Semiconductors: Not a Rising Tide

Core Conclusion

The semiconductor industry is not experiencing a broad, synchronized recovery. AI-related demand is accelerating, supported by upward revisions to hyperscaler capex and a doubling of CoWoS advanced packaging capacity. However, the non-AI segment is stalled by excess mature-node capacity and persistent inventory overhangs in automotive and industrial end markets. Investors pricing in a uniform cyclical rebound are misjudging the depth of this divergence. The actionable implication is to treat AI infrastructure pullbacks as entry points, while deferring cyclical exposure until demand signals visibly turn. A second implication is that exposure to pure-play foundries and analog names reliant on broad unit recovery carries asymmetric downside if the macro environment weakens further.

Evidence: The AI Supply Chain Is Running Hot, but Non-AI Is Still Idling

Cloud capital expenditure guidance has been revised higher multiple times, with 2026 AI server unit shipments now expected to grow over 50% year-on-year. In parallel, TSMC’s CoWoS advanced packaging capacity plan has more than doubled. These data points confirm a genuine investment cycle, not a temporary burst. The investment consequence is that component constraints — especially in packaging and high-bandwidth memory — will persist, supporting pricing power for the few suppliers that can scale.

On the non-AI side, global semiconductor sales on a three-month moving average remain in year-on-year contraction. Analog and power discrete shipments have not recovered to 2022 levels. Automotive semiconductor inventory days are still above normal, and industrial demand is subdued by weak factory automation spend. The drawdown phase is ending, but a restocking cycle has not begun. That gap implies revenue disappointment risk for names where consensus already assumes a second-half inflection.

Key Risks

A macroeconomic slowdown would hit the non-AI recovery thesis and also compress valuation multiples for AI stocks. Geopolitical tension and technology export controls could further fracture supply chains and limit access to critical tools. A separate risk is that AI capital expenditure fails to translate into adequate returns, triggering a mean-reverting de-rating of the entire AI semiconductor complex. Finally, mature-node capacity overbuild in China and legacy fabs creates structural price competition — a margin headwind that cyclical upturns alone will not cure.

Investment Implications

For AI-exposed semiconductor companies, interim corrections driven by rotation or capex-return anxiety should be used to build positions, because the underlying capacity build-out is multi-year and likely undersupplied in 2026. For pure cyclical laggards tied to broad-based unit recovery, the burden of proof remains on demand. Until auto and industrial order patterns turn, early positioning carries negative carry and downside risk. The cost of waiting for confirmation is limited compared to the cost of being wrong prematurely.