36Kr | Three Key Signals from WAIC 2026: Computing Power Restructuring, Agent Delivery, and AI Hardware Closed Loop

How to ensure the intelligence we already possess operates long-term

Understanding specific people and organizations, entering the real world?

Seventy years ago, a group of young scholars first proposed the concept of "artificial intelligence" at the Dartmouth Conference.

Seventy years later, by the Huangpu River, the artificial intelligence industry is no longer satisfied with merely teaching machines to think and speak; it has begun to hand over this capability to enterprise systems, intelligent terminals, and robots, attempting to allow AI to truly participate in the operation of human society.

From July 17 to 20, the 2026 World Artificial Intelligence Conference and the High-Level Meeting on Global Governance of Artificial Intelligence were held in Shanghai. This year's conference, themed "Intelligent Partners, Co-Creating the Future," was spread across four venues in three locations: Expo, Zhangjiang, and West Bank. The exhibition area surpassed 100,000 square meters for the first time, with over 1,100 companies showcasing more than 3,000 exhibits, and over 300 products making their debut. The two major tracks of intelligent computing and embodied intelligence gathered more than 200 companies.

This is the largest WAIC ever held, and it may also be the most complex in terms of technology routes and product forms. Large models, super nodes, intelligent agent operating systems, AI phones, humanoid robots, dexterous hands, near-storage computing chips, and various industry solutions all appeared within the same exhibition system. At first glance, they belong to completely different technological tracks, but together they present a clear evolutionary path:

AI is evolving from a model that can generate answers into a system that can understand goals, mobilize resources, perceive environments, and deliver results.

In the past three years, large models have proven capable of writing articles, generating images, coding, and understanding videos, as well as handling increasingly complex reasoning tasks. However, by 2026, the key question facing the industry is no longer "Can AI do it?" but rather, "Can AI do it long-term, stably, and at low cost?" Successfully completing a task once in testing is entirely different from a system handling tens of thousands of business requests daily.

A robot completing a jump on stage is not the same capability as it continuously performing handling, sorting, and loading tasks in a factory for months. Transitioning artificial intelligence from capability demonstration to real production 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 promptly detected.

When the large model craze first emerged, the industry was most easily drawn to parameters, rankings, and model releases. Now, more and more companies are focusing their efforts on the systems behind the models and the real tasks in front of them. Models have not disappeared, nor have they become unimportant. On the contrary, they are being embedded into larger technological architectures, much like electricity, chips, and operating systems.

Through this WAIC, we have observed three migrations occurring in the AI industry.

01 Computing power is no longer just about chips, shifting from parameter competition to system competition

One of the most attention-grabbing products at this WAIC is the Huawei Atlas 950 super node prototype.

The key to super nodes is not how many chips are packed into a cabinet, but whether it can reduce the cost of data exchange between chips. High-speed interconnection, shared memory, task scheduling, and software stacks collectively determine whether a large number of independent chips form a team fighting alone or a computing whole that can work collaboratively.

However, as model sizes increase and reasoning tasks become more complex, a single chip is increasingly unable to independently determine final performance. The speed at which data is transmitted between chips, whether hundreds or even thousands of chips can collaborate efficiently, whether memory is sufficient, how tasks are scheduled, and whether software tools are mature all affect the actual utilization of computing power.

Even if a chip's theoretical peak is high, if a significant amount of time is spent waiting for data, cross-card communication, and task switching, it is difficult to translate that into effective model training and reasoning capability. It's 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 output.

AI computing power is undergoing a similar transformation. The basic unit of competition is gradually shifting from individual chips to servers, super nodes, and clusters; the dimensions of competition are also expanding from computing performance to high-speed interconnection, memory, heat dissipation, software ecology, energy efficiency, and overall operation and maintenance. Especially in the reasoning era, computing power systems are no longer just facing a few large-scale training tasks lasting months, but a large number of requests that vary in length and have significant concurrency fluctuations.

Intelligent agents need to repeatedly call models and tools, reasoning models need to generate longer thought processes, and multimodal models must simultaneously handle text, audio, images, and video. These changes will increase token consumption and raise higher demands for response speed, unit cost, and system stability.

Public data shows that by 2025, the scale of China's AI-related industries has surpassed one trillion yuan, with the overall penetration rate of AI in key industries exceeding 80%. The daily token call volume in China increased from 100 billion at the beginning of 2024 to 100 trillion by the end of 2025, and it continues to grow rapidly in 2026.

Compared to model parameters, token call volume is closer to the true level of the AI industry. Parameters represent the capabilities a model may possess, while tokens indicate that these capabilities are being utilized. Only when models are embedded in search, programming, office work, customer service, marketing, scientific research, and manufacturing processes will tokens continue to be generated. When the daily token calls reach hundreds of trillions, industry competition will naturally shift from "who has the largest model" to "who can produce intelligence at lower costs and higher efficiency." This is also the core background for the rise of super nodes.

It is not just about packing more chips into the same cabinet, but about minimizing the distance between chips, allowing a large number of processors to share data, memory, and tasks, thereby improving the effective computing power of the entire system. For domestic AI chip companies, this change brings both opportunities and raises the competitive threshold.

In the past, a company only needed to design a chip capable of running mainstream models to have a chance 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, internet, software platforms to cluster operation and maintenance. In other words, domestic computing power competition is shifting from "Is there a chip?" to "Can a system be formed?"

For an AI chip to truly enter a production environment, it also requires compilers, operator libraries, framework adaptation, performance tuning, and developer tools. Model architectures are rapidly changing; today's mainstream operators and precision formats may soon be replaced by new reasoning methods. Therefore, the value of open architectures lies not only in reducing licensing restrictions but also in allowing chips to continuously evolve and access more R&D resources.

However, whether it can ultimately form competitiveness still depends on mass production, software adaptation, customer migration costs, and long-term operational data. In the AI infrastructure field, any grand technological narrative must ultimately be tested against the same set of metrics: how many effective tokens can be generated per kilowatt-hour, how long a device can operate stably, how much time is needed for model migration, and whether it can quickly recover after a failure.

The true threshold of the computing power industry has never been just about lighting up chips, but about turning chips into productivity that customers can continuously use. This is why intelligent computing platforms, model service platforms, compilers, reasoning engines, and computing power scheduling systems are becoming increasingly important. If chips are the engine, super nodes and clusters are the vehicles, then software platforms determine whether this vehicle is easy 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 hope that computing power systems can be as stable, transparent, and easy to use as cloud services. Whoever can lower the barriers to using computing power will have a better chance of attracting developers and customers.

This round of computing power competition is also changing the cloud computing industry. Traditional cloud services primarily charge based on CPU, storage, and bandwidth, while AI clouds are increasingly providing services based on model calls, token counts, and task results. Computing power itself is gradually retreating to the background; what customers ultimately purchase is the speed of model training completion, the latency of reasoning responses, and the processing capacity for specific business tasks.

In the computing power industry, the term "Token Factory" is becoming a common concept to describe this shift. It focuses not just on how many chips have been deployed, but on how many usable tokens a system can stably generate under given power and hardware conditions. In this logic, chip peak performance is not the only metric; the overall system's throughput, energy efficiency, stability, and utilization rate determine the final output. The WAIC of 2026 further illustrates that while the performance of individual chips remains important, it is no longer sufficient to solely determine the competitiveness of an AI infrastructure.

The future leaders in computing power may not belong solely to companies with the strongest chips, but rather to those that can organize chips, interconnections, software, energy, and customer needs into a system. Model training shaped the previous stage of the computing power market, while large-scale reasoning will redefine the infrastructure of the next stage.

02 Intelligent agents enter enterprises, with the real barriers being context and delivery

If super nodes address how intelligence is produced, then intelligent agents focus on how intelligence is used. Agents are one of the most densely appearing concepts at this WAIC. From agent operating systems and intelligent agent phones to enterprise intelligent agents and multi-agent collaboration platforms, more and more companies are no longer satisfied with having AI answer questions; they hope AI can understand goals, break down tasks, invoke tools, and complete work with minimal human intervention.

The shift from chatbots to intelligent agents may seem like just a change in product names, but it corresponds to two completely different product logics. Traditional chatbots wait for users to pose clear questions and then return text, images, or code. They can assist humans in completing a specific step, but most of the work still requires human connection. Humans 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 task.

The goal of intelligent agents is to receive a more abstract task, then autonomously break it down into steps, seek information, invoke tools, and continuously adjust based on intermediate results. For example, users no longer ask AI to "write a marketing copy" but rather 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 the dissemination effects and continuing to optimize.

Each of these steps may be completed by different models, software, and data systems. Therefore, the value of intelligent agents is not merely adding another method for generating content but attempting to become an intermediary layer that organizes different tools and processes. This is why operating systems, orchestration platforms, memory systems, and context engineering are beginning to receive attention.

For an intelligent agent to truly enter an enterprise, it needs to understand at least three types of information. The first type is world knowledge, which is the common information learned during general model training; the second type is enterprise knowledge, including product information, brand specifications, customer information, business data, and historical projects; the third type is organizational operating rules, such as who can access what data, what content requires approval, which departments a task should go through, and who is responsible when risks arise.

General large models often can only solve the first type of problem well. They know the general rules of marketing, manufacturing, finance, and retail but may not know how a specific company defines its brand, what a particular customer has purchased in the past, or which people need to approve a contract. The true assets accumulated by enterprises over the years are often not fully documented in public files. They are scattered across databases, meeting notes, emails, employee experiences, historical plans, and numerous implicit rules. Many decisions even lack standard processes and can only rely on familiar business personnel for judgment.

This creates the core barrier for intelligent agents entering enterprises: models understand the world but do not understand organizations. Therefore, as foundational model capabilities gradually become a public supply, the competitive focus of enterprise AI is shifting towards context. Context is not equivalent to simply placing some documents 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 rights, and under what circumstances the model should invoke which information.

Incorrect enterprise context can be even more dangerous than having no context at all. If a model does not know the answer, it may prompt the user to provide additional information; however, if it references an outdated pricing policy, an expired contract, or incorrect customer records, it may make seemingly reasonable but ultimately completely incorrect judgments.

This is also the starting point for Tezign Technology's demonstration of the GEA enterprise-level intelligent agent architecture. GEA consists of four layers: intention, orchestration, skills, and context, where the context layer attempts to transform the scattered brand data, content assets, and business experiences of enterprises into a unified factual source that AI can call; the orchestration layer coordinates different foundational models and modular skills based on tasks. Tezign hopes to transform enterprise AI from a one-time delivery project into a capability system that can continuously participate in enterprise operations and deliver results.

According to the company, it currently serves over 180 enterprise clients, with its platform covering over one million professional users across 50 countries and regions globally, and it has been widely implemented in scenarios such as insight research, content growth, design creation, and product innovation.

Tezign has made a representative judgment: many core issues in enterprises are not merely mathematical problems with standard answers.

New product directions, brand strategies, user insights, and creative designs all require decision-makers to explore multiple possibilities before making judgments based on organizational constraints. To this end, Tezign has also developed a Creative Reasoning Model aimed at open business problems, hoping to enable intelligent agents to not just quickly converge on one answer but to first diverge, then judge and execute.

Whether this exploration can truly improve the quality of enterprise decision-making still requires more real business results for validation. However, it at least reveals a core contradiction faced by enterprise intelligent agents: models excel at generating answers, but what enterprises truly need is to undertake the process.

An answer is usually just a piece of text, while the process includes goal confirmation, information collection, plan comparison, permission approval, task execution, result evaluation, and accountability tracing.

When intelligent agents begin to participate in the process, they are no longer just a tool for improving efficiency but, to a limited extent, become a "digital employee" of the organization. This will change the form of enterprise software. In the past, employees needed to learn the menus and buttons of different software, constantly switching between CRM, ERP, content systems, databases, and collaboration tools. Intelligent agents may become a new interaction layer above these software. Employees only need to describe their goals, and the intelligent agent, after understanding the intent, invokes different systems to complete queries, analyses, and operations. Traditional software will not disappear, but more interfaces may retreat to the background, with human intent becoming the new entry point. However, from chatbots to digital employees, there is still a long way to go in building reliability.

If a chatbot answers a question incorrectly, the user can ask again; however, if an intelligent agent deletes an incorrect file, sends a wrong quote to a customer, or modifies production parameters, the consequences will directly impact the real world. Therefore, enterprise intelligent agents will not immediately leap from auxiliary tools to fully autonomous systems. A more realistic path is to first assist employees in searching, summarizing, and generating content, then invoke tools under human confirmation, and finally, possibly complete some low-risk tasks independently within defined boundaries.

Whether intelligent agents can enter core processes ultimately depends on three conditions: whether they can understand the real context of the enterprise, whether the execution process can be observed and audited, and whether errors can be promptly detected and halted. Models determine how smart an intelligent agent is, while context and governance mechanisms determine whether it is trustworthy. After enterprises introduce intelligent agents, they also need to rethink the evaluation methods for AI projects.

In the past, the success of a software could usually be measured by usage rates, user numbers, and renewal rates. Intelligent agents require a more complex set of metrics. Enterprises need to know their task completion success rates, average costs, manual takeover rates, and types of errors, and they also need to judge whether they have saved employee time or merely shifted work from execution to checking and correcting.

Even if an intelligent agent generates outputs quickly, if each output requires employees to double-check, its real efficiency improvement may be quite limited. This is also the dividing line for enterprise AI transitioning from demo to production environments. In demonstrations, intelligent agents only need to complete a carefully designed task; in enterprises, they must face vague instructions, missing data, system failures, permission restrictions, and constantly changing business rules.

Achieving success once is not difficult; maintaining success is the true technical threshold. Intelligent agents will also bring new organizational issues. When a digital employee can simultaneously undertake research, writing, data analysis, and task execution, traditional job boundaries may become blurred. Enterprises need to redefine the division of labor between humans and AI and decide who is responsible for the results of intelligent agents.

In the future, it is more likely that an intelligent agent will not completely replace a position, but rather an employee will manage multiple intelligent agents, or multiple specialized intelligent agents will collectively serve a team. The value of humans will shift from personally completing every step to setting goals, providing judgments, managing exceptions, and bearing responsibility. This also means that deploying intelligent agents in enterprises is never just a technical project.

It is equally an organizational engineering project. If an enterprise's existing data is chaotic, processes are vague, and there is a lack of collaboration between departments, intelligent agents will not automatically solve these problems; they may even exacerbate existing issues. Only when enterprises can clearly describe their business can machines truly understand and participate in it. This may be the most important competition for enterprise AI in the next stage: not who can quickly access a new model, but who can turn the model into a long-term operational part of the organization.

03 AI Gains a Body, Transitioning from Perceiving the World to Forming Real-Time Closed Loops

At this WAIC, the most intuitive heat still belongs to robots. Over 200 embodied intelligence companies gathered to exhibit humanoid robots, quadrupedal robots, dexterous hands, joint modules, sensors, and training platforms, collectively forming an ever-expanding industrial chain. Compared to a few years ago, the discourse system of the robotics industry has undergone significant changes. In the past, the focus of humanoid robot demonstrations was on "how human-like they are": can they walk, dance, run, and perform complex actions? Today, the industry is increasingly concerned with whether they can work.

Once robots enter factories, can they adapt to parts of different sizes? When production lines change, do engineers need to reprogram? How close can the speed of task completion be to that of humans? What are the failure rates and maintenance costs during continuous operation?

This indicates that embodied intelligence is transitioning from action performance to task delivery. The reason large models are changing robotics is that they provide a new technological path for machines to understand language and environments. Traditional industrial robots typically rely on preset programs. Under fixed positions, fixed rhythms, and fixed workpiece conditions, they can repeat the same action with high precision, but once the environment or items change, engineers need to re-tune them.

Embodied intelligence aims to give robots a certain degree of generalization ability. Humans can directly tell robots to "clean the table" or "put different parts into corresponding boxes," and the robots will convert natural language into action plans. However, the real world is far more complex than the language world. If a language model makes a judgment error, it can regenerate; if a robot makes a judgment error, it may break items, interrupt production lines, or even harm those nearby. Therefore, for robots to truly gain working capabilities, they need to form a complete "perception-understanding-decision-execution-feedback" closed loop.

First, they need to understand the environment through cameras, LiDAR, force sensors, and tactile systems; then break down human goals into specific actions; and finally execute through joints, motors, and dexterous hands, making real-time corrections based on results. This is also why robot competition cannot solely focus on large models. The body structure, reducers, actuators, sensors, batteries, control systems, and data acquisition capabilities also determine whether products can be implemented. For factories, whether robots resemble humans is not important; what matters is whether they can consistently create value. Industrial scenarios may become the first market where embodied intelligence achieves large-scale implementation.

Factory environments are relatively structured, task objectives are clear, and it is easier to calculate input-output ratios. Robots do not need to possess general intelligence from the start; as long as they can stably complete specific tasks such as handling, sorting, quality inspection, loading and unloading, and patrol inspection, they have the opportunity to form a commercial closed loop. However, even in industrial scenarios, the large-scale implementation of robots still faces challenges.

An action success rate of 99% sounds high, but if a production line needs to execute tens of thousands of actions daily, the remaining 1% still means a significant number of anomalies and manual interventions. Therefore, industrial customers care less about how stunning a demonstration is and more about whether the task success rate remains stable after hundreds of hours of continuous operation. Robot companies are also re-evaluating the balance between products and business models.

Fully general humanoid robots have the greatest imaginative space but need to solve issues related to perception, decision-making, embodiment, data, and costs simultaneously. In contrast, starting from specific tasks such as handling, sorting, and patrol inspection is more likely to form revenue and data closed loops. Robots accumulate data through real scenarios, which in turn trains models and improves hardware, potentially providing a realistic path for embodied intelligence to gradually achieve general capabilities.

Past intelligent hardware often waited for human operation. Users opened apps, pressed buttons, and then received results from the screen. The new generation of AI terminals hopes to continuously perceive users' voices, images, locations, and physical states, proactively providing assistance at the right time. The premise of this shift is the continuous enhancement of edge AI capabilities.

The cloud can provide stronger general reasoning capabilities, while edge computing offers advantages such as low latency, local operation, privacy protection, and personalization. As model compression, chip performance, and terminal computing power continue to improve, more and more AI tasks will be dynamically allocated between the cloud and local devices. Tasks that do not require real-time responses and have larger computational loads can be assigned to the cloud; tasks involving continuous perception, personal data, and immediate feedback are more suitable for completion on the device locally.

Especially in data involving health, images, sounds, and locations, the significance of edge computing is not just to increase speed. It also determines whether devices can operate without continuously uploading sensitive information.

Of course, there is still a long way to go from consumer-grade neural devices to widely validated health solutions. EEG signals are easily affected by actions, wearing methods, and individual differences, and the actual effects of acoustic feedback also require long-term, standardized user data and research validation.

However, neural perception products at least provide a way to observe the future of AI hardware. Robots attempt to understand the external world, while smart wearables begin to try to understand a person's internal state. Although they seem to belong to completely different tracks, the underlying logic is becoming increasingly similar: continuously collecting signals, assessing the current state, generating actions or feedback, and then continuously adjusting based on new results.

This means that AI is transitioning from a question-and-answer tool to a continuously operating closed-loop system. It appears not only in the bodies of robots in factories but may also exist in smartphones, glasses, headphones, and devices that people wear daily. Once AI gains a body, the standards for evaluating industries will also change. An error in a chat product may only be an inaccurate answer; however, an error in a device that controls a robot or continuously reads bodily signals may lead to real-world consequences. Therefore, the closer AI gets to the physical world, the higher the requirements for safety, explainability, and reliability.

Intelligent devices need to clarify which data is processed locally and which data needs to be uploaded; robots need to know under what circumstances they should stop actions; intelligent agents need to request human confirmation before executing high-risk tasks. These capabilities are not additional auxiliary functions but are necessary conditions for AI to transition from demonstration to large-scale application.

This WAIC featured over 300 debut products and also promoted a batch of industrial scenarios and cooperation projects. However, for the AI industry, the number of debuts is no longer the only important metric. The market is no longer lacking in new models, new intelligent agents, and new robots; what is truly scarce is systems that can continuously produce results. How much repetitive labor can be reduced? How long can costs be recouped after deploying robots? What is the utilization rate and stability of computing systems? Does an AI hardware truly solve user problems? These questions may not be as eye-catching as technical demos, but they determine whether an AI company can cross the threshold of commercialization.

Thus, the competitive unit of the AI industry has also shifted from individual models to complete systems. This change can be summarized as three migrations: from individual chips to computing power systems capable of producing effective tokens; from answering questions to understanding enterprise context and undertaking business processes; from existing in chat boxes to possessing perception, action, and real-time feedback capabilities in physical terminals.

None of these three migrations are easy to accomplish. They require collaboration between algorithms and hardware, between technology companies and industry clients, and between product innovation and reliability building. Therefore, WAIC 2026 is not declaring that AI has become omnipotent. On the contrary, it indicates that the entire industry has finally begun to face more complex issues than training large models: how to ensure the intelligence already possessed operates long-term, how to enable it to understand specific people and organizations, and how to allow it to enter the real world while remaining reliable and controllable.

(Source: WeChat Public Account 36Kr)