The AI hard-tech correction has released some risks, with allocation focus shifting to rotation and diversification

Overview

The report uses ten categories of market and fundamental signals to assess the maturity of the China AI hard-tech correction cycle. Previously strong AI hardware, momentum, small-cap, and growth styles reversed quickly, while valuations and speculative sentiment cooled significantly, but market concentration, financing leverage, and institutional technology holdings remain elevated. Goldman Sachs judges that the correction has released a substantial portion of risk and that the long-term fundamentals of artificial intelligence have not been damaged, but near-term market volatility may persist; therefore, the investment focus should shift from a single hard-tech exposure to rotation, diversification, and stock-specific returns.

Core views

First, the correction at the price and style levels has been relatively sufficient, with extreme divergences between soft tech and hard tech, as well as momentum and small-cap factors, narrowing significantly, suggesting that the pace of drawdowns may begin to slow. Second, risks have not been fully cleared: returns and turnover in A-share hard tech remain highly concentrated, STAR50 absolute valuations are still not cheap, margin financing balances are above historical norms, and public mutual fund technology allocations are near historical highs. Third, supportive factors are increasing, including retail sentiment returning to neutral, improved signals from buybacks and insider transactions, declining policy-tightening risk, and the national team turning to net buying. Fourth, global artificial intelligence capital expenditure remains strong, but the marginal momentum of capex and earnings upgrades is beginning to weaken; capex guidance and commercialization progress in the upcoming earnings season will determine the direction of the next phase. Fifth, at the portfolio level, investors should gradually position in H-share soft tech, policy beneficiaries and domestic substitution, oversold stocks with earnings upgrades, as well as IPO and shareholder return themes.

Analysis framework

The report assesses whether correction risks have been digested from ten dimensions: return dispersion and factor rotation, market breadth and concentration, valuation, financing leverage, retail sentiment, institutional positioning, corporate actions, policy statements, national-team trading, and capital expenditure and earnings revisions, combining historical cycles, cross-market comparisons, high-frequency fund-flow indicators, and text-quantification tools.

Methodology notes

• Comprehensive risk monitoringTen-Signal Framework for the AI Hard-Tech Correction

  Assess the correction cycle across multiple dimensions including pricing, risk appetite, fund flows, positioning, policy, and fundamentals.

  The ten signals cross-validate one another, avoiding reliance solely on price declines to judge whether the market has bottomed, and distinguish risks already released from those that remain elevated.

• Quantitative factor analysisReturn Dispersion and Style Rotation Analysis

  Compare relative return deviations between hard tech and soft tech, as well as factors such as momentum, small caps, and growth.

  Relevant styles have quickly reversed from extreme levels in the first half and moved close to long-term averages, implying that the pace of selling driven purely by price momentum may slow.

• Valuation analysisForward P/E and PEG Comparison

  Examine absolute valuations, growth-adjusted valuations, and relative premiums or discounts versus global peers at the same time.

  A-share hard-tech valuations have fallen back near long-term averages, and PEG is more reasonable, but absolute valuation risk for the STAR50 and valuation risk relative to global peers have not been fully eliminated.

• Investor behaviorA-share Retail Sentiment Indicator

  Integrate 13 high-frequency inputs including margin financing balance, new account openings, IPO subscriptions, ETF flows, and turnover velocity.

  A five-day moving average of the median z-score over the past year is used to measure sentiment; the current level is about 0 standard deviations, significantly cooler than about 1 standard deviation one month ago.

• Policy text analysisPolicy Risk Barometer

  Use large language models to quantify the direction, frequency, and intensity of policy statements and regulatory announcements.

  The model classifies policy signals as tightening, neutral, or easing and scores them as mild, moderate, or extreme; results show that market overheating and policy-tightening risks peaked in the first quarter of 2026 and then returned to a neutral range.

• Fund-flow analysisNational Team Tracker

  Estimate national-team holdings and weekly net purchase size, and compare them with subsequent market returns.

  Historically, large weekly net purchases exceeding 1.5 standard deviations have often been associated with medium-term market bottoms, but whether current buying strength has reached that threshold still requires ongoing observation.

• Asset allocationHard-Tech and Soft-Tech Rotation Model

  Assess the recovery potential of soft tech based on relative returns, valuations, earnings, and commercialization trends.

  The model indicates that H-share Internet and soft tech may continue to recover part of their relative underperformance versus hard tech over the coming months.

Asset mapping & comparison

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

• China A-shares

  Maintain Overweight; the main beneficiary market of liquidity, policy support, and the national-team stabilization mechanism.

  Strengths

  Long-term artificial intelligence fundamentals remain strong, policy risk has become neutral, and the national team has recently turned to net buying.

  Weaknesses

  Hard-tech returns and turnover are highly concentrated, and margin financing balances remain above historical norms.

  Comparison

  Compared with H-shares, A-shares have higher exposure to hard tech, small caps, and retail trading.

  Risks

  Further deleveraging, quantitative funds reducing risk, renewed policy tightening, and downward revisions to earnings expectations.

• A-share AI hard tech

  Medium- to long-term structural preference, but concentration should be controlled and positioning should be selective in the short term.

  Strengths

  Benefits from global artificial intelligence capital expenditure, domestic substitution, and supply-chain earnings growth.

  Weaknesses

  STAR50 absolute valuations remain high, fund holdings and financing leverage are concentrated, and earnings-upgrade momentum is slowing.

  Comparison

  Valuations and relative returns have fallen significantly from June highs, but the valuation discount versus global AI peers has narrowed.

  Risks

  Second-order changes in capital expenditure growth, reversal of the global AI trade, and continued unwinding of crowded trades.

• H-share Internet and soft tech

  Tactical attractiveness has improved; suitable for gradual additions to participate in rotation from hard tech to soft tech.

  Strengths

  Previously lagged significantly, valuations are near cyclical lows, food-delivery subsidy losses are narrowing, and commercialization of cloud computing, agents, and token consumption is improving.

  Weaknesses

  Still faces uncertainty around capital expenditure, regulation, and earnings delivery.

  Comparison

  Compared with A-share hard tech, valuations are lower, position crowding is smaller, and it may continue to recover the relative return gap.

  Risks

  Artificial intelligence commercialization falls short of expectations, competition intensifies, and regulation changes.

• Policy beneficiaries and domestic substitution themes

  Can provide structural growth and diversified returns across AI and non-AI areas.

  Strengths

  Covers 15th Five-Year Plan strategic directions such as artificial intelligence self-sufficiency, biotechnology and pharmaceuticals, service consumption, advanced manufacturing, and agriculture.

  Weaknesses

  There are differences in the pace of policy delivery and conversion into corporate earnings.

  Comparison

  Recently, domestically oriented AI companies have outperformed export-oriented AI companies.

  Risks

  Policy support weaker than expected, execution delays, export restrictions, and industry competition.

• Oversold technology and non-technology stocks with earnings upgrades

  Used to refocus on earnings delivery and capture stock-specific alpha.

  Strengths

  Some companies with solid fundamentals and upgraded earnings expectations have been excessively sold off amid valuation and positioning resets.

  Weaknesses

  Earnings quality needs to be verified company by company; screening cannot rely solely on price declines.

  Comparison

  Compared with pure thematic trades, these assets are more directly supported by earnings revisions.

  Risks

  Results miss expectations, earnings upgrades reverse, and valuations continue to compress.

• China and Hong Kong IPOs

  Can provide short-term alpha and relatively low-correlation returns.

  Strengths

  Median phased returns after year-to-date new listings are relatively high.

  Weaknesses

  Return persistence may be limited, with significant stock-specific differences and allocation constraints.

  Comparison

  Compared with mature listed stocks, IPO returns depend more on offering pricing and short-term supply and demand.

  Risks

  Excessive offering valuations, declining liquidity, selling pressure after lock-up expirations, and reversal of market sentiment.

• High-shareholder-return stocks

  Suitable as core diversified portfolio assets and sources of cash return.

  Strengths

  Dividends and buybacks provide relatively stable cash returns, and the theme has performed relatively resiliently during this correction.

  Weaknesses

  Growth elasticity is usually weaker than artificial intelligence themes.

  Comparison

  Compared with high-growth technology stocks, return volatility is lower and visibility of cash returns is higher.

  Risks

  Deterioration in earnings and cash flow, dividends or buybacks falling short of expectations.

Key data

• Average first-half gain of China AI stocks33%Refers to the China AI stock universe defined in the report.
• Drawdown of major growth indices from recent highsMore than 20%Includes the STAR50, ChiNext, and CSI 1000.
• First-half return gap between STAR50 and HSTECHMore than 100%The degree of return dispersion is close to the extreme level seen in early 2021.
• Concentration of STAR50 year-to-date returnsTop ten outperformers contributed 90%Shows that return concentration in A-share hard tech remains high.
• Share of A-share turnoverTechnology 26%, ChiNext 14%, STAR50 6%Turnover concentration in related sectors is at a high level in recent years.
• June valuation peakChiNext at 26x forward P/E; median STAR50 constituent at 50xAfter the correction, A-share hard-tech valuations have fallen back to around or below long-term averages, but STAR50 absolute valuations remain elevated.
• A-share margin financing balanceRMB2.6 trillionEquivalent to 5.5% of free-float market capitalization, below the previous peaks of RMB3 trillion and 6.0%, but still above historical norms.
• Financing concentrationStocks in the highest financing-balance quantile account for 30% of the total balanceThis is the highest level in history, and concentration risk in AI hard-tech leverage remains prominent.
• A-share retail sentiment0 standard deviations within the past-year rangeDown from about 1 standard deviation one month ago, with risk appetite shifting to neutral to slightly weak.
• Domestic public mutual fund assetsTotal assets close to RMB40 trillion, of which about RMB7 trillion is allocated to equitiesEquity allocation accounts for about 6.6% of total A-share market capitalization and 15% of free-float market capitalization.
• Change in buybacks in the third quarter of 2026Number of buyback announcements up 35% year over year, amount up 59% year over yearOn a quarterly annualized basis, reflecting improved management confidence.
• National-team holdings and recent tradingHoldings of about RMB5 trillion, net purchases of more than RMB140 billion over the past three weeksHoldings account for about 5% of total A-share market capitalization; estimated net sales were about RMB1.5 trillion over the previous six months.
• Artificial intelligence spending by major global cloud vendorsMay exceed USD900 billion in 2026 and reach USD1.3 trillion in 2027Equivalent to about 1.7% and 2.3% of combined China and U.S. GDP, respectively.
• Capital expenditure and earnings upgrades2026 and 2027 capital expenditure expectations have been raised by 32% and 75% year to date; hard-tech earnings expectations have been raised by 12% and 22%The momentum of capital expenditure and earnings revisions has recently begun to slow from elevated levels.
• Post-listing performance of A-share and H-share IPOsMedian returns 1 month and 3 months after listing year to date are 75% and 68%, respectivelyThe report believes IPOs can still provide short-term alpha and low-correlation returns.
• Forecast cash returns to shareholders by Chinese listed companiesMay exceed RMB4 trillion in FY2026Including dividends and buybacks, expected to reach another record high.

Impact & implications

Strategically, investors should not simply view this drawdown as a reversal of the long-term AI thesis, nor should they assume that risks have been fully cleared. A more reasonable approach is to maintain medium- to long-term core exposure to A-shares and AI hard tech, while reducing concentration in high-valuation, high-leverage, and crowded names; gradually add to H-share soft tech where valuations are at cyclical lows and artificial intelligence commercialization is improving; and broaden exposure to 15th Five-Year Plan policy beneficiaries, domestic substitution, oversold non-tech stocks with earnings upgrades, IPOs, and high-shareholder-return assets. Future excess returns are more likely to come from sector rotation and stock-specific fundamentals than from a single AI hardware beta.

Risks

• A-share financing deleveraging may still be in its early stage, and leverage is highly concentrated in AI hard tech.
• Absolute valuations of hard-tech assets such as the STAR50 remain elevated, and the discount versus global AI peers has narrowed significantly.
• Public mutual fund allocations and overweights to the technology sector are near historical highs, and crowded trades have not been fully unwound.
• Momentum in global artificial intelligence capital expenditure and China hard-tech earnings upgrades is slowing from elevated levels, and second-order changes may weigh on valuations.
• Policy statements are currently becoming neutral, but sudden tightening in history has triggered reversals in China equity bull markets.
• Another rapid reversal in the global AI trade could affect the China market through high correlation and risk-appetite channels.
• The scale of recent national-team buying has not yet clearly reached the historical threshold commonly associated with confirmation of a medium-term bottom.
• IPO returns, soft-tech rotation, and improvement in artificial intelligence commercialization may fall short of expectations.

What to watch

• Whether relative returns among hard tech and soft tech, momentum, small-cap, and growth factors stabilize.
• Changes in concentration reflected by STAR50 return contribution, technology turnover share, and individual-stock correlations.
• STAR50 absolute valuation, PEG, and the valuation gap between China AI stocks and global peers.
• Margin financing balance, share of free-float market capitalization, and AI hard-tech financing concentration.
• Retail sentiment indicators, new account openings, ETF fund flows, turnover velocity, and small- and mid-size order flows.
• Public mutual fund technology positioning, cash ratios, and trading activity of quantitative strategies.
• Listed-company buybacks, unusual trading warnings, and net buying or selling by major shareholders and executives.
• Whether the Policy Risk Barometer shifts from neutral back toward tightening or easing.
• Whether weekly net buying by the national team exceeds the historical key signal threshold of 1.5 standard deviations.
• Capital expenditure guidance, artificial intelligence development progress, and commercialization roadmaps during the earnings season of major Chinese and U.S. cloud vendors.
• Whether China hard-tech earnings revisions continue to slow, and whether soft-tech earnings revisions can improve cyclically.