The pullback in the China AI theme looks more like healthy deleveraging than a bubble bursting

Overview

This report focuses on the investment view after the recent pullback in the China AI theme. JPMorgan notes that market discussion has centered on whether the AI ecosystem has entered a bubble-bursting phase, but its conclusion is that this decline is more consistent with “healthy deleveraging.” The report tests the resilience of the AI capex cycle from four angles—balance sheets, progress in LLM capabilities, hardware supply constraints, and A-share liquidity indicators—and recommends maintaining a bias toward large-cap, high-quality AI names.

Core views

The core views include: first, the balance sheets of major Chinese and U.S. cloud vendors remain healthy, with debt-to-asset ratios of around 40%, far below the extreme leverage seen in historical bubble cycles; second, LLM capabilities continue to improve, Chinese models are catching up with leading U.S. models, and new application scenarios may still drive ARR and infrastructure demand growth; third, AI hardware supply bottlenecks will not disappear in the short term, and the next two quarters are not the key window for testing capex sustainability; fourth, leverage in A-share AI trading has already been cleared out to a considerable extent, and ETF flows turned into inflows from early July, led by technology and hardware ETFs.

Analysis framework

The report uses a strategy research framework that combines market liquidity indicators with fundamental validation: on the liquidity side, it tracks A-share 5-day rolling turnover, the share of margin trading in turnover, the sector structure of margin trading, and ETF net inflows; on the fundamentals side, it compares cloud vendor debt-to-asset ratios, bond ratings, LLM capability evolution, domestic and overseas AI hardware capacity release timelines, and assesses post-correction allocation value alongside valuation multiples and valuation dispersion.

Methodology notes

• Liquidity analysisDeleveraging and trading structure observation

  Use turnover, margin trading share, and ETF fund flows to judge whether thematic trading is merely undergoing leverage cleanup.

  The report argues that A-share turnover has declined but not collapsed, and the share of margin trading in the IT sector has retreated from highs, indicating speculative leverage has been squeezed out, but this is not equivalent to a withdrawal of long-term capital.

• Fundamental validationThree checks on AI capex sustainability

  Validate the AI investment cycle through balance sheets, model capabilities, and hardware supply bottlenecks.

  Through cloud vendor leverage ratios, LLM capability improvements, and the supply timelines for HBM, advanced process nodes, and advanced packaging, the report rebuts the narrative that the AI bubble is about to burst.

• Valuation analysisFTM P/S and valuation dispersion

  Observe forward price-to-sales ratios and the degree of valuation differentiation among China AI names.

  The report states that after the correction, overall forward P/S multiples and valuation dispersion for China AI names are at more reasonable levels, which better supports selective positioning in large-cap, high-quality names.

Asset mapping & comparison

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

• Zhongji Innolight - A (300308.SZ)

  One of the top picks in the China AI ecosystem, categorized under the global supply chain theme.

  Strengths

  Benefits from AI hardware and global supply chain demand; listed by the report as a top pick.

  Weaknesses

  May be affected in the short term by liquidity volatility in the AI theme and valuation repricing.

  Comparison

  Together with Victory Giant-H, JCET-A, and Weichai-H, it belongs to the report’s preferred AI names related to the global supply chain.

  Risks

  Downward revisions to AI capex expectations, faster-than-expected easing of hardware supply bottlenecks, and continued deleveraging in thematic trading.

• Victory Giant Technology - H (2476.HK)

  One of the top picks in the China AI ecosystem, categorized under the global supply chain theme.

  Strengths

  The report includes it among high-quality names benefiting from the global AI supply chain.

  Weaknesses

  H-share risk appetite and fund flows in the tech hardware sector may bring short-term volatility.

  Comparison

  Together with Zhongji Innolight, JCET-A, and Weichai-H, it reflects the report’s hardware supply chain allocation approach.

  Risks

  AI hardware orders missing expectations, valuation compression, and changes in the pace of global supply chain capacity expansion.

• JCET - A (600584.SS)

  One of the top picks in the China AI ecosystem, related to advanced packaging and the global supply chain.

  Strengths

  The report mentions that JCET’s 7nm advanced packaging plant is expected to progress in 2H28, highlighting its medium-term capacity story.

  Weaknesses

  The capacity release timeline is distant, and short-term earnings validation may be insufficient.

  Comparison

  Compared with model application companies, JCET is more geared toward the AI hardware and packaging capacity chain.

  Risks

  Delays in advanced packaging construction, demand volatility, technology iteration, or intensifying competition.

• NAURA - A (002371.SZ)

  One of the top picks in the China AI ecosystem, categorized under the domestic substitution theme.

  Strengths

  Benefits from domestic demand for AI hardware and semiconductor equipment in China.

  Weaknesses

  High expectations may make valuation more sensitive to order timing.

  Comparison

  Together with AMEC-A, Baidu, Zhipu, and Iluvatar CoreX, it falls under the report’s localization plays.

  Risks

  Slower-than-expected localization progress, volatility in semiconductor capex, and changes in policy and supply chain constraints.

• Baidu.com (9888.HK; BIDU)

  One of the top picks in the China AI ecosystem, categorized under domestic substitution and large-model application.

  Strengths

  As a Chinese hyperscaler and AI model-related company, it benefits from improving LLM capabilities and expanding application scenarios.

  Weaknesses

  AI commercialization and ARR growth still require continued validation.

  Comparison

  Compared with Zhipu AI and Iluvatar CoreX, Baidu combines the characteristics of a cloud vendor, large-model player, and internet platform.

  Risks

  Model capability catch-up falling short of expectations, a longer AI investment payback period, and pressure on advertising or cloud businesses.

• Bank of China - H (3988.HK)

  One of the top picks in non-AI sectors, used for short-term sector rotation positioning.

  Strengths

  The report believes rotation into non-AI sectors may continue in July, and bank stocks can benefit from defensiveness and policy expectations.

  Weaknesses

  It has low correlation with the AI main theme, and its medium-term growth elasticity may be weaker than that of the technology sector.

  Comparison

  Together with CICC-H, Innovent, BYD-H, and CR Land, it forms the report’s non-AI top picks.

  Risks

  Changes in the macro growth and interest rate environment, pressure on bank asset quality, and unmet policy expectations.

Key data

• Cloud vendor debt-to-asset ratioabout 40%Major Chinese and global hyperscalers are all around this level, below the peak of China’s property cycle and leverage levels associated with the U.S. internet bubble.
• Historical peak debt-to-asset ratio of China property developersabout 100%Used in the report to compare with the current leverage level of the AI infrastructure layer.
• Average leverage of U.S. telecom operators during the internet bubbleabout 230%Used in the report to show that current AI cloud vendors are still far from the extreme leverage seen in historical bubble cycles.
• Peak A-share 5-day rolling turnover rateabout 6.5%It has already retreated meaningfully, but remains above the bottom levels of past bull markets.
• Margin trading as a share of turnover in the A-share IT sectorfrom about 12% down to 8% to 9%Indicates that the most leverage-concentrated part of AI trading has already undergone significant clearing.
• MXCN index base-case target for end-2026100The report maintains this target; the bear-case target is 80.
• CSI-300 index base-case target for end-20265,200The report maintains this target; the bear-case target is 4,000.
• Consensus EPS growth forecast year over year13% and 24%The report says this growth outlook supports the MXCN and CSI-300 targets, together with generally supportive liquidity conditions.
• Timing for easing of AI hardware supply bottleneckspossibly as early as late 2027 to 2028The report believes the next two quarters are not the window for validating capex sustainability.

Impact & implications

The investment implication is that, in the short term, the AI theme may still be affected by liquidity volatility and rotation into non-AI sectors, but the risk-reward profile has improved after the pullback. The report prefers large-cap, high-quality companies within the AI ecosystem, while also focusing on two main lines: the global supply chain and domestic substitution. Non-AI areas may continue to be supported in July by policy expectations, broker and insurance earnings, and healthcare performance, but the report believes the AI ecosystem could outperform again during the August earnings season on the back of second-quarter results and second-half guidance.

Risks

• Deleveraging in the AI theme may not be fully over, and short-term liquidity volatility may continue to pressure valuations.
• If AI capex is revised down, it could weaken earnings expectations for hardware and supply chain names.
• If progress in LLM capabilities slows, growth in new application scenarios and ARR may fall short of expectations.
• If hardware supply bottlenecks ease faster than expected, pricing power and the urgency of capex may decline.
• In July, policy expectations and rotation into non-AI sectors such as brokers, insurers, and healthcare may continue to divert funds away from AI.
• The index targets and company views in the report depend on consensus earnings growth, liquidity, and macro assumptions, and therefore carry forecast uncertainty.

What to watch

• Whether the A-share 5-day rolling turnover rate continues to fall toward historical bottom ranges.
• Whether margin trading in the IT sector as a share of turnover stabilizes around 8% to 9% or continues to decline.
• Whether A-share ETF net inflows can continue, especially fund flows into technology and hardware ETFs.
• Whether AI companies’ second-quarter results and second-half guidance outperform expectations during the August earnings season.
• Whether the gap between Chinese LLM capabilities and leading U.S. models continues to narrow.
• Capacity release milestones for Huawei Ascend, CXMT HBM, TSMC 3nm, Micron HBM, and SK Hynix memory supply.
• Whether MXCN and CSI-300 continue to be supported by EPS growth expectations and liquidity conditions.
• Whether rotation into non-AI sectors weakens after July and whether funds return to the AI ecosystem.