The market is pricing in an AI peak, but industry data has not confirmed it.

Over the past few weeks, global AI assets have experienced a significant correction since 2026. Korean storage heavyweight stocks, the US semiconductor index, and AI infrastructure companies that had seen substantial gains earlier have all retraced, with some core targets falling more than 20% from their recent peaks.

However, price changes are not synchronized with fundamental shifts. TSMC's Q2 revenue and gross margin remain at high levels, Microsoft has not cut its AI infrastructure spending, and the proportion of S&P 500 companies beating earnings expectations in early Q2 reports remains significantly above historical averages. The real change in the market is not that AI demand has suddenly disappeared, but rather that the pricing logic has shifted from capital expenditure expansion to the realization of orders, profit margins, and cash flow.

I. Stock prices have already priced in a cyclical top; earnings have not yet confirmed a turning point

JPMorgan data shows that the Korean market retreated by about 25% from its recent highs, the SOX index fell back by approximately 20%, and some storage and AI supply chain targets dropped between 20% and 50%. Meanwhile, global equity indices remain near highs, momentum factors have given up much of their year-to-date gains, and semiconductor technical indicators are quickly approaching oversold territory.

The first two items are already reflected in prices, while the third has not yet been supported by financial data. European semiconductors' relative stock prices have fallen noticeably, but earnings expectations for the next 12 months remain resilient, and the gap between price performance and EPS revisions is widening.

The main contradiction in semiconductors today remains valuation and positioning, rather than an earnings collapse. Only when orders, prices, gross margins, and capex guidance weaken simultaneously will this retracement be confirmed as the start of an industrial cycle reversal.

II. TSMC still shows no signs of cooling AI demand

In Q2 2026, TSMC achieved revenue of $40.2 billion, at the upper end of its original guidance of $39-40.2 billion; gross margin reached 67.7%, slightly above the upper end of original guidance; operating profit margin was 60.3%, also exceeding the previous guidance range.

The company expects Q3 revenue to further rise to $44.6-45.8 billion, with a midpoint of approximately $45.2 billion, representing continued growth of about 12% compared to actual Q2 revenue. Gross margin guidance is 65%-67%, remaining at high levels.

Core Verification

TSMC's Q3 revenue guidance continuing to grow quarter-over-quarter means AI demand is still being realized in wafer revenue. Even considering initial costs at overseas fabs, advanced packaging investments, and exchange rate fluctuations, the company's gross margin can still be maintained above 65%, and the profitability of advanced process and high-performance computing products has not yet shown a turning point.

AI supply constraints exist not only in wafer manufacturing but also involve HBM, advanced packaging, network interconnects, and power support. Expanding capacity in a single link does not mean the entire computing system immediately enters oversupply. TSMC's earnings report at least negates a more pessimistic assumption: that AI chip demand has deteriorated rapidly on the order side.

III. The key to the memory cycle is not expansion plans, but effective supply

Memory is the segment with the greatest divergence in this adjustment. The market has begun trading new supply for 2027-2028, including wafer fabs, advanced packaging plants, and HBM capacity expansions by Samsung, SK Hynix, and Micron. At the same time, increased capacity from Chinese memory manufacturers adds medium-to-long-term price pressure to traditional DRAM and NAND.

The issue is that capital expenditure does not equal effective supply. Memory expansion requires factory construction, equipment introduction, process certification, and yield ramp-up. HBM also requires advanced packaging support and customer validation, with expansion cycles typically longer than ordinary memory products. The market can trade supply growth in advance, but short-term prices ultimately depend on sellable capacity, not planned capacity.

JPMorgan believes that AI server demand, priority allocation of wafers to HBM, and slow release of new capacity may prolong DRAM and NAND supply-demand tightness until 2028. Its core logic is that HBM continues to occupy high-quality wafers and packaging resources, thereby squeezing supply for ordinary DRAM and enterprise-level NAND.

At this stage, the market has reacted strongly to distant supply, but prices and earnings for 2026-2027 have not deteriorated to the same extent. Therefore, memory stocks may continue to exhibit high volatility, but one cannot conclude that the industrial cycle has ended solely based on stock price retracements.

IV. Cloud vendors are still expanding; the market now demands capital returns

In the first phase of AI trading, capital expenditure itself was positive news. Increased data center budgets from Microsoft, Meta, Alphabet, and Amazon mean larger order opportunities for GPUs, storage, network equipment, and power infrastructure. Now, the market begins asking whether these investments can translate into revenue and cash flow.

Microsoft's Q3 FY2026 capital expenditure was $31.9 billion, with about two-thirds used for short-cycle assets like GPUs and CPUs; operating cash flow was $46.7 billion, and free cash flow was $15.8 billion. Azure revenue grew 40% year-over-year, and Microsoft's AI business annualized run-rate exceeded $37 billion, up 123% year-over-year.

The company expects next quarter's capex to exceed $40 billion and anticipates FY2026 natural year capex of approximately $190 billion, with about $25 billion coming from component price increases. Even with continued increases in GPU, CPU, and storage supply, Microsoft still calculates that compute constraints will last at least until the end of 2026.

Valuation Logic Has Shifted

For TSMC, memory, advanced packaging, and network equipment companies, cloud vendor capex remains revenue and orders; for the cloud vendors themselves, the same expenditure means higher depreciation, lower short-term free cash flow, and stricter capital return requirements.

Subsequent market focus will be on three indicators: whether AI revenue growth can sustainably outpace depreciation growth, whether compute utilization can offset the pressure infrastructure investment places on gross margins, and whether capex growth can ultimately translate into free cash flow growth. Cloud vendors have not stopped expanding, but the simple relationship of "higher capex equals higher valuation" has ended.

V. Cooling inflation can only alleviate valuation pressure, not replace earnings growth

JPMorgan views marginal disinflation as an important macroeconomic condition for market stabilization in this round. The logic is that falling energy prices and short-term inflation momentum help relieve upward pressure on bond yields and monetary policy, thereby improving the valuation environment for cyclical stocks and high-profit tech assets.

US CPI fell 0.4% month-over-month in June, rising 3.5% year-over-year; core CPI was flat month-over-month, up 2.6% year-over-year. Short-term price momentum has clearly cooled, but overall inflation remains above policy targets.

This means the discount rate pressure facing tech stocks may no longer rise rapidly, but one cannot assume interest rates will drop immediately. For AI assets with still-high valuations, subsequent upside must rely on earnings growth, not merely valuation expansion. A more accurate description is not "a loose cycle has begun," but rather "the tail risk of worsening inflation has decreased."

VI. Earnings season is becoming the new pricing anchor

As of mid-July, about 10% of S&P 500 companies have disclosed Q2 results, with 88% beating earnings expectations, higher than the ten-year average of 76%. The market expects S&P 500 overall EPS to grow approximately 24.7% year-over-year in Q2.

Early samples are concentrated in financials and a few large enterprises, so they do not yet represent the full tech earnings season, but at least indicate that corporate earnings have not deteriorated in line with stock price adjustments.

Stock price decline, valuation compression, and continued EPS upgrades improve the risk-reward ratio; stock price decline accompanied by simultaneous EPS downgrades implies the retracement may just be the beginning of earnings deterioration.

Semiconductors currently resemble the former scenario. Traditional software, business services, and some media companies face different risks: AI may lower development barriers, compress headcount, and weaken subscription pricing, meaning their earnings pressure comes from business model changes, not crowded positions.

VII. AI assets enter a structural screening phase

Although semiconductors and software are both categorized under technology, AI's economic impact on them differs. For semiconductors, AI enhances chip value, storage capacity, network bandwidth, and packaging complexity, with primary risks stemming from valuation and supply cycles; for some software and business services, AI may reduce per-task costs, decrease user seats, and weaken pricing systems, with risks arising from revenue structure and pricing power.

The segments with greater allocation value in the next phase are not all AI-related assets, but those that simultaneously meet three conditions:Earnings expectations are still being upgraded, effective supply remains constrained, and valuations have undergone substantive compression.

Conclusion

This round of AI adjustment has released a considerable amount of momentum and valuation pressure. TSMC's revenue and gross margin remain strong, Microsoft continues to increase AI infrastructure investment, and the proportion of S&P 500 early earnings beats exceeds historical averages, none of which support the judgment that the AI infrastructure earnings cycle has ended.

However, the market will not return to a stage where only capex is watched while ignoring capital returns. Whether semiconductors have truly bottomed still depends on three variables: whether core manufacturers can continue to raise revenue and gross margin guidance; whether cloud vendor AI revenue can outpace depreciation and capex; and whether HBM and traditional memory prices can maintain resilience before new supply is released.

The AI infrastructure earnings cycle has not yet been falsified; semiconductors are moving from a valuation compression phase into an earnings verification phase. The next rally will not come from all AI assets re-expanding valuations, but from a select few companies that can consistently realize orders, profit margins, and cash flow.

Data Sources: JPMorgan Global Equity Strategy, TSMC Company Announcements, Microsoft Company Announcements, FactSet, and US Bureau of Labor Statistics. Data current as of late July 2026.

Disclaimer: This article is for industry research and information exchange purposes only and does not constitute any investment advice. Markets carry risks; investment requires caution.