China AI value chain: J.P. Morgan expects China AI profits to expand sharply, with value progressively shifting from hardware toward cloud and downstream monetization

Overview

This is a China AI value-chain initiation that models how expanding intelligence consumption translates into revenue and operating profit for hardware, cloud, and models plus applications. J.P. Morgan is constructive on the sector, arguing that hardware captures early buildout economics while cloud and downstream businesses should account for an increasing share of earnings growth as utilization, paid usage and gross margins improve.

Core views

J.P. Morgan expects China’s AI value chain to expand as better model capability, lower inference costs, longer reasoning chains, richer context and multi-step agents drive about 60x growth in inference-token consumption between 2026E and 2030E. Enterprises are expected to account for approximately 69% of 2030E token consumption, supported by recurring workflow use cases such as coding, research, customer service and operations; consumers provide the remaining 31% through creation, productivity and personalized services. The report estimates more than US$1.4tn of global AI end-customer spending by 2030E, with China representing roughly one-quarter, and frames that demand as extending beyond traditional software budgets toward labor, service and workflow economics. The report separates the value chain into hardware, cloud, and models plus applications, while emphasizing that layer revenues overlap and should not be added as end-customer spending. Its bottom-up model follows one dollar of cloud investment through three revenue-recognition cycles: hardware earns when capacity is procured, cloud earns recurring revenue over the five-year service life of that capacity, and downstream companies monetize the intelligence and workflows served by the installed base. China hyperscaler AI investment is modeled at US$777.6bn cumulatively over 2026E-30E, rising from US$100.0bn in 2026E to US$184.8bn in 2029E and 2030E. Hardware revenue therefore equals annual cloud investment and reaches the same cumulative US$777.6bn; hardware operating margin rises from 30% to 40% over the period. Hardware earns first because suppliers are paid at procurement. It is forecast to generate US$286.1bn, or approximately 53%, of the US$543.8bn cumulative 2026E-30E industry operating-profit pool. Hardware operating profit rises from US$30.0bn in 2026E to US$73.9bn in 2030E, but its share of annual industry profit falls from 93% to 30% as cloud and downstream monetization scale. J.P. Morgan does not characterize the current upstream concentration as a mispricing by itself: it reflects the fact that hardware recognizes meaningful economics during the capacity-buildout phase. Continued training, inference and agent workloads still support a large absolute hardware profit pool, even as incremental value shifts elsewhere. Cloud is the recurring monetization bridge between installed capacity and intelligence consumption. The model assumes each capex cohort generates cloud revenue for five years starting the year after procurement, with utilization ramping from 60% in the first revenue year to 80% in years three through five. Full-utilization revenue yield rises from 0.50x annual capex in 2026E by 4% annually to 0.585x by 2030E. Cloud revenue is therefore projected to grow from US$20.0bn in 2026E to US$249.1bn in 2030E, or 12.5x, as multiple investment cohorts operate simultaneously. Cloud EBITA margin is assumed at 12% in 2026E and 15% thereafter, yielding operating profit growth from US$2.4bn to US$37.4bn. Cloud returns remain a key system constraint: weak yield, utilization or serving efficiency could curb future capacity expansion, while attractive returns permit more investment and a larger downstream pool. Models plus applications are the report’s main source of medium-term earnings growth but also its most back-ended and sensitive layer. The cloud revenue mix shifts from 45% training and model R&D / 55% inference in 2026E to 15% / 85% by 2030E. Because inference spend is downstream cost of goods sold, downstream revenue is constrained by the capacity that cloud can physically serve. With downstream gross margin assumed to improve from 30% to 50%, revenue rises from US$15.7bn in 2026E to US$423.4bn in 2030E, or 26.9x. Training and model R&D absorb US$9.0bn in 2026E against downstream gross profit of only US$4.7bn, producing a US$5.9bn operating loss. The layer is projected to turn profitable in 2028E, with US$6.2bn of operating profit and roughly 40% gross margin, before reaching US$132.0bn of operating profit in 2030E on US$423.4bn of revenue and a 50% gross margin. The resulting profit mix changes materially. Total annual operating profit rises from US$26.5bn in 2026E to US$243.3bn in 2030E; hardware grows 2.5x, cloud 15.6x, and downstream moves from loss-making to 54% of annual industry profit. Downstream contributes approximately 32% of cumulative five-year operating profit but 64% of the US$217bn increase in annual industry profit between 2026E and 2030E. About 77% of modeled downstream cumulative profit arrives in 2030E, making funding capacity, cash burn and access to capital important selection criteria through the pre-inflection period. J.P. Morgan explicitly treats the 2030E pool as a scenario outcome rather than a precise point forecast because modest changes in gross margin, training intensity or paid-adoption pace can materially alter the outcome. J.P. Morgan argues that current market value broadly reflects today’s revenue distribution rather than the modeled earnings transition. Models plus applications represent approximately 14% of its estimate of China-AI-attributable listed market value, broadly aligned with the layer’s approximately 12% share of 2026E value-chain revenue. Its forecast, however, has downstream revenue share rising to approximately 49% by 2030E and downstream operating-profit share reaching approximately 54%. The report therefore focuses on whether paid usage, gross profit and cash generation provide evidence that the later profit shift will become visible to the market. The relevant near-term indicators are cloud utilization, recurring API and subscription revenue, paid application users, gross margin after inference and delivery costs, operating leverage and cash flow. The stock-selection framework is Opportunity × Scarcity × Value Capture. Opportunity measures the economic pool around a company’s workloads; Scarcity tests which capabilities remain hard to replicate or substitute; and Value Capture asks whether those advantages turn into customer spending, gross profit, cash conversion and returns. The report expects scarcity to migrate over time from raw capacity and frontier model intelligence toward serving efficiency, proprietary enterprise context, permissions, workflow ownership and transaction execution. It also cautions that scarce assets do not automatically create equity value: companies must demonstrate monetization and retained economics after paying for compute, inference, deployment and other stack layers. Alibaba, Tencent and Zhipu are J.P. Morgan’s preferred exposures because they occupy different points in the modeled sequence. Alibaba combines T-Head silicon, cloud capacity and a model franchise, and J.P. Morgan’s incremental-capacity reconstruction estimates a 26.6% after-tax project IRR versus 10% for the normalized industry cohort, contingent on stronger revenue yield and faster utilization. Tencent is the preferred application exposure because its distribution, proprietary context and transaction execution may remain necessary even as general models absorb more interface functionality; the report believes these AI application assets are largely unpriced. Zhipu is the preferred differentiated-intelligence exposure, with the central test being whether relative capability sustains API pricing, paid calls, retention and gross profit sufficient to fund continued model development. The report initiates SenseTime, Phancy and Xunce at Overweight and assumes Kingsoft Office at Overweight. It initiates or assumes Neutral ratings on MiniMax, JST, Meitu, Kingdee and Manycore, citing varying combinations of uncertain monetization, feature substitutability, early commercialization, valuation and growth constraints. It assumes Yonyou at Underweight because its enterprise workflow assets have not yet translated into comparable consolidated growth, profitability and cash generation. Sector-wide risks are slower-than-expected intelligence-consumption growth, erosion of company-specific scarcity, and weaker conversion of adoption into gross profit and cash flow.

Analysis framework

J.P. Morgan starts with projected token consumption and end-customer AI spending, then builds a bottom-up value-chain model from cloud investment. It traces the same investment through hardware procurement, five-year cloud revenue cohorts, and downstream revenue constrained by inference capacity. It then evaluates companies using Opportunity × Scarcity × Value Capture, testing strategic positioning against observable indicators such as utilization, recurring revenue, paid adoption, gross profit, retention, cash flow and incremental returns.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

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

  The report follows cloud investment through hardware procurement, recurring cloud monetization and downstream intelligence revenue to show how economics are shared across linked layers.

• Industry AnalysisSupply-demand framework

  Token-consumption and inference-capacity demand model.

  Projected growth in AI workloads, utilization and inference demand drives the report’s estimates for capacity needs, cloud revenue and downstream monetization.

• Competition & strategyEconomic Moat and Competitive Advantage

  Opportunity × Scarcity × Value Capture framework.

  The framework assesses the size of a company’s AI opportunity, the durability of assets that are hard to replicate, and whether those assets translate into revenue, gross profit, cash flow and returns.

• Corporate Fundamentals and FinanceFree cash flow analysis

  Financial-conversion sequence from revenue to gross profit, operating leverage, CFO/FCF and incremental returns.

  The report treats top-line AI adoption as insufficient evidence unless it produces gross profit after inference and deployment costs, cash conversion and acceptable returns on investment.

• Valuation methodsP/E and PEG Valuation

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

  The stock summary applies forward P/E multiples to forecast earnings for Alibaba, Tencent, Zhipu and several other covered companies.

Asset mapping & comparison

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

• Alibaba (BABA / 9988.HK)

  Preferred integrated exposure across hardware, cloud and models plus applications.

  Strengths

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

  Weaknesses

  Cash conversion may be deferred by a rising investment base.

  Comparison

  J.P. Morgan estimates c.26.6% after-tax incremental project IRR versus 10% for the normalized industry cohort.

  Risks

  Competitive cloud pricing, internal silicon failing to improve unit economics, and sustained investment delaying cash returns.

• Tencent (700.HK)

  Preferred application exposure.

  Strengths

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

  Weaknesses

  AI contribution may remain difficult to identify in reported group results.

  Comparison

  The report considers Tencent’s AI application assets largely unpriced relative to the broader China-AI-attributable listed market.

  Risks

  AI engagement may not translate into advertising, merchant or transaction revenue; development and inference costs may absorb gross profit.

• Zhipu (2513.HK)

  Preferred differentiated-intelligence exposure; Overweight.

  Strengths

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

  Weaknesses

  Economics depend on sustained relative capability and gross profit sufficient to fund continued model development.

  Comparison

  The report identifies Zhipu as its preferred China model exposure, while MiniMax remains Neutral pending stronger evidence of capability, pricing and gross-profit conversion.

  Risks

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

• SenseTime (20.HK)

  Initiated at Overweight as an infrastructure, deployment and multimodality exposure.

  Strengths

  SenseCore links capacity with production AI workloads.

  Weaknesses

  Long-term value depends on generalizing capabilities beyond internal workloads.

  Comparison

  External adoption, utilization and margin progression are the central validation metrics.

  Risks

  Technical improvements may not scale into external adoption, reusable revenue or acceptable returns.

• Phancy (6682.HK)

  Initiated at Overweight as a heterogeneous-compute optimization and deployment exposure.

  Strengths

  Scheduling, model adaptation and deployment address fragmented accelerator environments.

  Weaknesses

  Productization and operating conversion remain key tests.

  Comparison

  The report sees potential for economics to shift from capacity-heavy deployment toward optimization-heavy reusable software.

  Risks

  Capabilities may not convert into gross-margin expansion, deployment efficiency or customer reuse.

• Xunce Technology (3317.HK)

  Initiated at Overweight as an enterprise AI deployment and productization exposure.

  Strengths

  Enterprise context, governance, integration and Token/TokenOS module reuse.

  Weaknesses

  Implementation-heavy delivery must become scalable software economics.

  Comparison

  Module reuse, deployment time, revenue per employee and cash collection are key tests.

  Risks

  Enterprise deployment demand may remain labor-intensive rather than generating recurring software and usage revenue.

• Kingsoft Office (688111.CH)

  Assumed at Overweight as a productivity-application exposure.

  Strengths

  Large user base, distribution, document context and workflow ownership.

  Weaknesses

  AI monetization must exceed AI-serving costs.

  Comparison

  The report views access to increasingly interchangeable intelligence as able to expand paid productivity use while lowering serving cost.

  Risks

  Paid conversion, incremental gross profit and broader workflow penetration may not develop as expected.

• MiniMax (100.HK)

  Neutral foundation-model exposure.

  Strengths

  Strong usage growth and multimodal and agentic positioning.

  Weaknesses

  Durable capability, pricing power and gross-profit conversion require further proof.

  Comparison

  Unlike Overweight-rated Zhipu, the report views MiniMax’s evidence as more weighted toward usage growth than differentiated financial conversion.

  Risks

  Litigation proceedings with US studios, competition, sustained R&D investment, commercialization uncertainty, and computing-infrastructure and supplier dependence.

• Yonyou Network Technology (600588.CH)

  Assumed at Underweight as an enterprise-application exposure.

  Strengths

  Enterprise state, permissions and complex workflow assets.

  Weaknesses

  Consolidated growth, profitability and cash generation have not shown comparable value capture.

  Comparison

  The report sees a sizable AI opportunity but weaker financial conversion than required for its investment case.

  Risks

  AI product and contract activity may continue to fail to translate into revenue, margin and cash conversion.

Key data

• China AI token-consumption growthc.60x from 2026E to 2030EDriven by broader adoption, improved capability, lower inference costs and more complex agentic workflows.
• Enterprise share of 2030E China token consumptionc.69%Consumers account for the remaining c.31%.
• Global AI end-customer spending>US$1.4tn in 2030EChina is estimated to represent roughly one-quarter of demand.
• China AI hyperscaler investmentUS$777.6bn cumulatively in 2026E-30EModeled from US$100.0bn in 2026E to US$184.8bn in 2029E and 2030E.
• China AI industry 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$423.4bn in 2030EA 26.9x increase, supported by a larger inference pool and gross-margin expansion from 30% to 50%.
• Downstream operating profitUS$(5.9)bn in 2026E to US$132.0bn in 2030EThe model projects the first positive downstream operating profit in 2028E at US$6.2bn.
• Alibaba incremental-capacity project IRRc.26.6% after taxVersus 10% for the normalized industry cohort in J.P. Morgan’s reconstruction.

Impact & implications

J.P. Morgan expects all three layers to grow, but distinguishes early hardware earnings visibility from the larger medium-term earnings trajectory in cloud and models plus applications. It argues that the key investment debate is not simply whether AI spending rises, but whether capacity earns acceptable returns, falling inference costs expand downstream gross profit, and individual companies retain scarce charging points and convert usage into cash flow.

Risks

• Intelligence consumption could grow more slowly than J.P. Morgan expects.
• Company-specific sources of scarcity could erode as capacity expands and model capabilities become more interchangeable.
• AI adoption may convert weakly into gross profit and cash flow after inference, development, deployment and support costs.
• Cloud revenue yield, utilization or serving efficiency could remain inadequate, limiting economically sustainable capacity investment.
• Downstream forecasts are sensitive to gross margin, training intensity and the pace of paid adoption, with much of modeled profit concentrated in 2030E.

What to watch

• Growth in cloud utilization, serving efficiency, revenue yield and cloud segment margins.
• Recurring API and subscription revenue, paid application users, AI paid conversion and usage-linked revenue.
• Gross margin after inference and delivery costs, operating leverage, cash flow and returns on incremental capital.
• For models, relative capability, pricing, token and ARR growth, official-endpoint economics and harness adoption.
• For middleware, module reuse, deployment time, revenue per employee, cross-industry replication and cash collection.
• For applications, retention, NRR, transaction or action volume, AI-related ARPU or ACV, and whether machine-generated usage offsets seat compression.