China AI application research shows early evidence of workflow-driven monetization

Overview

This report summarizes the key takeaways from J.P. Morgan’s China AI application research tour. The research covered autonomous driving, independent model development, enterprise workflow software, and vertical AI applications. The core conclusion is that China’s AI application landscape is showing early evidence of workflow-driven monetization. Foundation models remain important, but in some enterprise scenarios they increasingly look like rentable and routable inputs, while true value capture depends more and more on proprietary data, business process control, customer integration, deployment capabilities, and pricing power.

Core views

The report’s core views include: first, model access may become more replaceable in some enterprise use cases, and customers are more willing to pay for workflow layers that can be embedded into business processes and complete tasks; second, vertical AI applications in insurance and financial risk, enterprise data integration, cross-border marketing, and similar areas are showing signals of recurring revenue and value-based pricing; third, ADAS is closer to scale than L4 robotaxi, while L4 has shown early commercialization potential in some cities and overseas deployments but remains constrained by regulation and unit economics; fourth, neutral specialist vendors can win in some OEM and enterprise procurement scenarios, while platforms can still benefit through cloud, maps, computing power, and infrastructure; fifth, AI infrastructure demand is expanding from frontier training into inference, domestic stack adaptation, memory, storage, and enterprise workflow execution; sixth, consumer AI and agentic e-commerce still lack sufficient valuation evidence.

Analysis framework

The report uses a combination of research meeting notes and investment mapping. It first distills management commentary and business signals from eight company interactions into six observations, and then compares them with prior assumptions about model capability, application monetization, infrastructure, and autonomous driving economics. The focus of the analysis is not to restate non-public numbers, but to judge which signals can be translated into an investment framework for listed companies and which variables still need confirmation through financial statements, customer evidence, segment disclosures, and third-party verifiable data.

Methodology notes

• Commercialization frameworkOwn the scenario, rent the model

  In enterprise AI, model capability is a necessary input, but proprietary data, workflow control, user interface, feedback loops, and deployment experience determine long-term value capture.

  This framework explains why enterprise customers can route tasks across multiple models based on price, performance, and availability, while still paying the application and workflow layers that can be embedded into business processes.

• Stage classificationNarrative, pilot, early commercialization, scale

  The report distinguishes different AI use cases by commercialization maturity rather than treating all AI applications as being at the same stage.

  Vertical and enterprise AI applications are seen as lying between early commercialization and scale, ADAS is closer to scale, L4 robotaxi is in early commercialization in some cities but broad deployment is still at the pilot stage, while consumer AI and agentic e-commerce remain more narrative-driven.

• Infrastructure demandTask volume elasticity assumption

  The bullish infrastructure thesis depends on whether growth in AI task volume outpaces the decline in per-task cost.

  If the expansion in task volume driven by inference, deployment, and enterprise workflow execution exceeds the cost compression caused by model efficiency gains, demand for computing power, cloud, memory, storage, and domestic chips will be better supported.

• Autonomous driving assessmentLayered assessment of ADAS and L4 robotaxi

  ADAS and L4 robotaxi have different commercialization paths, regulatory constraints, and investability, and cannot be evaluated with the same valuation logic.

  ADAS benefits from OEM mass production, real-road data closed loops, and scalable software margins; L4 robotaxi requires auditable city-level contribution profit, license expansion, vehicle cost reduction, safety records, and regulatory continuity.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AI infrastructure, computing power, domestic chips, memory, and storage

  Clearest beneficiary direction

  Strengths

  Benefits from deployment expansion, inference growth, multi-model routing, domestic stack adaptation, and enterprise workflow execution, without needing to accurately predict the eventual winning foundation model.

  Weaknesses

  The investment thesis depends on continued growth in task volume and utilization; if capacity loosens or efficiency improves too quickly, pricing and demand elasticity may come under pressure.

  Comparison

  Compared with the model layer, infrastructure is more model-agnostic; compared with the application layer, revenue mapping is more direct but also more affected by supply-demand cycles.

  Risks

  If model architecture efficiency, sparsification, edge inference, or declining per-task costs outpace task volume expansion, this could weaken cloud, chip, memory, and storage indicators.

• Workflow-layer and vertical AI applications

  Selectively positive

  Strengths

  In data-intensive scenarios such as insurance and financial risk, enterprise data integration, cross-border marketing, and operations automation, they are showing signals of recurring revenue, value-based pricing, and operating-level profitability.

  Weaknesses

  Some companies have high customer concentration, wallet share from major clients remains small, implementation and human-assisted components may still be heavy, and profit quality may not be fully software-like.

  Comparison

  Compared with foundation models, the workflow layer is closer to customer business outcomes; compared with infrastructure, scalability and gross margin require more case-by-case validation.

  Risks

  Enterprises or platforms may build workflow orchestration capabilities in-house, and customers may renegotiate once model costs decline, leading to margin compression at the application layer.

• Foundation models and model-led companies

  View is more nuanced and requires proof of vertical defensibility

  Strengths

  In high-value workflows such as coding, agents, and enterprise software automation, model quality, reliability, context length, tool use, and multi-step task completion rates can still affect customers’ willingness to pay.

  Weaknesses

  Multi-model routing and low switching costs for raw APIs may weaken pricing power for general model access, and simply leading on benchmarks is not enough to prove a long-term moat.

  Comparison

  Compared with the application layer, the model layer can more easily demonstrate usage growth; but compared with workflow companies that control the scenario, retention and pricing power need more disclosure-based validation.

  Risks

  If usage growth comes with gross margin compression, customers route tasks by price, and model vendors cannot demonstrate vertical retention or workflow control, the valuation thesis will weaken.

• ADAS and L4 robotaxi

  Selectively constructive

  Strengths

  ADAS is closer to mass production and OEM installation, with real-road data loops and the potential for scalable software margins; some L4 operators claim improved city-level economics and pricing in overseas deployments.

  Weaknesses

  Profitability at the L4 robotaxi company level remains distant, and regulatory barriers, vehicle costs, safety records, and city replication capability remain key uncertainties.

  Comparison

  ADAS is easier to underwrite than L4 robotaxi because its revenue contribution and supplier economics are closer to visible mass production.

  Risks

  Accidents could lead to license suspensions lasting multiple quarters, vehicle cost reduction could stall, or city-level contribution profit may fail to generalize across regions.

• Large platforms, cloud, maps, and distribution infrastructure

  Mildly positive as infrastructure suppliers, more selective on platforms’ own models

  Strengths

  Platforms such as Tencent and Alibaba can still benefit through cloud, maps, computing power, data infrastructure, and enterprise distribution, especially when neutral specialist vendors also consume platform infrastructure.

  Weaknesses

  In OEM and enterprise procurement, customers may prefer neutral, customizable specialist model or application vendors that are not tied to competing ecosystems.

  Comparison

  The platform value of infrastructure and distribution remains intact, but platforms’ own models do not necessarily have a superior share in the application layer relative to neutral specialist vendors.

  Risks

  If platforms regain enterprise entry points by subsidizing model access, bundling cloud and application tools, or leveraging distribution, neutral application vendors will face pricing pressure.

• Consumer AI and agentic e-commerce

  Continue to apply a discount

  Strengths

  The long-term opportunity could be large, and valuation upside may still open up if high retention, paid conversion, and transaction uplift can be proven.

  Weaknesses

  Most companies currently avoid positioning consumer AI as a near-term business model because of low loyalty, intense competition, weak willingness to pay, and rapid product imitation.

  Comparison

  Compared with enterprise and vertical AI applications, consumer AI currently lacks underwritable evidence of revenue quality and gross margin.

  Risks

  If they can rely only on downloads, demos, or product launches without proving repurchase, gross margin, and transaction uplift, valuation credit for listed companies will remain limited.

Key data

• Research coverage8 meetingsCovered autonomous driving, independent model development, enterprise workflow software, and vertical AI applications.
• Core commercialization signalsRecurring revenue, value-based pricing, operating-level profitabilityMainly from vertical areas such as insurance and financial services risk, enterprise data integration, and cross-border marketing, but most figures are undisclosed and require verification.
• Model-layer observationLow switching costs for raw APIs in some enterprise scenariosCustomer stickiness comes more from enterprise data integration, workflow control, and deployment depth than from pure access to foundation models.
• Autonomous driving stageADAS is close to scale; L4 robotaxi is in early commercialization or pilot stage in some cities and overseas deploymentsBroader valuation recognition requires third-party verifiable unit economics, license expansion, vehicle cost reduction, and safety records.
• Infrastructure mappingComputing power, cloud orchestration, domestic chips, memory, storage, and inference infrastructureDeployment-driven and model-agnostic demand makes infrastructure the clearest beneficiary direction among listed companies in the current evidence set.
• Consumer AI stageStill at the narrative stageCurrent evidence is insufficient to support significant valuation credit for listed companies; retention, paid conversion, repurchase, gross margin, and transaction uplift need to be monitored.
• Company prices and ratings involved9988.HK HK$126.00/OW; BABA $131.47/OW; 0100.HK HK$663.00/OW; 0700.HK HK$439.00/OW; 2513.HK HK$1010.00/OWThe prices listed in the report are all closing prices as of 2026-05-21.

Impact & implications

The investment implication is to expand the AI investment framework from 'who owns the strongest foundation model' to 'who can own the scenario, workflow, and data.' Infrastructure remains the clearest mapping to listed companies, because multi-model routing, inference growth, and domestic stack adaptation will continue to consume computing power, memory, storage, and cloud resources. Application-layer opportunities are more concentrated in vertical workflow companies that can convert model capability into customer value. Model-led companies need to prove retention, pricing power, gross margin, and workflow control, while autonomous driving needs auditable unit economics and regulatory continuity to support valuations.

Risks

• Key data on profitability, margins, and break-even are mostly based on management commentary, and some come from private companies, so they may not be verifiable in audited financial statements, prospectuses, or listed company segment disclosures.
• Scenario and workflow moats may be less durable than expected, as enterprises and platforms may build orchestration capabilities in-house or compare vendors more aggressively and renegotiate once inference costs fall.
• Infrastructure demand elasticity may be weaker than the research suggests; if model efficiency, sparsification, edge inference, or architectural improvements reduce per-task computing intensity faster than deployment expands, related beneficiaries may underperform expectations.
• Robotaxi remains constrained by regulation, and even if city-level economics improve, it may still be difficult to replicate across regions because of differences in geography, weather, safety standards, and regulatory regimes.
• The economics of vertical AI applications may be more service-like than they appear, as manual implementation, human-assisted workflows, customer concentration, or low wallet share may weaken software-like scaling ability.
• Platforms may once again strengthen their distribution and bundling advantages, putting neutral specialist vendors under pressure in procurement share and pricing.

What to watch

• Whether listed cloud and hardware companies disclose more clearly the contribution of inference, deployment, and domestic chip mix to growth.
• Whether model companies can demonstrate improving net retention, stable or improving pricing, low churn, and maintained gross margins in a multi-model routing environment.
• Whether coding, agents, and enterprise software automation become defensible high-frequency workflow verticals for model companies.
• The revenue quality, customer concentration, implementation intensity, pricing structure, net retention, and gross margin after excluding human support for vertical AI application companies.
• ADAS OEM installation rates, supplier revenue contribution, mass-production cadence, and software margins.
• L4 robotaxi city-level contribution profit, expansion into licensed cities, vehicle cost reduction, safety records, and regulatory continuity.
• Retention, paid conversion, repurchase, gross margin, and measurable transaction uplift for consumer AI and agentic e-commerce.
• Cloud inference growth, OEM design wins, enterprise procurement share, and the breakdown between infrastructure revenue and model or application revenue for platforms such as Tencent and Alibaba.