Three Key Signals of WAIC 2026: Computing Power Restructuring, Agent Delivery, and the Closed Loop of AI Hardware
Seventy years ago, a group of young scholars first proposed the concept of "artificial intelligence" at the Dartmouth Conference.
Seventy years later, by the banks of the Huangpu River, the artificial intelligence industry is no longer satisfied with enabling machines to think and speak, but has begun to embed this capability into enterprise systems, smart terminals and robots, attempting to make AI truly participate in the operation of human society.
From July 17 to 20, 2026 World Artificial Intelligence Conference and the High-Level Meeting on Global AI Governance was held in Shanghai. With the theme of "Smart Partners, Co-Creating the Future", this conference is spread across four venues in three areas: Expo, Zhangjiang and West Bund. Its exhibition area has exceeded 100,000 square meters for the first time, with more than 1,100 enterprises presenting over 3,000 exhibits and over 300 products making their global debut. The two tracks of intelligent computing and embodied intelligence each gather more than 200 enterprises.
This is the largest WAIC in history, and possibly the one with the most complex technical routes and product forms. Large models, supernodes, agent operating systems, AI phones, humanoid robots, dexterous hands, near-memory computing chips and various industry solutions all appear in the same exhibition system. At first glance, they belong to completely different technical tracks, but when put together, they present a clear evolution path:
AI is evolving from a model that can generate answers into a system that can understand goals, mobilize resources, perceive the environment and deliver results.
Over the past three years, large models have proven their ability to write articles, generate images, code, understand videos, and handle increasingly complex reasoning tasks. But in 2026, the key question facing the industry is no longer "can AI do it", but "can AI do it continuously, stably and at low cost". A model successfully completing a task once in a test is two entirely different things from a system handling tens of thousands of business requests every day.
A robot completing a single jump on stage is not the same capability as it performing handling, sorting and loading/unloading tasks in a factory for consecutive months. For artificial intelligence to transition from capability demonstration to real production, it requires a complete set of infrastructure beyond the model: how computing power is organized, how data flows, how tools are invoked, how permissions are controlled, how results are evaluated, and how errors are detected in a timely manner.
When the boom of large models first emerged, what easily attracted the most attention in the industry were parameters, rankings and model releases. Today, more and more companies are focusing their efforts on the systems behind the models and the real tasks ahead of them. Models have not disappeared, nor have they become less important. On the contrary, they are being embedded into larger technical architectures, just like electricity, chips and operating systems.
Through this WAIC, we have witnessed three ongoing transformations in the AI industry.
01
Computing power is no longer measured solely by chips: shifting from parameter competition to system competition
On the exhibit list of this WAIC, one of the most talked-about products is the real Huawei Atlas 950 supernode.
The key to a supernode does not lie in how many chips are packed into a single cabinet, but in whether the cost of data exchange between chips can be reduced. High-speed interconnection, shared memory, task scheduling and software stacks jointly determine whether a large number of independent chips form a disjointed group working on their own, or a coordinated computing whole.
However, as model scales increase and reasoning tasks become more complex, a single chip is increasingly unable to independently determine the final performance. The data transmission speed between chips, the efficient collaboration of hundreds or even thousands of chips, sufficient memory, proper task scheduling and mature software tools all affect the actual utilization rate of computing power.
No matter how high the theoretical peak of a single chip is, if a large amount of time is spent waiting for data, cross-card communication and task switching, it can hardly be converted into effective model training and reasoning capabilities. This is like a factory: having the fastest workers does not necessarily mean it has the highest output. Raw material supply, workstation arrangement, transportation routes and management systems can all become bottlenecks that determine production.
AI computing power is undergoing a similar transformation. The basic unit of competition is gradually shifting from a single chip to servers, supernodes and clusters; the dimensions of competition have also expanded from computing performance to high-speed interconnection, memory, heat dissipation, software ecosystem, energy efficiency and overall operation and maintenance. Especially in the reasoning era, computing power systems are no longer only facing a small number of large-scale training tasks that last for months, but a huge number of requests with varying durations and obvious concurrency fluctuations.
Agents need to repeatedly invoke models and tools, reasoning models need to generate longer thought processes, and multi-modal models also need to process text, audio, images and videos simultaneously. These changes will all increase Token consumption, and put forward higher requirements for response speed, unit cost and system stability.
Public data shows that the scale of China's AI-related industries has exceeded 1 trillion yuan in 2025, and the overall AI penetration rate in key industries has exceeded 80%. China's daily average Token invocation volume has grown from 100 billion at the beginning of 2024 to 100 trillion by the end of 2025, and continues to increase rapidly in 2026.
Compared with model parameters, Token invocation volume is closer to the real development level of the AI industry. Parameters represent the potential capabilities a model may have, while Tokens mean these capabilities are actually being used. Tokens can only be continuously generated when models are embedded into search, programming, office work, customer service, marketing, scientific research and manufacturing processes. When the daily invoked Token volume reaches the level of hundreds of trillions, industry competition will naturally shift from "who has a larger model" to "who can produce intelligence at lower cost and higher efficiency". This is the core background behind the rise of supernodes.
It is not just about packing more chips into the same cabinet, but about shortening the distance between chips as much as possible, allowing a large number of processors to share data, memory and tasks, and improving the effective computing power of the entire system. For domestic AI chip companies, this change not only brings opportunities, but also raises the threshold of competition.
In the past, as long as a company designed a chip that could run mainstream models, it had the opportunity to enter the market. In the future, what customers need may no longer be a single card, but a complete set of solutions ranging from chips, servers, interconnection, software platforms to cluster operation and maintenance. In other words, the competition for domestic computing power is evolving from "whether we have chips" to "whether we can form a complete system".
At this WAIC, Agile Intelligence made its global debut of the Epoch series RISC-V AI computing power supernode solution. This complete system, consisting of the Epoch cloud high-performance AI chip, ELink high-speed interconnection architecture, full-liquid cooling solution, full-stack software ecosystem and large-scale cluster networking, delivers far superior performance than traditional AI servers. According to the company, this solution supports high-bandwidth interconnection of 32 to 128 cards in a single cabinet, and adopts an orthogonal backplane-free architecture to replace traditional cable connections, which improves signal transmission quality, significantly reduces transmission latency, and enhances computing power density and reliability. ELink can support both vertical scaling within a supernode and horizontal networking between nodes, further expanding computing power resources to clusters with tens of thousands of cards.
At present, Agile Intelligence's Epoch series AI chips, acceleration cards and related hardware have entered mass production and batch delivery, with customers covering the internet, telecom operators, finance and energy industries. The company has also joined hands with industry chain partners to build the world's first full-stack RISC-V supernode AI Token factory. This reflects that the competition for domestic AI computing power has entered a new stage from "whether we can manufacture chips" to "whether we can deliver complete infrastructure".
RISC-V provides an alternative path for this competition. Unlike private instruction sets controlled by a single enterprise, RISC-V is an open instruction set. Theoretically, different companies and research institutions can design processors, expand functions and improve the software ecosystem on this basis. However, openness does not automatically equal maturity.
For an AI chip to truly enter the production environment, it also requires compilers, operator libraries, framework adaptation, performance optimization and developer tools. Model architectures are evolving rapidly, and today's mainstream operators and precision formats may soon be replaced by new reasoning methods. Therefore, the value of an open architecture lies not only in reducing authorization restrictions, but also in enabling chips to continuously evolve and access more R&D resources.
But whether it can ultimately form competitiveness still depends on mass production, software adaptation, customer migration costs and long-term operation data. In the field of AI infrastructure, any grand technical narrative must ultimately be tested by the same set of indicators: how many effective Tokens can be generated per kilowatt-hour of electricity, how long can a device operate stably, how much time is required for model migration, and whether it can recover quickly after a failure.
The real threshold of the computing power industry has never been simply powering on chips, but turning chips into productivity that customers can use continuously. This is why intelligent computing platforms, model service platforms, compilers, reasoning engines and computing power scheduling systems are becoming increasingly important. If chips are engines, supernodes and clusters are vehicles, then software platforms determine how easy these vehicles are to drive.
Model companies hope to shield the differences between different chips, developers want to migrate models at the lowest possible cost, and enterprise customers expect computing power systems to be as stable, transparent and easy to use as cloud services. Whoever can lower the threshold of using computing power will have a better chance of winning over developers and customers.
This round of computing power competition is also transforming the cloud computing industry. Traditional cloud services are mainly billed by CPU, storage and bandwidth, while AI cloud increasingly provides services based on model invocations, Token volume and task results. Computing power itself is gradually moved to the background, and what customers ultimately purchase is the speed of model training completion, the latency of reasoning responses, and the processing capability of a certain business task.
In the computing power industry, "Token Factory" is becoming a widely used concept to describe this shift. It no longer focuses solely on how many chips are deployed, but on how many business-usable Tokens a system can stably generate under given power and hardware conditions. Under this logic, the peak performance of chips is not the only indicator. The throughput, energy efficiency, stability and utilization rate of the entire system determine the final output. The 2026 WAIC further demonstrates that while the performance of a single chip remains important, it is no longer sufficient to independently determine the competitiveness of an AI infrastructure.
The future leader in the computing power sector will not necessarily belong to the company with the most powerful chips, but more likely to the enterprise that can integrate chips, interconnection, software, energy and customer demands into a well-functioning system. Model training has shaped the computing power market in the previous stage, and large-scale reasoning will redefine the infrastructure in the next stage.
02
Agents enter enterprises: the real barrier lies in context and delivery
If supernodes solve the problem of how intelligence is produced, then agents address the problem of how intelligence is used. "Agent" is almost one of the most frequently appearing concepts at this WAIC. From agent operating systems and agent-equipped phones to enterprise agents and multi-agent collaboration platforms, more and more companies are no longer satisfied with letting AI answer questions, but hope that AI can understand goals, break down tasks, invoke tools, and complete work with minimal human intervention.
From chatbots to agents, it may seem like just a change in product names, but behind it lie two completely different product logics. Traditional chatbots wait for users to raise clear questions, and then return text, images or code. They can help people complete a certain part of the work, but most of the work still requires humans to connect the dots. People not only need to know how to ask questions, but also need to judge whether the answers are reliable, and copy the answers to the next step of work.
The goal of agents is to receive a more abstract task, then independently break it down into steps, search for information, invoke tools, and continuously adjust according to intermediate results. For example, instead of asking AI to "write a marketing copy", users can ask it to "develop a marketing campaign for a new product". This task may require researching market trends, analyzing user feedback, identifying target audiences, proposing creative directions, generating content for different channels, and finally monitoring communication effects for further optimization.
Each of these links may be completed by different models, software and data systems. Therefore, the value of agents is not just adding a new way to generate content, but trying to become an intermediate layer that organizes different tools and processes. This is why operating systems, orchestration platforms, memory systems and context engineering are gaining increasing attention.
For an agent to truly enter an enterprise, it needs to understand at least three types of information. The first type is world knowledge, that is, public information learned during the training of general models; the second type is enterprise knowledge, including product materials, brand guidelines, customer information, business data and historical projects; the third type is organizational operation rules: who can access what data, what content requires approval, which departments a task needs to go through, and who is responsible when risks arise.
General large models can usually only solve the first type of problem well. They understand the general principles of marketing, manufacturing, finance and retail, but may not know how a specific company defines its brand, what a certain customer has purchased in the past, or who needs to approve a contract. The real assets accumulated by enterprises over the years are usually not fully recorded in public documents. They are scattered in databases, meeting minutes, emails, employee experiences, historical plans and a large number of implicit rules. Many decisions do not even have standard processes, and can only rely on people familiar with the business to make judgments.
This forms the core obstacle for agents to enter enterprises: models understand the world, but not the organization. Therefore, as foundational model capabilities gradually become public offerings, the focus of enterprise AI competition is shifting to context. Context is not equivalent to putting some files into a knowledge base. It also needs to address whether the information is up-to-date, whether different sources conflict, which business object the data belongs to, who has access permissions, and under what circumstances the model should invoke which piece of information.
Incorrect enterprise context can be even more dangerous than no context. If a model does not know the answer, it may prompt the user to supplement information; but if it cites an expired price policy, outdated contract or incorrect customer record, it may make seemingly reasonable but completely wrong judgments.
This is the starting point for Tezign to showcase its GEA enterprise-level agent architecture at this event. GEA consists of four layers: intent, orchestration, skill and context. The context layer attempts to convert the scattered brand data, content assets and business experience of the enterprise into a unified source of facts that can be invoked by AI; the orchestration layer coordinates different foundational models and modular skills according to tasks. Tezign hopes to turn enterprise AI from a one-time delivered project into a capability system that can continuously participate in enterprise operations and deliver results.
According to the company, it has served more than 180 enterprise customers so far, with its platform covering over 1 million professional users in 50 countries and regions around the world, and has been widely deployed in scenarios such as insight research, content growth, design creation and product innovation.
Tezign put forward a quite representative judgment: many core problems in enterprises are not mathematical problems with standard answers.
New product direction, brand strategy, user insight and creative design all require decision-makers to explore multiple possibilities first, and then make judgments in combination with organizational constraints. To this end, Tezign has also developed a Creative Reasoning Model for open-ended business problems, hoping that agents will not just quickly converge to a single answer, but first diverge for exploration, then make judgments and execute.
Whether this exploration can truly improve the quality of enterprise decision-making still needs to be verified by more real business results. But at least it reveals a core contradiction faced by enterprise agents: models are good at generating answers, but what enterprises really need is to manage the entire process.