The investment thesis for China's AI industry is undergoing a subtle shift: the market, which previously focused on model capabilities, parameters, and leaderboard rankings, should now pay closer attention to who controls customer data, workflow entry points, deployment capabilities, and pricing power. Eight meetings covering autonomous driving, independent model development, enterprise workflow software, and vertical AI applications delivered a common signal: while models remain important, in some enterprise scenarios, they are beginning to resemble replaceable inputs. Analysts including Yao Cheng from JPMorgan Securities (China) noted in a May 22nd industry research report on China's AI sector: "Some AI applications are beginning to show preliminary signs of commercial viability, particularly in workflow-intensive and data-rich verticals." The key point is not that AI applications can finally generate revenue, but where monetization is appearing first: not in general-purpose chatbots or simple API calls, but in process-heavy, data-rich, and results-measurable scenarios like insurance and financial risk, enterprise data integration, and cross-border marketing. This also alters some assumptions about autonomous driving and the application layer. Previously, a more conservative framework suggested robotaxis were nearer-term cost centers, AI application monetization was distant, and the clearest public market exposure was in infrastructure and computing power. The current view suggests ADAS is closer to mass production and scale, with some L4 autonomous deployment reportedly showing city-level economic improvements; vertical AI applications are also beginning to show early evidence of recurring revenue, value-based pricing, and operational profitability. However, much of this evidence comes from management commentary, and many companies are not publicly listed, remaining distant from auditable, replicable public market validation. The investment implications, however, become more focused: computing power, AI infrastructure, domestic chips, memory, and storage represent directions least reliant on betting on a single large model's success; the application layer should be evaluated based on proprietary data, workflow ownership, customer retention, and pricing power; model companies must prove they are not merely being called via APIs but can control high-value workflows. Consumer AI and agent-driven e-commerce have yet to provide sufficiently robust evidence on paid users, retention, and transaction growth, so their valuation narratives still require discounts. Model capability remains important, but moats for selling pure APIs are thinning The most consistent signal comes from enterprise scenarios: customers care about whether tasks can be completed, integrated into existing workflows, and leverage proprietary data, not necessarily about using the largest model. Some companies are already distributing requests to different domestic and frontier models based on price, performance, and specific tasks. Others admit that the cost of simply switching APIs is not high; what truly retains customers is enterprise data integration, workflow transformation, and post-deployment business stickiness. This is uncomfortable for model-layer valuations. Usage growth may continue, but whether the 'rent' stays with the model vendor becomes another question. If customers can switch between multiple models with low friction, pricing pressure on generic APIs will become increasingly evident. Of course, models are not without value. For tasks like coding, agents, and enterprise software automation, reliability, context length, tool-calling ability, and multi-step task completion rates directly impact business outcomes. If model companies can control the user interface, workflow memory, and data feedback loops, they still have a chance to retain the economics of high-value scenarios. The earliest monetizing applications are not general AI, but process-heavy, data-rich vertical scenarios Insurance and financial services risk, enterprise data integration, and cross-border marketing were among the strongest application signals in this research. Their commonality is clear: customers pay not for "using AI," but for reduced risk, improved efficiency, better marketing conversion, and decision automation. Such scenarios are more amenable to value-based pricing. As long as results are measurable, suppliers have an opportunity to convert AI capabilities into recurring revenue, rather than one-time project fees. However, traditional software valuation logic cannot be directly applied here. Some businesses may still have a strong service component: high concentration of large customers, small share of customer wallet, and early deployments requiring significant manual implementation or assistance. If these costs are fully accounted for, gross margins and scalability may not be as strong as they appear. Therefore, a more accurate assessment is: some AI applications have moved from narrative into early commercial validation but have not yet proven they possess durable, scalable software economics. The computing power theme becomes clearer: the more replaceable models are, the more infrastructure resembles a common winner If enterprises use multiple models, model exclusivity decreases, but demand for computing power, storage, memory, cloud orchestration, and inference infrastructure does not disappear. This is where the infrastructure logic is clearer. Regardless of which model ultimately leads, training, inference, enterprise deployment, ADAS development, simulation, data processing, and managed services will consume computing resources. Model competition may even increase the frequency of experimentation and deployment, thereby boosting demand for inference and supporting infrastructure. The most critical assumption is the elasticity of task volume. An optimistic scenario requires AI task volume growth to outpace the decline in cost per task. If model efficiency, sparse computing, edge inference, or architectural improvements cause unit computing power consumption to decline faster, infrastructure benefits would be diluted. The current framework leans more toward the former: demand is expanding from frontier training to inference, adaptation to domestic technology stacks, memory, storage, and enterprise workflow execution. For public markets, this remains the clearest AI exposure. Autonomous driving is no longer just a cost center, but L4 is not yet proven The change in autonomous driving is that the purely cash-burning narrative is beginning to loosen. ADAS and L4 robotaxis should be viewed separately. ADAS, relying on automaker mass production, real-world road data feedback loops, and the potential for software-like batch profit margins, is closer to scaling. Suppliers can follow the increase in vehicle assembly rates, forming more visible revenue contributions. L4 robotaxis remain much earlier stage. Some operators have cited city-level economic improvements, declining vehicle costs, and better overseas deployment economics, but company-level profitability remains a later-stage goal, and regulatory hurdles have not disappeared. Success in one city does not guarantee replication across different geographies, weather conditions, and regulatory environments. For L4 companies, evidence that can truly change investment viability is specific: auditable city-level unit economics, expansion of licensed cities, sustained decline in vehicle costs, safety records, and regulatory continuity. Any license suspension triggered by an accident could set back commercialization by several quarters. Neutral specialist companies can penetrate automakers and enterprises; platforms may not dominate the application layer Platforms, with their clouds, traffic, ecosystems, and distribution interfaces, seemed poised to capture most of the AI application value. However, in enterprise and automaker procurement, neutrality is becoming a variable. Some non-platform model or application companies are securing orders from automakers and enterprises not because they have larger ecosystems, but because customers view 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 specialist firms. This does not weaken the benefit logic for platforms like Tencent and Alibaba in cloud, maps, computing, and data infrastructure. What truly needs distinction is that a platform acting as an infrastructure supplier and a platform's own model securing orders in the application layer are two different things. When customers prioritize neutrality, customization, and deep integration, platform models do not necessarily have a natural advantage. Procurement decisions will rely more on trust, integration depth, and vertical delivery capabilities. Consumer AI and agent-driven e-commerce still lack hard proof The evidence for consumer AI is currently weak. Common issues include low user loyalty, intense competition, insufficient willingness to pay, and fast product imitation. Agent-driven e-commerce and AI advertising are also at an earlier stage. In cross-border marketing, current optimizations still stem more from recommendation algorithms; large language model-driven agent commercialization lacks sufficiently verifiable revenue evidence. The focus here should not be on downloads, product launches, or demo effects, but on retention rates, paid conversion rates, reuse rates, gross margins, and measurable transaction growth. Without these metrics, consumer AI narratives struggle to support higher valuations. What truly needs verification is retention, pricing, and margins, not the AI narrative itself For model companies, a key question in the coming quarters is: after customers adopt multi-model routing, can the model layer continue to capture value? An optimistic scenario requires seeing rising net retention, stable or improving pricing, low churn, and stable gross margins amid multi-model competition. Coding may be the clearest testing ground. It is high-frequency, value-quantifiable, and has potential for developer interface lock-in. Model-leading names like Zhipu AI and MiniMax cannot rely solely on benchmark performance; they must also prove vertical scenario control, workflow control, and repeatable quality of use. For application companies, core metrics are revenue quality, customer concentration, implementation intensity, net retention, pricing structure, and true gross margins after deducting customer support and manual assistance costs. For infrastructure companies, the key is whether growth can be more clearly attributed to inference, deployment, and domestic chip adaptation. If task volume expansion continues to outpace unit cost declines driven by efficiency gains, computing power, domestic chips, memory, and storage remain the clearest AI investment themes.
Comments