On November 10, Wu Minghui, founder, CEO, and CTO of MingLue Technology, gave a speech at the 2026 China Golden Forum. When discussing whether AI investment exists in a bubble, he focused the conversation on the cost-benefit for enterprises. In his view, the focus of AI investment in the first half was to create more capable models. In the second half, enterprises need to calculate how many tokens an Agent consumes to complete a business task and how much return this expenditure can bring. Agents taking on more work, the enterprise needs to calculate the cost of each individual task. Wu Minghui used two hypothetical accounts in his speech. For a company with tens of thousands of employees, if each employee consumes 100,000 yuan worth of tokens annually, the annual cost of inference would reach several billion yuan. For a consumer-level AI application with a daily active user base of 100 million, even if each user initiates one inference request per day, the platform would still bear the cost of 100 million requests per day. These two accounts correspond to different payment logic. Consumer-level applications that rely on free services to accumulate users need to find income sources that can cover the cost of inference. How much enterprises are willing to pay for AI can be linked to specific business outcomes. Wu Minghui therefore concluded that the enterprise market would be an important scenario for the commercialization of Agents. Whether the cost of model calls can be controlled as the business scale expands will affect how many positions and processes enterprises can deploy Agents to. Programming is an example Wu Minghui used to illustrate this change. From assisting programmers in completing code, to handling a development task independently, to multiple Agents collaborating in the same work group, the instructions given to AI can be shorter, but the underlying execution chain becomes longer. Judging whether this working method is cost-effective by looking only at the price per call is difficult. Enterprises also need to add up the costs of multiple calls within a single task and measure them against the delivery results. What model enterprises use also depends on where the data can go. MingLue Technology has applied multi-Agent collaboration in its internal R&D. According to Wu Minghui, engineers organize multiple Agents to collaborate on programming through a self-developed open-source platform called Octo. Octo supports private deployment, and different Agents can join the team as Bots. Personnel are responsible for judgment and calibration, and the knowledge, methods, and feedback formed during collaboration can be reused for subsequent tasks. Wu Minghui referred to the mode of connecting multiple Agents into the same network for collaboration as IoA, which stands for Internet of Agents. After Agents enter the workflow, they will also come into contact with enterprise documents and code repositories. Wu Minghui mentioned that enterprises need to know what data Agents can access and also need to understand which business contexts are sent to the cloud each time a model is called. He called the ability of enterprises to control their own data, Agent permissions, and model usage "enterprise sovereign AI." MingLue Technology hopes to help large enterprises build organizational intelligence platforms based on this. For enterprises, this also means evaluating on a case-by-case basis which data can be passed on to external models with tasks and which processes need to remain within their own environment for handling. Model selection can also be refined to the task. For many routine office tasks, Wu Minghui believes that cutting-edge open-source models are already capable. When dealing with more complex problems, enterprises can still call more powerful models. In specific business scenarios, he thinks it is also possible to enhance the performance of open-source models through post-training, allowing them to approach the capabilities of cutting-edge closed-source models in that scenario at a lower cost. MingLue Technology has already applied its internal accumulated token lean management capability to customer service. The charging model and computing power ownership of active Agents. When Agents start to proactively check emails, organize schedules, or complete research at night, the cost of inference will increase with the number of tasks they initiate autonomously. Wu Minghui explained the contradiction with two charging models. Charging based on token usage means the more active the Agent, the higher the cost borne by the user. Adopting a fixed monthly fee means that every additional run of the Agent will incur extra costs for the service provider. Wu Minghui proposed that using user-owned computing power is a path he is optimistic about. After enterprises purchase servers, they can arrange the frequency of Agent operations themselves, utilizing the already invested computing power on tasks that require continuous execution. Following this logic, enterprises also need to evaluate whether the equipment investment can be amortized over the long term and stable task volume. Wu Minghui also drew on the programmatic advertising ecosystem he encountered earlier at Secondhand System. The advertising technology market includes DSPs that represent advertisers to optimize ad spend, SSPs that help media increase traffic revenue, and ad trading platforms that facilitate transactions between buyers and sellers. In his view, the token market may also form similar divisions. The providers of computing power hope to increase their revenue, while the users want to complete the same tasks at a lower cost. In between, there may be a platform that offers matching services. Especially for users, they need someone to help them understand the amount of Token consumed by each task and optimize the model and cost accordingly.

From selecting a model to purchasing computing power, enterprises ultimately need to evaluate the specific work. How far can a task be handled by an Agent, how much computing power it consumes, and how much business value the delivered results can bring. These data will influence which AI applications enterprises decide to scale from pilot testing to full deployment.