-effectiveness rather than theoretical capabilities;
• The era of brute-force hardware accumulation is over: Vertical industries require customized computational solutions, and demand for domestically produced computing resources is increasing. Simply buying more GPUs is no longer sufficient.

For example, the separation of CPUs and AI acceleration cards leads to slow data transfer and high energy consumption, which is a bottleneck for domestic AI development. Super-node clusters aim to address these issues by consolidating resources, improving scheduling, and reducing inefficiencies.

V. Future Competition: From "Hardware Parameters" to "Full Stack Ecosystems"

The trends at WAIC indicate that the competition in the AI industry will no longer revolve around who has the largest model parameters or the most powerful chips, but rather on:

1. Optimized software/hardware integration: Chip design focuses on high-speed interconnectivity, large memory, and high bandwidth;

2. Comprehensive ecosystem services: Providing integrated solutions from chips to cluster management;

3. System-level efficiency: Improvements in resource scheduling and inference engine performance to maximize computational power utilization.

In summary, companies that can integrate hardware, software, and ecosystems will have a competitive advantage in the AI industry.

Conclusion

Domestic AI development is transitioning from a focus on model parameters to an emphasis on computational infrastructure and efficiency. The breakthroughs in large models are inseparable from adequate computational resources, and the competition in this area has shifted from quantity to quality and collaboration. For end-users, this means more stable and intelligent AI services, driven by ongoing investments and strategic efforts within the entire industry in software and hardware ecosystems.