J.P. Morgan frontline research: The "profitable moment" of Chinese AI has arrived, and the value of the workflow layer is beginning to surpass basic models.

The investment logic of China's AI industry is undergoing a subtle shift: while the market used to focus on model capabilities, parameters, and rankings, it is now more important to focus on who controls customer data, work flow entry points, deployment capabilities, and pricing power. Conferences covering autonomous driving, independent model development, enterprise workflow software, and vertical AI applications have given a common signal: models are still important, but in some enterprise scenarios, models are beginning to look like a replaceable input. According to Wind Trading Platform, Morgan Securities (China) analyst Yao Cheng and others released a core judgment in a research report on China's AI industry on May 22: "Some AI applications are beginning to show preliminary signs of commercial value, especially in vertical fields with intensive workflows and abundant data." The key point of this statement is not that AI applications can finally be monetized, but where monetization first appears: not in general chatbots or simple API calls, but in scenarios such as insurance and financial risk, enterprise data integration, and cross-border marketing, which are process-intensive, data-rich, and measurable. This also changes some assumptions about autonomous driving and application layers. The previous more conservative framework was that Robotaxi was more like a cost center in the short term, AI application monetization was still early, and the real clear exposure to the market was in infrastructure and computing power. Now, ADAS is closer to mass production and scale, and there are claims of economic improvements in city-level deployments of some L4 autonomous driving; vertical AI applications are also starting to show early evidence of recurring revenue, value pricing, and profitability on the operational level. However, most of this evidence comes from management calibers, and many companies are not yet listed, still some distance away from auditable and reproducible public market validation. The investment implications are instead more focused: computing power, AI infrastructure, domestic chips, memory and storage are directions that do not need to bet on the outcome of a single large model competition; at the application layer, proprietary data, ownership of the workflow, customer retention, and pricing power should be considered; model companies must prove that they are not just being called on, but can control high-value workflows. Consumer AI and intelligent agent e-commerce have not yet presented sufficient evidence of hard payment, retention, and transaction increment, so the valuation story still needs to be discounted. Model capabilities are still important, but the moat of selling APIs alone is thinning The most consistent signal comes from enterprise scenarios: customers care about whether the task can be completed, integrated into existing processes, and mobilize their own data, rather than necessarily using the largest model. Some companies have already allocated requests to different domestic models and advanced models based on price, performance, and specific tasks. Some companies admit that the cost of switching simple APIs is not high, the real reason customers stay is enterprise data access, workflow transformation, and business stickiness after deployment. This is uncomfortable for the valuation of model layers. Usage growth can continue to exist, but whether the rental stays in the hands of model manufacturers is another issue. If customers can switch between multiple models with low friction, pricing pressure on general APIs will become more evident. Of course, models still have value. In tasks such as coding, intelligent agents, and enterprise software automation, reliability, context length, tool invocation capability, and multi-step task completion rate directly affect business results. Model companies still have the opportunity to maintain the economic viability of high-value scenarios if they can control user interfaces, workflow memories, and data feedback loops. The applications that are monetized first are not general AI, but vertical scenarios that are process-intensive and data-rich Insurance and financial risk services, enterprise data integration, and cross-border marketing are the strongest application signals this time. Their common point is clear: customers pay not for "using AI," but for risk reduction, efficiency improvement, marketing conversion improvement, and decision automation. These scenarios are more conducive to value pricing. As long as the results are measurable, suppliers have the opportunity to convert AI capabilities into recurring revenue, rather than one-time project fees. But traditional software valuation logic cannot be directly applied here. Some businesses may still have strong service attributes: high customer concentration, a small share in the customer's wallet, early deployment requires a large amount of manual implementation or human assistance. If these costs are fully accounted for, gross margins and scalability may not be as good as they appear on the surface. So the more accurate judgment is: some AI applications have moved from narratives to early commercial validation, but have not proven their enduring and scalable software economics. The mainline of computing power is clearer instead: the more replaceable models, the more like common winners are basic infrastructure If enterprises use multiple models, the exclusivity of models will decrease, but the demand for computing power, storage, memory, cloud orchestration, and inference infrastructure will not disappear. This is where the logic of infrastructure is clearer. Regardless of which model ultimately leads, training, inference, enterprise deployment, ADAS development, simulation, data processing, and hosting services all require computing resources. Model competition may even increase the frequency of experimentation and deployment, thereby increasing demand for inference and supporting infrastructure. The key assumption is the elasticity of task volume. The optimistic scenario requires that AI task volume grow faster than the cost of single task decreases. If model efficiency, sparse computation, edge inference, or architectural improvements make unit computing consumption decrease faster, infrastructure benefits will be weakened. Currently, this framework tends to favor the former: demand is expanding from cutting-edge training to inference, adaptation of domestic technology stacks, memory, storage, and enterprise workflow execution. For the public market, this is still the clearest exposure to AI. Autonomous driving is no longer just a cost center, but L4 still has not passed The change in autonomous driving is that the previous pure money-burning narrative is starting to loosen. ADAS and L4 Robotaxi need to be separated. ADAS relies on whole vehicle production, real road data loop, and potential software batch profit margins, making it more closer to scale. Suppliers can follow the increase in installation rates by vehicle manufacturers to form a more visible revenue contribution. L4 Robotaxi is still much earlier. Some operators have claimed city-level economic improvements, vehicle cost reductions, and better economic deployment overseas, but company-level profitability is still a later goal, and regulatory barriers have not disappeared. Running successfully in one city does not mean it can be replicated in different geographies, weather conditions, and regulatory environments. For L4 companies, concrete evidence that can truly change the feasibility of investment is specific: auditable city-level unit economics, expansion to licensed cities, sustained vehicle cost reductions, safety records, and regulatory continuity. Any licensing suspension due to an accident could push back the commercialization process by several quarters. Neutral specialized companies can penetrate into car manufacturers and enterprises, and platforms may not necessarily dominate the application layer Platforms originally had cloud, traffic, ecosystem, and distribution interfaces, so it seemed that they should capture most of the value of AI applications. However, in procurement from enterprises and automakers, neutrality is becoming a variable. Some non-platform model or application companies can enter whole vehicle manufacturers and enterprise orders, not because they have a larger ecosystem, but because customers see them as more customizable and neutral suppliers. Platform clouds, maps, and infrastructure can still be used as input layers, but deployment context and customer workflows may be controlled by neutral specialized companies. This does not weaken the logic of benefit for platforms like Tencent and Alibaba in cloud, mapping, computing, and data infrastructure. The real difference is: as infrastructure suppliers, and platforms with their own models taking orders in the application layer, are two different things. When customers prioritize neutrality, customization, deep integration, platform models may not necessarily have an advantage. Procurement decisions will rely more on trust, integration depth, and vertical delivery capabilities. Consumer AI and intelligent agent e-commerce are still missing a hard bill Consumer AI seems to have weak evidence at the moment. Common problems include low user loyalty, intense competition, insufficient willingness to pay, and fast product imitation speeds. Intelligent agent e-commerce and AI advertising are also still in the early stages. In cross-border marketing, the current main optimizations still come from recommendation algorithms, and the intelligent agent commercialization driven by large language models still does not have sufficient verifiable income evidence. What needs to be seen here is not download volume, product releases, or demo effects, but retention rate, conversion rate, reuse rate, gross margin rate, and measurable transaction increments. Without these indicators, the consumer AI story is difficult to support a higher valuation. What really needs to be verified is retention, pricing, and gross profit, rather than the AI narrative itself For model companies, the key question for the next few quarters is whether the model layer can continue to capture value after customers route through multiple models. In the optimistic scenario, we need to see an increase in net retention rate, stable or increased pricing, low churn rate, and stable gross margin rates in the face of competition from multiple models. Coding may be the clearest testing ground. It is high-frequency, quantifiable in value, and has potential for locking in developer interfaces. Model-leading targets like KNOWLEDGE ATLAS and MiniMax cannot rely solely on benchmark test performance, but also need to prove vertical scenarios, workflow control, and quality of repeatable use. For application companies, the core indicators are revenue quality, customer concentration, implementation intensity, net retention rate, pricing structure, and true gross profit margin after deducting customer support and human assistance. For infrastructure companies, the key is whether the growth can be more clearly attributed to inference, deployment, and adaptation of domestic chips. If task volume expansion continues to be faster than unit cost reductions from efficiency improvements, computing power, domestic chips, memory, and storage are still the clearest mainline for AI. This article is retrieved from Wall Street News, edited by GMTEight: Chen Wenfang.