Author | Sun Rui, WeChat | lombredeau
Agents are bringing a new wave of demand to cloud computing.
Baidu's Q2 2026 financial report shows that Baidu AI Cloud infrastructure revenue reached CNY 7.3 billion, a year-on-year increase of 50%; among which GPU cloud revenue surged by 283% year-on-year, further accelerating from the 184% growth in the previous quarter. Baidu stated that the growth in GPU cloud revenue reflects the continuous increase in market demand for public cloud AI computing.
Shifting the focus one level up from computing power, it becomes evident that the application boundaries of AI are expanding from point solutions to entire enterprise business processes.
Since the beginning of this year, AI office applications have become the most fiercely contested To B scenario among tech giants. Multiple major companies have started integrating their product lines and launching their own office agents, further embedding agents into the daily work of enterprises.
However, office work is only a part of enterprise business. Most key tasks that determine business outcomes often occur outside office software. Core business links such as R&D, production, risk control, and design involve a large number of complex and highly specialized tasks.
On September 16, at the 2026 Intelligent Economy Forum, Baidu AI Cloud proposed the concept of the "Industrial Agent Operating System" for the first time, organizing general agents, professional agents, and customized agents to enable them to run collaboratively and in a closed loop within enterprise businesses.
What Baidu AI Cloud aims to do is not just about how enterprises use agents, but how to organize scattered agent capabilities, embed them into enterprise business systems, and ultimately form a sustainably operating intelligent production system.
Office is the Entry Point, Core Industrial Scenarios are the Next Stop
The reason office agents exploded first is that their tasks are relatively standardized, and the payment models are mature.
Therefore, office work naturally becomes the entry point for agents to enter enterprises.
However, truly high-value tasks in enterprises often require more complex capabilities. For example, in automotive R&D, a single R&D task may simultaneously involve the enterprise's historical vehicle model data, design specifications, simulation tools, R&D knowledge bases, and specific engineering processes; financial risk control requires understanding internal enterprise rules, industry knowledge, and real-time data; quality inspection, production scheduling, and supply chain management in manufacturing also have their own complex business logic.
These tasks deeply embedded in the industrial chain are difficult to be directly completed by an office agent. This is because what enterprises need is not simply integrating AI capabilities, but enabling AI to truly understand how the company conducts its business.
Baidu AI Cloud's judgment on industrial agents is precisely built on this. At the forum on September 16, Shen Dou, Executive Vice President of Baidu Group and President of Baidu AI Cloud Business Group, categorized enterprise tasks into three types: 20% daily operations, 30% professional domain tasks, and 50% core tasks that determine business outcomes. He believes that the capabilities to perform these tasks will gradually be internalized by the corresponding agents in the future.
During this forum, Baidu launched a development platform based on DuMate and announced the opening of the DuMate ecosystem. Providers of capabilities and products such as connectors, skills, suites, and hardware can join the platform, which promises not to charge commissions from partners. By opening up the development, invocation, and connection capabilities of agents, Baidu attempts to further organize scattered agent capabilities into the enterprise's business system.
Behind this corresponds exactly Baidu's understanding of industrial agents. Office agents mostly solve how enterprise employees use AI and let AI participate in daily work, whereas industrial agents need to further enter the enterprise's business systems and production operation processes, handling tasks that require the joint participation of enterprise data, professional knowledge, business rules, and existing systems.
From Personal Agents to the Enterprise "Agent Operating System"
The difficulty with industrial agents lies in the fact that complex tasks in enterprises are never completed independently by a single software.
Behind a real business task, there are often data, knowledge, processes, models, tools, and computing power simultaneously. In the past, these capabilities were scattered across ERP, CRM, databases, knowledge bases, and various professional software, and enterprise employees were responsible for stringing them together.
With the emergence of agents, this connector role has begun to be taken over by AI.
However, if there are only independent agents in the enterprise, the problem is not truly solved. Each agent may have its own data, tools, and working methods, which will ultimately form new information silos.
The core of the industrial agent operating system proposed by Baidu is to organize these scattered capabilities.
Specifically, the industrial agent can be understood in three layers.
The first layer is the general agent, which handles general tasks such as writing, making PPTs, searching for information, and code development. DuMate belongs to this layer.
The second layer is the professional agent, which provides capabilities for specific industries and professional fields, such as agents for industrial decision-making, marketing, and R&D scenarios. Baidu has currently formed a professional agent matrix including FM Agent, Baidu Shengsuan, MeDo, Hogee, and Baidu Comate.
The third layer is the customized agent. It further integrates an enterprise's own data, knowledge, systems, and business processes to solve specific problems that only exist within that company.
The prerequisite for the three layers of agents to truly play their role is that they can be organized.
For example, after an automotive R&D task enters the system, the operating system first understands the task and the enterprise's business background, and then invokes the corresponding professional agent; the professional agent continues to connect with the enterprise's knowledge base, historical data, and simulation tools to complete calculations and deductions; finally, the results re-enter the R&D process. For the enterprise, what it gets is no longer an AI-generated answer, but a result that can continue to be executed in the business process.
The "operating system" is not intended to replace the original ERP, CRM, financial, or supply chain systems, but to organize these systems along with the data, processes, and experience accumulated in them, enabling agents to invoke them.
This will also bring about a change: in the past, enterprises purchased individual standalone software; in the future, more and more capabilities may be encapsulated as Skills and invoked by agents on demand.
Therefore, the key to industrial agents is whether these agents can form a continuously running system around real businesses.
Baidu AI Cloud already has extensive experience in industrial implementation.
According to data disclosed by Baidu AI Cloud, it has served over 800 financial institutions, covering all systemically important banks, and has ranked first in the market share of China's financial large models for two consecutive years; in the automotive sector, Baidu AI Cloud claims to have covered China's mainstream car manufacturers and maintained a leading position in the autonomous driving R&D solutions market; in the field of embodied intelligence, it has served more than 50 related enterprises.
The significance of these industrial customers is that what they provide is not a standardized office demand, but a large number of real business processes. For agents to enter these processes, they must solve the connection problems among data, knowledge, permissions, tools, models, and computing power.
To Get Industrial Agents Running, AI Cloud Must Also Be Upgraded
After agents enter the industry, AI cloud needs corresponding upgrades.
Traditional AI applications are often one-question-one-answer, with one invocation corresponding to one computation. However, in the industrial agent scenario, a complex task may require the participation of multiple models, multi-round reasoning, invoking enterprise knowledge bases and databases, and then invoking external systems through MCP connectors, while also needing to continuously save task states, memory, and execution trajectories.
This makes Agent Infra a new challenge for AI cloud. The answer provided by Baidu AI Cloud is the full-stack capability of "Chip-Cloud-Model-Agent": the underlying chips provide computing power, the AI cloud is responsible for computing power and resource scheduling, the models provide intelligent capabilities, and finally, the three types of agents bring these capabilities into specific business scenarios.
The scarce value of full-stack capabilities is also reflected here.
In the past, when enterprises deployed large models, the focus was on solving training and inference; when agents start running continuously, enterprises also need to solve how to schedule models, how agents collaborate, how tools are invoked, how tasks are tracked, and how data and permissions flow securely.
This is also the reason why Baidu AI Cloud is looking at the industrial agent operating system and the full-stack AI cloud together this time.
At the computing power layer, Kunlunxin has accumulated over ten years of experience, and the third-generation P800 has currently achieved large-scale cluster deployment in fields such as internet finance, manufacturing, automotive, and telecommunications operators.
At the Agent Infra layer, Baidu AI Cloud provides more than 150 SOTA models, long-chain reasoning optimization, built-in sandbox and risk control capabilities, focusing on solving the elasticity, security, and transparency issues during the operation of agents.
At the agent layer, when agents generate more real business data, this data can continue to be accumulated as enterprise knowledge and capabilities, driving agent iteration; as the scale of agent operation expands, it will bring more model invocations and computing power demands.
Thus, "Chip-Cloud-Model-Agent" forms a cycle from the underlying infrastructure to enterprise production and operation.
When AI transforms from a tool into a participant in enterprise business processes, the cloud is no longer just the infrastructure carrying AI, but begins to become the base for the continuous operation of agents.
Ultimately, what is formed by "Chip-Cloud-Model-Agent" is not just a technical chain from computing power to applications, but a productivity growth flywheel that allows AI capabilities to continuously enter enterprises, accumulate experience, and in turn drive the next round of computing power and model demands.
Looking back at history, every technological paradigm shift has given birth to a new "operating system" layer. For example, Windows in the PC era, and iOS and Android in the mobile era.
The essence of an operating system has never been a product, but a structural dependency. And the dependency in the agent era may be deeper than in any previous round. Baidu AI Cloud's first proposal of the "Industrial Agent Operating System" is precisely a forward-looking bet on the infrastructure layer that will run AI in future enterprises.
This battle has just begun.