China’s AI sector has entered a new phase of value capture: shifting from training to inference, and from potential to profitability.

Overview

Morgan Stanley’s in-depth report highlights that China’s artificial intelligence industry has officially entered Phase 2.0. The defining characteristic of this phase is no longer merely catching up on technological capabilities, but rather leveraging speed, cost efficiency, and system-level integration to rapidly translate AI capabilities into tangible economic value and corporate profits. The report underscores that the focus of AI development has shifted from model training to inference, and from the lab to the real economy. Although, in the short term, AI’s contribution to macroeconomic growth may be constrained by labor‑adjustment frictions, in the medium to long term it will serve as a critical lever for mitigating demographic aging and boosting total factor productivity. From an investment perspective, the firm recommends focusing on power infrastructure, semiconductor localization, and leading application‑layer companies with clear monetization pathways.

Core views

Macro Perspective: AI is a medium- to long-term productivity driver, not a short-term cyclical boost. Our research estimates that over the next decade, AI could cumulatively raise China’s total factor productivity by roughly 3 percentage points, lifting potential GDP in 2035 by an additional 3.5 percentage points compared with a no‑AI baseline. However, in the near term (2026–2027), AI’s net contribution to headline GDP growth may be modest, or even slightly negative. While AI‑related capital expenditure cycles—such as data center construction—are expected to add 0.2–0.3 percentage points to growth, this boost will largely be offset by labor market disruptions during the transition. Given low profit margins among Chinese firms, companies are likely to prioritize cost cutting over output expansion, leading to faster displacement of white‑collar and certain service‑sector jobs than new job creation, thereby intensifying deflationary pressures. Industry Trend 1: Electricity has become the new bottleneck, with energy storage emerging as the next growth frontier. As AI workloads shift from training to high‑frequency inference, data center power demand is rising exponentially, and the key constraint has shifted from computing capacity itself to the stability and flexibility of electricity supply. We project that global data centers’ annual incremental deployment of energy storage systems (ESS) will surge from its 2025 baseline, reaching approximately 321 GWh by 2030. In addition to energy storage, gas turbines and high‑voltage grid equipment such as transformers will also benefit from the ongoing expansion of global power systems. Chinese equipment exporters enjoy a distinct competitive edge in this space. Industry Trend 2: Semiconductor self‑sufficiency is advancing, with competition shifting toward cost‑performance. Affected by export controls, China’s AI chip market is accelerating its localization. By 2030, the domestic AI chip market is projected to reach $67 billion, with local supply penetration climbing from 41% in 2025 to 86% by 2030. Although gaps remain in advanced process nodes, domestic manufacturers have delivered highly competitive total cost of ownership (TCO) for inference workloads through system‑level optimization, multi‑chip packaging, and software ecosystem adaptation. Procurement decisions are pivoting from chasing peak theoretical performance to emphasizing deployable cost efficiency and supply chain resilience. Industry Trend 3: Embodied intelligence and autonomous driving are gaining rapid traction. In the humanoid robotics sector, China holds a leading position, leveraging its manufacturing base and advantages in real‑world data. Deliveries of Chinese humanoid robots are expected to double to 28,000 units in 2026, and following the maturation of model capabilities after 2028, large‑scale adoption will commence, unlocking substantial long‑term market potential. In autonomous driving, L2+ features are rapidly becoming standard, with penetration set to rise from 25% in 2025 to over 50% by 2030; meanwhile, the commercialization of L4 robotaxis is accelerating under policy support, and by 2030 they are expected to account for roughly 8% of China’s ride‑hailing fleet. Internet and Software: Full‑stack capabilities build moats. Our analysis suggests that companies with end‑to‑end “chip‑cloud infrastructure‑foundation model‑application” capabilities will prevail. Alibaba is regarded as the best‑positioned full‑stack AI platform, with Alibaba Cloud and the Tongyi Qianwen model at the forefront. In contrast, pure‑play software and SaaS companies face the risk of disruption by AI agents, putting their valuations under pressure, as AI lowers the barrier to personalized software development and may erode the terminal value of traditional software.

Analysis framework

The research report employs an analytical framework that integrates top-down macroeconomic projections with bottom-up industry‑level validation. At the macro level, analysts deploy a total factor productivity (TFP) model to quantify AI’s contribution to long‑term potential GDP, while simultaneously examining short‑term frictional costs through the lens of labor market dynamics—such as youth unemployment and pressures on the middle‑income cohort. This analysis underscores AI’s quintessential J‑curve characteristics: high upfront investment and adjustment costs in the near term, coupled with substantial long‑term returns. At the industry level, the report constructs the “China AI 65” equity universe, leveraging a proprietary Global AI Mapping Survey to track firms’ AI exposure and its financial materiality. Beyond static AI themes, analysts place particular emphasis on the rate of change—identifying which companies are seeing accelerating improvements in both their AI relevance and earnings expectations. By comparing the financial performance of enablers (e.g., semiconductors, power) with adopters (e.g., application‑tier firms), they find that adopters’ margin expansion is outpacing revenue growth, confirming that, in the early stages, AI primarily delivers efficiency‑driven gains. Finally, in specific sector selection, the report applies a supply‑demand framework and bottleneck analysis. For instance, in the power sector, it identifies a shift in demand from “power availability” to “power flexibility,” thereby highlighting opportunities in energy storage and grid infrastructure. In semiconductors, by analyzing supply chain restructuring amid geopolitical constraints, the report zeroes in on equipment vendors and chip design firms with compelling domestic substitution narratives.

Methodology notes

• Macroeconomic frameworkOthers

  Total Factor Productivity (TFP) and Potential GDP Estimation

  The research report employs the TFP model to assess the extent to which technological progress (AI) enhances an economy’s long-term productive capacity. In simpler terms, it quantifies how much additional GDP can be generated solely through improvements in technical efficiency, holding labor and capital inputs constant. This approach helps readers understand why AI represents a long-term positive driver while delivering only gradual short-term results.

• Industry/Industrial Analysis FrameworkSupply-and-Demand Framework

  Bottleneck Shift Analysis (From Computing Power to Electricity)

  The research report points out that the bottleneck in industrial development has shifted from “whether or not we have chips” to “whether or not we have sufficient electricity.” When a particular link—such as power supply—becomes the limiting factor constraining overall expansion, investment opportunities in that segment are often the most substantial. This reflects the quintessential logic of supply-side bottleneck analysis.

• Company Fundamentals and Financial FrameworkProfit Quality Analysis

  Profit Margin Expansion vs Revenue Growth

  The research report distinguishes between two types of financial impacts driven by AI: first, direct revenue growth; and second, cost reduction and margin expansion. Data indicate that, at present, Chinese AI adopters are deriving greater benefits from the latter, suggesting that investors should prioritize companies that can leverage AI to achieve substantial cost efficiencies and productivity gains, rather than focusing solely on revenue growth rates.

Asset mapping & comparison

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

• Alibaba (BABA.N)

  A best-in-class full-stack AI platform, featuring a self-developed chip (T-Head), Alibaba Cloud, the Qwen large language model, and a wide array of application scenarios.

  Strengths

  Full-stack capabilities drive cost optimization and performance synergy; Alibaba Cloud holds the No. 1 market share in China.

  Comparison

  Compared with Baidu and Tencent, Alibaba boasts a higher degree of integration across its infrastructure and model layers.

  Risks

  Competition in the cloud computing market is intensifying (e.g., ByteDance).

• MiniMax (0100.HK) / Z.ai (2513.HK)

  A leading provider of foundational large models in China, benefiting from rising API pricing and rapid growth in ARR.

  Strengths

  The model boasts high intelligence density and low inference costs; recently, it has demonstrated pricing power, moving beyond mere commoditization.

  Weaknesses

  Computational power constraints remain a key challenge.

  Risks

  Technological iteration is rapid, and competition is intense.

• Contemporary Amperex Technology Co., Limited (300750.SZ) / Sigen Electric Co., Ltd. (002028.SZ) / Yingliu Group Co., Ltd. (603308.SH)

  Key beneficiaries of power infrastructure. CATL stands to gain from surging demand for energy storage; Synergy and Yinfeng stand to benefit from grid‑equipment supply and gas‑turbine exports.

  Strengths

  A world-leading battery and equipment manufacturer, benefiting from the urgent demand for flexible power in AI data centers.

  Risks

  Fluctuations in raw material prices and overseas trade policies.

• Cambricon (688256.SH) / Hygon Information (9903.HK) / SMIC (0981.HK)

  Core beneficiaries of semiconductor localization. Cambricon and Hygon provide domestically produced AI accelerator cards; SMIC serves as the cornerstone of advanced-node manufacturing.

  Strengths

  Against the backdrop of geopolitical dynamics, it boasts advantages in supply-chain security and policy support; its cost-effectiveness stands out in inference scenarios.

  Weaknesses

  Advanced process technologies are still constrained by upstream equipment, such as lithography machines.

  Comparison

  Cambricon enjoys strong traction in cloud-based inference, while Hygon boasts a competitive edge in supply-chain resilience.

  Risks

  U.S. export controls have been further tightened.

• Beisen Holdings (9669.HK) / Meitu Inc. (1357.HK) / Stone Technology (688169.SS)

  AI application-layer companies demonstrate significant margin expansion or revenue growth driven by AI.

  Strengths

  Beisen’s AI-driven ARR is expected to grow by 10x; Meitu’s paid‑user conversion rate is improving; and Roborock’s algorithms are driving premiumization and margin expansion.

  Comparison

  The most attractive securities identified through risk-return analysis.

  Risks

  Weak macroeconomic consumption is weighing on demand.

Key data

• The cumulative boost to TFP from AI~3pptOver the next decade, this will partially offset the impact of population aging.
• Potential GDP Growth Rate in 2035+3.5pptCompared to the scenario without AI adoption
• Semiconductor Self-Sufficiency Rate Forecast41% (2025) -> 86% (2030)Supporting more resilient and cost-effective deployment
• Earnings Before Interest and Taxes (EBIT) margin expansion among AI adopters12-13pptFrom approximately 4% in 2021 to 16–17% in 2027E.
• China AI Chip TAMUS$67 billionBy 2030, CSP demand is expected to dominate.
• Humanoid Robot Sales Forecast28,000 units (2026E)More than double the 2025 level, entering the scale-up phase after 2030.
• Annual incremental deployment of data center ESS~321GWhGlobal total capacity is projected to reach 85 GWh by 2030, with China accounting for a significant share.

Impact & implications

For the Chinese market, this means that the investment rationale must shift from “hype-driven themes” to “performance validation.” For infrastructure providers—such as those in power equipment, semiconductor manufacturing, and data centers—the upward cycle in capital expenditures driven by AI is well established. In particular, companies that address key pain points like “power‑supply bottlenecks” and “domestic substitution of computing power” are poised for sustained order growth and expanding market shares. At the application layer, AI’s value primarily lies in “cost reduction and efficiency gains.” Firms that can deeply integrate AI into their business processes, thereby significantly boosting gross margins or operational efficiency—such as Meitu, Beisen, and Roborock—are likely to see a revaluation of their valuations. By contrast, traditional software companies lacking distinctive data‑driven moats or easily replaceable by AI‑powered solutions face risks to their long-term terminal value. On the macro level, policymakers will need to strike a balance between fostering the widespread adoption of AI to maintain global competitiveness and mitigating the social pressures arising from labor displacement. Policies are expected to favor protecting labor‑intensive service sectors while permitting faster automation in high‑end cognitive domains.

Risks

• Labor substitution risk: If AI adoption accelerates too rapidly and policy responses prove inadequate, it could trigger widespread job losses, intensify deflationary pressures, and give rise to a negative feedback loop characterized by “micro-level efficiency gains coupled with macro-level demand contraction.”
• Semiconductor Supply Chain Constraints: Restricted access to upstream EDA tools and advanced manufacturing equipment could impede performance breakthroughs and capacity expansion for domestically produced AI chips.
• Delayed monetization of enterprise AI: Despite rising adoption rates, revenue generation from B2B applications remains in its early stages and uneven across sectors, which could weigh on the near-term earnings performance of relevant companies.
• Geopolitical Risks: Further escalation of export control policies could disrupt supply chains or restrict market access.

What to watch

• The pace and scale of AI adoption—particularly the progress in transitioning from pilot projects to large-scale production environments.
• Policy Response: The extent and direction of government measures to mitigate the employment impacts of AI, such as reskilling programs and social safety nets.
• Progress in Power Infrastructure: Deployment of Data Center Energy Storage Systems and Grid Flexibility Measures.
• Large‑model pricing trends: Monitor whether API prices continue to rebound, thereby confirming that the industry has moved beyond price wars and entered a phase of value‑driven growth.