In the interview, Lin Yuan repeatedly emphasized a critical issue: the enterprises you see today do not lack the willingness to adopt AI; rather, they dare not use it. Listed companies fear data leaving their secure domains, banks are constrained by regulations and worry about compliance, and the manufacturing industry considers processes and formulas as their lifelines. None of these enterprises can use public cloud large models. Every customer is anxious about whether data can remain within their domain, whether their original systems can remain untouched, and what tasks the AI Agents are actually running after entering the enterprise.
As the CEO of QingCloud, Lin Yuan has been fighting in the cloud infrastructure arena for over a decade. He has experienced the fierce competition in the public cloud market and encountered numerous pitfalls. In the AI era, he chose a path quite different from many startups: instead of engaging in a price war for public cloud tokens, he focuses on large enterprise (large B) customers, helping them build private, controllable, and proprietary AI infrastructure. Behind this judgment lies a more fundamental industry belief. He believes that the most stable demand in the AI era is not computing power, models, or price, but privacy. And this demand is rewriting the competitive rules of the entire enterprise AI market.
Recently, AIFenxi conducted an in-depth exchange with Lin Yuan. Centering on themes such as data sovereignty, enterprise AI infrastructure, and AI Agent management, Lin Yuan shared QingCloud's judgments and practices in transitioning from cloud computing to AI infrastructure.
Core Perspectives
- Privacy is becoming the most stable driving force for payment in the AI era.
Enterprises do not lack the willingness to adopt AI; rather, they dare not let high-value data leave their secure domains. Large enterprises still keep data, models, and computing power within their control through privatized approaches.
- What enterprises want is not a specific model service, but their own AI.
Large enterprise customers have too many business types to be solved by a single model. What they need is an internally operated AI infrastructure to support different departments, different models, and different permissions, etc.
- Tech giants will not consume the enterprise AI privatization market.
For tech giants, the public cloud and public model services are the main battlegrounds; privatized AI is not the highest priority business in many scenarios.
- The high price of GPUs has increased the willingness to pay for GPU management software.
The larger the hardware procurement amount and the more massive the cluster, the stronger the enterprise's demand for scheduling, model management, and O&M (Operations and Maintenance) software, and the higher their payment capacity will be.
- AIDC (AI Data Center) is a long-term asset in the AI era.
QingCloud regrets not building IDCs (Internet Data Centers) in the CPU era, and it must not miss the opportunity again in the AI era.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. The Greatest Payment Motivation for Enterprise AI Infrastructure is Data Privacy
AIFenxi: Transitioning from traditional cloud to AI infrastructure, what has changed and what remains the same?
Lin Yuan: From a technical perspective, the changes are actually not as significant as imagined. Building clouds in the past was also about infrastructure, dealing with networks, CPUs, storage, and scheduling. The design concepts of large-scale, stability, and performance have not undergone fundamental changes. Now the entire AI is also based on a cloud-native architecture, and we started working on cloud-native and open-source very early, so our technical accumulation is still on a major extension line. What has really changed significantly is that customers value the business-level value of infrastructure. In the past, when a customer expanded from 100 to 1,000 people, server resources did not grow linearly, and the number of business systems the customer needed to run was limited. We didn't care much about the customer's business; we just provided resources, and customers looked at whether your resources were abundant, cheap, and stable enough. But AI is different. AI must interact with data and business; otherwise, it has no value. Whether it is the growth of carbon-based employees or silicon-based employees in an enterprise, new demands will be placed on AI infrastructure. At the same time, users will highly value data privacy. In the past, building a cloud was like building a house; how to live in it was the customer's business. Now, every model call by an enterprise on the public cloud is exchanging its data sovereignty for AI value. This is our most important judgment on the market: in the AI era, whether it is To B or To C, everyone will eventually pay for privacy.
AIFenxi: Why does AI make data sovereignty so important?
Lin Yuan: Because to generate value, AI must touch the core data of the enterprise. Enterprises will definitely not let valuable data leave their secure domains. Data that leaves the domain must have already been judged and deemed safe to leave. If you want to create valuable AI scenarios, you have to face valuable data, so why should customers share it? Therefore, many enterprises today do not lack the willingness to adopt AI; rather, they dare not use it. Can listed companies hand over their data? Banks are constrained by regulations. The manufacturing industry considers processes and formulas as their lifelines. The demands are all urgent, but the first reaction is always: can the data stay within the domain?
AIFenxi: How to solve the problem of data leaving the domain?
Lin Yuan: Currently, there are two main models. The first is private deployment, where customers procure assets themselves, and we deploy the platform, keeping data, models, and computing resources within the customer's internal environment. The second is dedicated cloud, where customers do not want to spend too much cash flow at once; equipment is deployed in designated data centers, and we provide overall delivery. We also hold GPU assets for some customers, but the logic is not to buy equipment first and then look for demand, but to look at the customer first. If the customer quality is high enough and the cooperative relationship is stable enough, we can help them hold assets. QingCloud has experienced fierce competition in the public cloud market and later gradually clarified that its basic market is large enterprises, serving large customers in a private and dedicated manner. In the AI era, this route is clearer: enterprises need an AI infrastructure that can protect data sovereignty and enter real businesses.
AIFenxi: Can keeping data within the domain solve all privacy problems?
Lin Yuan: Keeping data within the domain is just the baseline, not the whole picture. Enterprises deploy both models and data internally, but AI Agents still need to connect to Feishu, WeCom, and call external Skills. If these connections transmit data out, it is still non-compliant. Therefore, enterprises also need to manage the tools, interfaces, and external services that Agents can call. The installation of any Skill must be audited, and Agents cannot be allowed to freely crawl content from the external network. Security is truly established only when the entire link is within the control scope.
02. What Enterprises Want is Not a Specific Model, but Their Own AI Infrastructure
AIFenxi: Entering the AI infrastructure field, why does QingCloud still regard large enterprises as core customers?
Lin Yuan: Because the needs of large and small enterprises are completely different. Small B customers are simple; perhaps one model is enough, and they will use whoever is cheaper. Large B customers are not like this. Look at state-owned enterprises, finance, and large-scale manufacturing; they have a particularly wide variety of businesses. An insurance customer might use Qwen for OCR, DeepSeek for backend Agents, Kimi for text analysis, and the engineering team needs something else. It is like an internally operated AI cloud. What enterprises want is their own AI, not a single model service. Only in their own environment, with data not leaving the domain, departments being isolated, and ROI being measurable, do they dare to use it. Moreover, the procurement logic of enterprises is completely different from that of small B customers in the public cloud. A physical machine in the public cloud must run 100 virtual machines to make money, and users also care more about whether it is cheap enough. But in a private cloud, enterprises are willing to pay for compliance and services; they already hold server assets and do not care about the price of a single output. For example, when running virtual machines, the number of virtual machines on a physical machine in a bank is clearly written in the management specifications and will not exceed the standard.
AIFenxi: How does QingCloud translate these needs into products?
Lin Yuan: We focus on large B customers and build three layers. The first layer is a heterogeneous computing power scheduling platform, which solves the management and scheduling of resources such as GPUs. Enterprises can see what cards are available, how they are allocated, and how many resources different tasks use. The second layer is an enterprise-level Token platform, which can also be understood as an enterprise-level MaaS (Model as a Service) platform. Enterprises can deploy and manage different models to provide unified model inference services for various departments. The third layer is an enterprise-level Agent working platform. Ordinary employees do not care what model is called at the underlying level; they only care about whether the work can be completed. Employees can assign tasks to AI Agents, letting them call internal enterprise systems and resources for execution. Ultimately, it boils down to one sentence: help enterprises build their own AI, enabling them to dare to use AI, be able to use AI, and use AI well.
AIFenxi: Compared to the IaaS (Infrastructure as a Service) cloud platforms of the past, what is the difference?
Lin Yuan: The core of IaaS in the past was resource management, which is the first layer of today's GPU cloud platforms. In the AI era, IaaS cloud platforms extend upwards to Tokens and Agents, making the system thicker than before. But the truly difficult things are not in the technology itself. You see, GPU cloud is actually simpler than CPU cloud because the Kubernetes (k8s) architecture makes the underlying layer sufficiently standardized. The real difficulty is how to enter the enterprise's production environment. The means of production for enterprises have long existed in systems such as OA, ERP, CRM, email, and knowledge bases. AI must enter the enterprise context and connect with these existing systems to truly generate value.
03. Agents Are Also Employees; Enterprises Want to Know What Employees Are Doing
AIFenxi: By introducing QingCloud's Agent working platform, do the enterprise's original systems need to be rebuilt?
Lin Yuan: Customers have explicitly stated not to touch the existing systems. OA, CRM, ERP, email, and Wiki are just there. We do not transform them, but I can connect through Connectors, so the Agent can work within the context of the original systems.
AIFenxi: Why do enterprises pay special attention to Agent management?
Lin Yuan: Because the ones actually doing the work are Agents. Essentially, an Agent is the digital twin of an employee, inheriting the employee's permissions and theoretically capable of doing many things. But for enterprise managers, while they think the employees are excellent, they also particularly want to know what the employees are doing. Therefore, enterprises need three things. First, observability: I need to know which Agents are working, who is using them, and what they are busy with. Second, controllability: what an Agent can and cannot do must have boundaries. Third, high-risk operations must be confirmed by humans. Agents can plan and execute tasks, but they cannot decide on their own to delete data, modify core business information, or trigger important processes.
AIFenxi: Seeing this, the requirements for the Agent working platform from frontline employees and enterprise managers are completely different?
Lin Yuan: Completely different, but both must be satisfied simultaneously. We sell to decision-makers, but the experience must serve users. For employees, it must be easy to use and willing to be used. For bosses, it must be controllable and have a framework to constrain it. If either end is missing, the product cannot stand. The users and procurement decision-makers of large B products are often not the same group of people. If the experience is poor, employees won't touch it; if the management is inadequate, bosses won't approve it.
04. Tech Giants Will Not Consume the Entire Enterprise AI Market
AIFenxi: In QingCloud's track, who will be the main players?
Lin Yuan: Roughly several categories. The first category is server manufacturers; they give away GPU management and scheduling platforms when buying servers, but this is not their main business. The second category is those who originally did containers. Containers are inherently difficult to sell in China, and many have also turned to AI. The third category is startups, but their logic is different from ours. They need to run revenue, seek valuations, and raise financing. We must return to normal business logic, and products must be able to make money.
AIFenxi: What about tech giants?
Lin Yuan: In the enterprise market, especially in the privatization segment, tech giants will not exert force; they are unwilling to enter this business. I have encountered large manufacturing enterprises that originally used tech giants' products, but later the giants stopped iterating. Because the strategic core of tech giants is always public cloud services, and that scale is too large. QingCloud is precisely to select those customers who truly need continuous privatized iteration. If a customer's scale is very large and they need to bind with tech giants, then we cannot do it, and we accept that.
AIFenxi: From the perspective of business models, what is better about GPU management in the AI era compared to CPU management in the past?
Lin Yuan: Several aspects. First, we are in the software business, and the value of software will definitely be linked to the procurement scale of the underlying hardware. In the past, a CPU server might only cost tens of thousands of RMB, but now a GPU server might cost millions of RMB. After the hardware procurement amount increases significantly, the enterprise's requirements for resource scheduling, model services, permission management, and O&M assurance will also improve synchronously. Second, the scale of resource usage is growing. A hundred-billion-level manufacturing enterprise used the cloud in the past just to run OA and ERP, and 50 servers were enough. Agents are essentially employees and productivity; the more people you need, the larger the usage. We have a biopharmaceutical customer that used to invest over 2 billion RMB in R&D annually. The costs came from the manpower investment of PhDs, imported experimental instruments and equipment, etc. After introducing GPUs, they conducted post-training of models to assist scientific research, using simulations to reduce real experiments, with servers running 24/7. In the first year, they only deployed a small cluster. The customer calculated it themselves and saved 400 million RMB in R&D expenses in one year. Third, customer opportunities in new industries; emerging customers in embodied intelligence and autonomous driving are being created. The superposition of these three things makes the enterprise privatized AI market larger, with higher average customer unit prices. Moreover, the customers' models, data, and business processes are precipitated on the platform, resulting in high stickiness. Our earliest private cloud customers have been renewing until now; in the AI era, the binding time will be longer and the service depth will be deeper.
AIFenxi: Beyond software, is QingCloud considering venturing into the hardware part?
Lin Yuan: In the CPU era, we did not build IDCs, and looking back now, it is somewhat regrettable. At that time, influenced by Silicon Valley thinking, we felt that tech companies should be asset-light. But after doing cloud for more than a decade, you will see a fact: renting other people's IDCs for a long time involves huge cumulative investments. If we had precipitated some high-quality assets early on, today's value would be completely different. In the AI era, we do not want to miss out again. AIDCs can carry AI computing power for a long time and are a kind of basic asset. It is like having a piece of land in the Second Ring Road of Beijing twenty years ago; would you buy it or not? At that time, you might have felt there was nothing around, but after urban expansion, it has value. AI is now going through this stage, finding places and building houses for Agents. It is hard to see an upper limit to human demand for AI, and computing power and power consumption will continue to grow in the future. In places with good energy and network conditions, high-quality AIDCs have long-term asset value.
AIFenxi: When did you start considering participating in AIDCs?
Lin Yuan: AIDCs require a large amount of capital, and it is impossible to invest entirely on our own. We started researching places like Ulanqab in Inner Mongolia two years ago, but we were constrained by capital. The current idea is to work together with large customers. Some customers themselves have infrastructure investment capabilities. QingCloud participates in early planning, approval, architecture design, and platform construction, helping customers connect data centers, GPU resources, and upper-level AI platforms. We do not necessarily have to hold all the assets, but we must gradually enter the asset layer and accumulate capabilities in AIDC planning, construction, and operation.
AIFenxi: From the perspective of the overall future trend of the AI industry, which layer has greater value, infrastructure or applications?
Lin Yuan: In the future, it will still be applications. Infrastructure is like electricity; when have you ever seen a power plant make particularly huge amounts of money? In the future, applications will take the most money because they are closest to value. It's just that the application market is like winning the lottery; the money all goes to the top products, and it is too difficult for startups to bet on this probability. So many people return to infrastructure as they go along, because the boundaries of infrastructure are relatively clear, and investing in it can yield some output. QingCloud excels in infrastructure; rather than chasing what we are not good at, it is better to deepen the things with stronger certainty.
AIFenxi: Then what will future applications look like?
Lin Yuan: I think the high probability is that they will be an integration of software and hardware. AI has made human interaction increasingly simple. Keyboards, mice, touch screens—now people are too lazy to even look at the screen and just speak directly. When you speak with your mouth, you do not need software in the traditional sense. The software is still there, but it has become invisible, hidden behind the hardware. That new carrier could be glasses, watches, or earrings. The interaction method will definitely change, and it may not be the traditional software form we imagine.