Summary of Key Points
At the 2026 World Artificial Intelligence Conference (WAIC), edge AI (artificial intelligence running on hardware such as smartphones, glasses, and robots that are close to users) became a consensus in the industry. In the past, large models mainly competed based on the size of their parameters on cloud-based server clusters. Now, the focus has shifted towards implementing them at the edge—making AI truly accessible to everyone by enabling it to function on devices closest to users. This transition reflects a deep collaboration between model manufacturers and chip manufacturers, who are working together to overcome challenges such as power consumption, heat dissipation, and computational limitations at the edge. The competition in edge AI has also evolved from simply ensuring that the technology can be implemented to actually solving real problems. The approach emphasizes developing general-purpose intelligence first before specializing for specific tasks, with both the edge and cloud working together (the edge handling privacy and real-time responses, while the cloud handles complex calculations and external information). 2026 is considered a pivotal year for the large-scale adoption of edge AI.
I. Edge AI Is on the Rise: Large Models Moving from Cloud Competition to Edge Implementation
Previously, discussions about large models centered around who had the most parameters or the strongest server clusters (for example, models with hundreds of billions of parameters). However, these models were all located in the cloud, and users had to connect to the internet to use them. Things have changed now—devices such as Nubia's AI-powered smartphones, Quark's AI glasses, and embodied robots at WAIC all incorporate large models directly into their hardware.
Why is this the case? Because edge devices are much closer to users. For instance, AI glasses can provide real-time translations and object recognition without waiting for cloud responses, and smartphones can process chat records locally, enhancing privacy. As the founder of BleeqUp AI glasses stated, edge devices use smaller, custom-designed models for everyday tasks, while cloud-based large models handle more complex scenarios (such as accessing specialized information). A report from the Beijing Zhiyuan Research Institute describes this as the second leap in the development of large models: from single-point intelligence in the cloud to distributed intelligence with collaboration between the edge and the cloud.
II. Edge AI Is Not a “Reduced Version of Cloud AI”: The Challenge Lies in Maximizing Intelligence with Limited Resources
Edge devices (such as smartphones and robots) have physical constraints—they cannot consume too much power (they need good battery life) and must not overheat (due to limited heat dissipation), and their computational power and memory are also inferior to those in the cloud. Therefore, edge AI requires a re-design of both software and hardware. For example, robots that need to make autonomous decisions must process data from cameras and sensors in real-time; using cloud models would result in high latency and increased energy consumption. Aixinyuanzhi’s “Embodied Brain Controller” has a computing power of 1500 TOPS (capable of quickly processing multi-modal data), but the key is its high efficiency—it can perform more intelligent tasks with the same amount of power (measured in “tokens per watt”). Close cooperation between chip and model manufacturers is essential; chip manufacturers need to adjust their computing units, and model manufacturers must optimize their models to ensure that AI functions effectively at the edge, not just simply being able to run.
III. The Competition in Edge AI: From “Deployment” to “Solving Real Problems”
The focus of edge AI competition has shifted from whether a model can be installed on a device to whether it can actually solve practical problems. Many demonstrations may look impressive (such as robots folding clothes), but they often represent specialized intelligence—that is, the model is trained with thousands of examples for one specific task and fails to perform well on other tasks, similar to early facial recognition models. FaceWall Intelligence proposes a “general-to-specific” approach: first developing a model with broad general knowledge (like a college student learning basic facts), and then optimizing it for specific scenarios. For example, a robot could be taught common sense such as “a cup is used to hold water” or “the floor should not be stepped on,” before being trained for household tasks like serving tea or sorting in a workshop. Such a model that can adapt to multiple contexts represents true general intelligence and has greater potential for the future.
IV. Edge-Cloud Collaboration Is the Best Solution: The Edge Handles Internal Tasks, the Cloud Handles External Ones
A purely edge-based or purely cloud-based approach has its drawbacks—the edge models have limited capabilities, and a fully cloud-based system poses privacy risks (data needs to be transmitted). The optimal solution is a combination of both. For instance, your AI smartphone could handle daily conversations locally (knowing your coffee preferences) and recognize photos without sending them to the cloud, while the cloud could provide information about the latest coffee shop promotions and calculate the best route. FaceWall Intelligence’s CEO describes this as “the edge handling internal tasks and the cloud handling external ones”—each plays a distinct role without replacing the other, together forming the foundation of intelligent devices. Qualcomm also believes that as hardware computing power increases, more large models will move to the edge, but the cloud will still remain an important complement.
V. 2026: The Year of Turning Point for Edge AI
The demonstrations at WAIC in 2026 indicate that this year marks not just the beginning of edge AI on devices but also three significant transitions:
1. From specialized to general intelligence: The focus is shifting from models designed for single tasks to those with generalized capabilities.
2. From cloud dominance to edge-cloud collaboration: The combination of both approaches will become the norm.
3. From technical showcase to practical application: AI is truly entering everyday devices (smartphones, glasses, robots) to solve real problems.
Those who can find a balance among power consumption, computational power, data management, and use cases will have an advantage in this long-term competition.
This news highlights how AI is evolving from a “faraway, advanced technology” to a practical tool that is already part of our lives. In the future, every intelligent device we use could be a small AI assistant.