Greater China Semiconductors: AI Infrastructure Buildout Drives Multi-Year Growth Across Foundry, ASIC, and Localization

Core Conclusion

The Greater China semiconductor supply chain is entering an extended upcycle driven by hyperscale AI infrastructure spending. Top 4 CSP capex surged 95% YoY in 1Q26, with total cloud capex reaching ~US$685bn in 2026e. This creates multi-year demand for foundry (TSMC), advanced packaging (CoWoS/SoIC), ASIC design services (Alchip, GUC), and China's domestic AI chip ecosystem. The bull case assumes cloud AI semi TAM grows to US$235bn in 2025e and US$753bn by 2030. We overweight TSMC, MediaTek (top pick), SMIC, Alchip, Aspeed, Hon Precision, WinWay, MPI, Cambricon, and Iluvatar.

Evidence Chain: AI Infrastructure Demand Is Real and Scaling

CSP capex remains robust and structurally higher. Top 4 cloud providers (Amazon, Google, Microsoft, Meta) posted 95% capex growth in 1Q26, and capex-to-EBITDA ratios stay stable at ~50%. Morgan Stanley's cloud capex tracker estimates US$685bn in 2026 from listed global CSPs alone, not including sovereign AI. This directly feeds semiconductor wafer and packaging demand.

TSMC is the primary bottleneck and beneficiary. AI semi revenue could reach 60% of TSMC's total by 2029e, and >30% as early as 2026e. TSMC likely doubled CoWoS and SoIC capacity in 2025 and will continue into 2026. Global CoWoS capacity is projected to reach 165kwpm by 2027e, up from ~370k wafers in 2024. NVIDIA remains the largest consumer (59% of 2026e CoWoS allocation), but Broadcom/Google TPU (20%), AMD (9%), and AWS (4%) are expanding rapidly.

ASIC projects are proliferating. Every major CSP now has a custom chip roadmap: Google TPU v7/v8, AWS Trainium 3, Microsoft Maia, OpenAI Nexus. Broadcom's TPU v7p (Ironwood) alone implies 145k CoWoS wafers in 2026e. MediaTek is designing Google's TPU v8t (Zebrafish) on 3nm. Alchip and GUC are key design partners for AWS and other hyperscalers. Aspeed, as the leading BMC provider, benefits from server complexity.

Investment implication: Overweight TSMC (12% upside), MediaTek (top pick, -13% PT upside but AI ASIC optionality), Alchip (6% upside), GUC (-10% upside), and Aspeed (12% upside). Testing equipment companies like Hon Precision, WinWay, and MPI are structural winners as chip testing time and pin counts rise. Hon Precision's handler market share is expected to grow from 28% (2024) to 42% (2027e), with EPS CAGR of 67% (2025-28e).

China AI Chips: Localization Accelerates Despite Export Controls

Demand is strong and self-sufficiency is rising. China AI GPU TAM is expected to reach US$67bn by 2030e, with domestic advanced node capacity supporting US$58bn of that. By 2030e, China AI chip self-sufficiency could hit 86%. Key domestic vendors—Cambricon, Iluvatar, MetaX—are gaining traction. Cambricon leads in inference performance and customer anchoring; Iluvatar leverages resilient supply chain with strong order visibility. Both are rated Overweight.

Inference economics favor domestic chips. Domestic AI accelerators offer lower total cost of ownership (TCO) and comparable cost per token vs. NVIDIA's China-compliant offerings. ByteDance token surges and rising NVIDIA 5090 gray-market prices confirm demand. Huawei's CloudMatrix 384 A3 SuperPod and Ascend 950 die shot illustrate system-level innovation.

Investment implication: Overweight SMIC (9% downside, but China's foundry enabler), Cambricon (7% upside, 90% revenue CAGR 2025-28e), Iluvatar (2% upside, 122% revenue CAGR). NAURA and AMEC (WFE) also benefit from China's capacity expansion. Risks include execution and further US export restrictions.

Key Risks and Divergences

1. Technology inflation and AI cannibalization: Rising wafer, OSAT, and memory costs pressure non-AI chip designers. Non-logic foundry utilization may only reach 80% in 2H26, implying non-AI demand is still weak.
2. Energy and capacity constraints: US power deployment limits (e.g., Rubin rack power 220kW per rack) and China's chip capacity caps could throttle growth.
3. Geopolitical escalation: US Executive Orders 14032 and 14105 could further restrict China's access to advanced nodes. Any Huawei/Iluvatar-related sanctions pose downside.
4. CoWoS competition: Intel EMIB could compete if it scales, though TSMC's roadmap (9.5x reticle) remains ahead.
5. Memory oversupply risk: NAND shortage is tight now, but NOR flash may normalize after 2026. DDR4 shortage persists but spot prices are capped.

Valuation and Trade Implications

The sector trades at elevated multiples: TSMC at 34.9x 2025e P/E, but 17.9x 2027e once earnings compound. Foundry median P/E is 39.6x 2025e, falling to 34.2x 2027e. AI-oriented names (Alchip 70.9x, GUC NM, Aspeed NM) are pricing in high growth, but EPS CAGRs of 50-100%+ justify premiums. We recommend focusing on:

• Core AI enablers: TSMC, MediaTek, ASMPT (7% upside, CPO equipment play)
• ASIC design: Alchip, GUC
• Testing equipment: Hon Precision (15% upside, 67% EPS CAGR), WinWay (22% upside, 95% EPS CAGR), MPI (15% upside, 94% EPS CAGR)
• China AI: Cambricon, Iluvatar, SMIC
• Memory niches: Macronix (23% upside), GigaDevice (-1% but structural)

Underweight names with limited AI exposure: UMC (-30%), Vanguard (-30%), WIN Semi (-74%), Realtek (6% but 18.7x 2025e, low growth).

Appendix: Key Valuation Metrics (May 7, 2026)