2026 World Internet Conference | Frost & Sullivan and LeadLeo Deeply Involved in Wuzhen Summit Agenda, Release of Blue Book and Group Standards

Xu Guangcan, Executive Director of Frost & Sullivan China

As the capabilities of large model systems continue to improve, artificial intelligence is shifting from primarily cloud-based training and inference to extending to terminals such as smartphones, computers, cars, robots, wearable devices, and smart homes. Endside AI can perform some or all of data processing and model inference locally, offering significant benefits in reducing response latency, minimizing cloud costs, protecting data privacy, and enhancing personalized experiences. At the same time, endside AI is not a single product or technology market but a composite industrial system composed of edge chips, storage, sensors, endside models, operating systems, development tools, terminal devices, and application services — currently, the industry has not yet reached a unified understanding of its market boundaries, industrial chain structure, and competitive landscape.

In light of this, the Blue Book, centered on the questions of "Is the technology innovative? Can the scenario be scaled? Can enterprises establish sustained competitiveness?", focuses on five aspects: definition and overview of the endside AI market, endside AI industrial chain and value distribution, key terminal markets and application scenarios, competitive landscape and industrial ecosystem, and benchmark cases and development trends, aiming to provide references for enterprises’ strategic layout, industrial cooperation, and investment decisions.

It is worth noting that endside AI and the core topic of this seminar, FlagOS heterogeneous computing power, are highly interconnected in their underlying logic: both face the same industrial pain points, such as "fragmented architecture and high adaptation costs" — the cloud requires a unified software stack for "one-time development, multi-chip operation", while the edge also needs to shield chip differences and reduce the transfer costs of models and applications. The unified technical foundation for heterogeneous computing power can be extended to the edge for reuse, supporting unified development and computing power coordination across cloud, edge, and end; the scaled implementation of endside AI will, in turn, open up new demand spaces for smart computing power operation models such as "Token Factory". The Blue Book systematically outlines this "cloud–edge–end" coordinated industrial mainline within a unified framework.

Gao Junxiao, Vice President of StarFan Intelligent, delivered a keynote speech titled "Enabling All Things, Creating the Future – Opportunities for Edge Chips in the Physical AI Era". His first judgment was that AI is evolving from a single-point application to an "intelligent all things" infrastructure — the application layer integrates into office and business systems, the device layer extends to smartphones, robots, cars, satellites, and unmanned systems, and the infrastructure layer (chips, storage, network, and servers) is redesigned around AI load. According to publicly available data he cited, global AI spending in 2026 reached $2.59 trillion, with an annual growth rate exceeding 70%. This is accompanied by the "hybrid AI" architecture becoming the mainstream: the cloud provides general capabilities, the edge supports industry- and region-specific closed-loop scenarios, and the terminal handles real-time perception and decision-making — the three dynamically allocate computing tasks based on time, latency, operating conditions, and cost.

In his view, the evolution path of AI is clearly visible: from perception AI (image and video recognition), to generative AI, to intelligent agent AI that can plan and execute tool usage, and finally to physical AI — completing the full closed-loop "perception–understanding–planning–action" in the real world, with robots, industrial and commercial applications, and driverless vehicles being typical applications. This is why endside AI has moved from single-point technology competition to system-level optimization: on the model side, it continuously "slims" through low-bit, sparse, knowledge distillation, and large-small model coordination; on the chip side, it uses heterogeneous CPU/NPU/GPU hybrid, dedicated NPU, and 3D stacking to enhance energy efficiency; on the engine side, it performs operator fusion, cross-platform adaptation, and performance analysis; on the system side, it moves toward multimodal fusion of vision and touch and "large-small brain" coordination — and all of this is underpinned by architectural breakthroughs in edge chips.

The answer provided by StarFan Intelligent is "not simply stacking computing power". Constrained by both design and power consumption, the edge cannot follow the path of server chips; it must leverage system-level capabilities to maximize the efficiency of limited computing power: through techniques such as mixed-precision quantization, the model size can be compressed by up to 85%, with precision loss controlled between 1% and 2%, and when combined with native IP and hardware-level mixed-precision support, the equivalent computing power increases by approximately 1.5 times, with power consumption kept below 25W.

At the end of his speech, he shifted his focus back to humans: In the future, humans will be responsible for setting goals, generating ideas, making judgments, and supervising, while AI will be responsible for understanding the scenario, planning, and optimizing, and robots will execute in the physical world and provide feedback, forming a closed-loop collaborative relationship; the stringent requirements of the real world regarding energy efficiency, cost, and reliability will, in turn, continue to define the architecture of edge chips. His vision is warm: "In the future, each person will have a truly own AI partner: responding more promptly, ideas more credible, privacy better protected, and capable of helping you work."