China AI value chain: China AI profit pool could expand sharply, with downstream models and applications becoming the main source of earnings growth by 2030E

Overview

JPMorgan initiates broader China AI coverage with a constructive view. Its central argument is that rapid growth in AI intelligence consumption expands the total profit pool across hardware, cloud and downstream models plus applications, but the mix of earnings should shift materially downstream after 2028E.

Core views

JPMorgan expects China inference-token consumption to grow about 60x between 2026E and 2030E as frontier models make more difficult tasks economically viable, falling inference costs broaden adoption, and longer reasoning, richer context and multi-step agents raise consumption per workflow. Enterprises are expected to account for about 69% of 2030E token demand, supporting recurring, workflow-embedded usage in coding, research, customer service and operations; consumers account for the remaining 31%. The report estimates global AI end-customer spending above US$1.4tn by 2030E, with China representing roughly one-quarter, including a China B2B opportunity of about US$250bn and consumer demand of about US$100bn. The report models China AI as three linked economic layers: hardware, cloud, and models plus applications. It starts with US$777.6bn of China cloud investment over 2026E-30E. Hardware recognizes revenue when capacity is purchased, cloud monetizes installed cohorts over five years, and downstream monetizes the inference capacity supplied by cloud. These are overlapping revenue claims on the same investment dollar and should not be added as end-customer spending; operating profit is the additive cross-layer measure. The approach constrains downstream revenue by physically available inference capacity rather than allocating a top-down TAM. Hardware remains the early-cycle earnings leader because it is paid first. JPMorgan forecasts hardware revenue of US$100.0bn in 2026E and US$184.8bn in 2030E, with operating profit rising from US$30.0bn to US$73.9bn and margins increasing from 30% to 40%. Hardware captures approximately US$286bn, or 53%, of the US$543.8bn cumulative 2026E-30E operating-profit pool. The report stresses that a later downstream shift does not make upstream value inappropriate today: hardware's large share reflects the timing of procurement earnings. Cloud is expected to compound as investment cohorts enter service and utilization ramps. The report forecasts cloud revenue increasing from US$20.0bn in 2026E to US$249.1bn in 2030E, or 12.5x, with operating profit rising from US$2.4bn to US$37.4bn. Its cloud model assumes utilization of 60%, 70% and 80% in revenue years one, two and three to five, respectively, and revenue yield rising from 0.50x of capex annually at full utilization in 2026E to 0.585x in 2030E. Sustainable cloud returns are a critical constraint: weak yields would slow investment, require higher utilization or pricing, or shift economics back toward infrastructure owners. Models plus applications are forecast to grow fastest, from US$15.7bn of revenue in 2026E to US$423.4bn in 2030E, or 26.9x. The report derives downstream revenue from inference spending and assumes gross margin rises from 30% to 50% as serving economics improve. The layer remains loss-making in 2026E and 2027E because training and model R&D absorb US$9.0bn in 2026E against only US$4.7bn of downstream gross profit. It turns profitable in 2028E, at US$6.2bn, as gross margin reaches about 40% and training falls to about 25% of downstream revenue. By 2030E, US$423.4bn of downstream revenue is projected to generate US$132.0bn of operating profit. This creates a pronounced profit-mix transition. Total China AI operating profit rises from US$26.5bn in 2026E to US$243.3bn in 2030E. Hardware's share falls from 93% to 30%, while models plus applications move from losses to about 54% of annual industry profit. Models plus applications contribute 32% of cumulative five-year operating profit but 64% of the approximately US$217bn increase in annual industry profit, versus about 20% from hardware and 16% from cloud. JPMorgan notes that about 77% of cumulative downstream profit arrives in 2030E, making funding capacity, cash burn and sensitivity to paid adoption, training intensity and gross margin especially important. The report does not characterize the current market-value mix as an obvious mispricing. Models plus applications represent about 14% of estimated China-AI-attributable listed market value, broadly matching their approximately 12% share of 2026E value-chain revenue. The question is whether the market will capitalize the projected move to a 49% downstream revenue share and approximately 54% profit share by 2030E before reported earnings reach their inflection. Evidence needed over the next 12-24 months includes cloud utilization, recurring API and subscription revenue, paid application users, gross profit after inference and delivery costs, and cash generation. JPMorgan applies an Opportunity × Scarcity × Value Capture framework to identify companies positioned to retain economics. Opportunity is the economic pool around a workload; scarcity is the capability that remains difficult to replicate; value capture tests whether monetization converts to gross profit, cash flow and returns. The report expects scarcity to migrate from raw capacity and frontier-model capability toward serving efficiency, proprietary context, permissions, workflow ownership and transaction execution. It also warns that application functionality can become easier to replicate as general models improve, so durable value depends on assets an agent still needs to complete the work. Alibaba is the preferred integrated infrastructure exposure because T-Head, cloud and its model franchise connect silicon, capacity and customer demand. JPMorgan's incremental-capacity reconstruction implies about 26.6% after-tax project IRR versus 10% for the normalized industry cohort, based on stronger revenue yield and faster utilization. Tencent is the preferred application exposure because distribution, proprietary context and transaction execution may retain value as model costs fall; the report believes these AI application assets are largely unpriced. Zhipu is the preferred differentiated-intelligence exposure, contingent on sustained relative capability supporting paid APIs, pricing, retention and gross profit sufficient to fund model development. Beyond the three preferred exposures, the report initiates SenseTime, Phancy and Xunce at Overweight. SenseTime and Phancy are viewed as ways to capture infrastructure optimization and deployment demand; Xunce is positioned to productize enterprise deployment through Token/TokenOS and reusable modules. Kingsoft Office is assumed at Overweight based on distribution and workflow ownership. MiniMax, JST, Meitu, Kingdee and Manycore are Neutral because monetization, valuation, growth expectations or feature defensibility require further proof. Yonyou is assumed at Underweight because its enterprise assets have yet to translate into comparable consolidated growth, profitability and cash generation.

Analysis framework

JPMorgan builds a bottom-up, linked value-chain model from cloud investment through hardware procurement, cloud utilization and inference-supported downstream revenue. It then assesses each company through Opportunity, Scarcity and Value Capture, using operating measures such as utilization, paid usage, pricing, recurring revenue, gross profit, retention, cash conversion and returns to test whether strategic positioning becomes financial value.

Methodology notes

• Industry AnalysisSupply-demand framework

  Three-layer AI value-chain model linking hardware, cloud and models plus applications.

  The report follows capacity investment through physical hardware, recurring cloud revenue and downstream inference consumption to estimate demand, revenue and profit by layer.

• Competition & strategyEconomic Moat and Competitive Advantage

  Opportunity × Scarcity × Value Capture framework.

  The framework asks how large a company's addressable workload is, what remains hard to substitute, and whether that advantage converts into monetization, gross profit, cash flow and returns.

• Corporate Fundamentals and FinanceFree cash flow analysis

  Financial conversion from revenue through gross profit, operating leverage, CFO/FCF and returns on incremental capital.

  The report treats usage or revenue as insufficient evidence unless it produces gross profit after AI costs and ultimately cash conversion.

• Valuation methodsP/E and PEG Valuation

  Company target prices use forward P/E multiples for several covered companies.

  The stock-summary table applies stated forward P/E multiples to forecast earnings for Alibaba, Tencent, Zhipu, Xunce, Kingsoft Office, MiniMax, JST, Kingdee and Yonyou.

• Valuation methodsEV/EBITDA valuation

  Company target prices use forward EV/EBITDA multiples for SenseTime and Phancy.

  The report applies 15x 2028E EV/EBITDA to the two infrastructure and deployment companies.

Asset mapping & comparison

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

• Alibaba (9988.HK)

  Preferred integrated hardware, cloud and model exposure.

  Strengths

  T-Head silicon, cloud capacity and model franchise enable model/infrastructure co-design.

  Weaknesses

  Cash conversion may lag project returns if investment remains elevated.

  Comparison

  Incremental after-tax project IRR of c.26.6% versus 10% normalized cohort.

  Risks

  Competitive cloud pricing, internal silicon without unit-economic gains, and delayed cash returns.

• Tencent (700.HK)

  Preferred AI application exposure.

  Strengths

  Distribution, proprietary context and transaction execution across consumer and commercial activity.

  Weaknesses

  AI engagement must convert into advertising, merchant or transaction revenue.

  Comparison

  JPMorgan considers its application assets largely unpriced within group valuation.

  Risks

  Inference and development costs may absorb incremental gross profit; limited disclosure could delay market recognition.

• Zhipu (2513.HK)

  Preferred differentiated-intelligence exposure.

  Strengths

  Relative model capability, paid API potential, serving optimization and proprietary harnesses.

  Weaknesses

  Economics depend on continued capability leadership and gross profit funding model development.

  Comparison

  Preferred model exposure; rated Overweight.

  Risks

  Model convergence, open-weight price competition and training costs rising faster than paid usage.

• SenseTime (20.HK)

  Covered infrastructure, deployment and multimodality exposure.

  Strengths

  SenseCore links capacity with production AI workloads.

  Weaknesses

  External adoption and scalable monetization remain key tests.

  Comparison

  Initiated at Overweight.

  Risks

  Utilization, external adoption, margin progression and returns may not scale as expected.

• Phancy (6682.HK)

  Covered compute-optimization and deployment exposure.

  Strengths

  Heterogeneous scheduling, adaptation and deployment address fragmented-chip complexity.

  Weaknesses

  Must turn capabilities into reusable software economics.

  Comparison

  Initiated at Overweight.

  Risks

  Productization, gross-margin improvement and deployment reuse may fall short.

• Xunce (3317.HK)

  Covered enterprise-deployment exposure.

  Strengths

  Enterprise context, governance, production integration, Token/TokenOS and reusable modules.

  Weaknesses

  Implementation-heavy work must become scalable recurring software.

  Comparison

  Initiated at Overweight.

  Risks

  Deployment time, revenue per employee, module reuse and cash collection may not improve.

• Kingsoft Office

  Covered productivity-application exposure.

  Strengths

  Distribution, document context and workflow ownership.

  Weaknesses

  AI paid conversion and incremental gross-profit delivery require validation.

  Comparison

  Assumed at Overweight.

  Risks

  AI monetization may not offset serving costs or generate expected workflow expansion.

• MiniMax (100.HK)

  Covered foundation-model exposure.

  Strengths

  Strong usage growth and multimodal and agentic positioning.

  Weaknesses

  Durable capability, pricing and gross-profit conversion remain unproven.

  Comparison

  Neutral while Zhipu is preferred.

  Risks

  Litigation with US studios, intensified competition, sustained R&D burden, uncertain commercialization and infrastructure dependence.

• JST (6687.HK)

  Covered commerce-workflow application exposure.

  Strengths

  Live orders, inventory, merchant rules and execution assets.

  Weaknesses

  Subdued e-commerce conditions, growth expectations and valuation offset workflow advantages.

  Comparison

  Assumed at Neutral.

  Risks

  AI workflow expansion may not overcome cycle and valuation constraints.

• Meitu (1357.HK)

  Covered visual-creation application exposure.

  Strengths

  Exposure to image, video, design, marketing and e-commerce creation workloads.

  Weaknesses

  Individual features are increasingly replicable by general models.

  Comparison

  Initiated at Neutral.

  Risks

  Incremental growth and monetization remain uncertain.

• Kingdee (268.HK)

  Covered enterprise-application exposure.

  Strengths

  Enterprise records, permissions and system-of-record position.

  Weaknesses

  AI-driven revenue acceleration and gross-profit conversion remain unproven.

  Comparison

  Assumed at Neutral.

  Risks

  AI attach, ACV, NRR and usage may not create clearer growth.

• Manycore (68.HK)

  Covered design and spatial-intelligence application exposure.

  Strengths

  Professional design workflows and proprietary 3D/spatial context.

  Weaknesses

  Commercialization remains early.

  Comparison

  Assumed at Neutral.

  Risks

  Core growth, enterprise NRR and new AI revenue need further evidence.

• Yonyou (600588.SH)

  Covered enterprise-application exposure.

  Strengths

  Enterprise state, permissions and complex workflows remain strategically valuable.

  Weaknesses

  Consolidated growth, profitability and cash generation lag the AI narrative.

  Comparison

  Assumed at Underweight.

  Risks

  AI product and contract activity may not translate into financial value capture.

Key data

• China AI token consumptionc.60x growth from 2026E to 2030EEnterprise demand is projected at approximately 69% of 2030E consumption.
• Global AI end-customer spending>US$1.4tn in 2030EChina is estimated to represent roughly one-quarter of demand.
• China AI cloud investmentUS$777.6bn over 2026E-30EAnnual investment rises from US$100.0bn in 2026E to US$184.8bn in 2030E.
• China AI operating profitUS$26.5bn in 2026E to US$243.3bn in 2030ECumulative 2026E-30E operating profit is US$543.8bn.
• Downstream revenueUS$15.7bn in 2026E to US$423.4bn in 2030E26.9x growth, supported by a gross-margin rise from 30% to 50%.
• Downstream operating profitUS$(5.9)bn in 2026E to US$132.0bn in 2030ETurns positive in 2028E at US$6.2bn.
• Downstream contribution32% of cumulative profit and 64% of annual-profit growthModels plus applications are projected to represent about 54% of 2030E industry operating profit.
• Alibaba incremental project IRRc.26.6% after taxVersus 10% for JPMorgan's normalized industry cohort.

Impact & implications

JPMorgan expects all AI layers to grow, but distinguishes hardware's more visible near-term earnings from cloud and downstream companies' larger medium- to long-term earnings trajectory. It argues that stock selection should focus on durable scarcity, proof of paid adoption and gross-profit conversion, funding through the pre-2028E earnings ramp, and valuation.

Risks

• China intelligence-consumption growth could be slower than JPMorgan expects.
• Company-specific scarcity could erode as capacity expands, model capabilities converge or application features become easier to replicate.
• AI adoption may not convert into gross profit and cash flow after model, inference, training, deployment and support costs.
• Cloud revenue yield or utilization could remain insufficient to sustain economically attractive capacity investment.
• Downstream companies may require external funding through the pre-2028E earnings ramp.
• Terminal-year outcomes are sensitive to paid adoption, gross margin and training intensity because most modeled downstream profit arrives in 2030E.

What to watch

• Token-consumption growth, especially enterprise adoption and agent workload intensity.
• Cloud utilization, revenue yield, serving efficiency and returns on incremental capacity.
• Paid API, subscription, usage and transaction revenue rather than capability announcements.
• Gross margin after inference and delivery costs, operating leverage and cash conversion.
• Evidence that applications retain proprietary context, permissions, workflow ownership and transaction execution as agents become user interfaces.
• Company-specific indicators including Alibaba cloud EBITA and free cash flow, Tencent incremental commercial profit, Zhipu paid API growth, and Xunce module reuse and deployment efficiency.