AI Hardware Expansion Anchored by TSMC’s Packaging Leadership and Capacity Growth

Core Conclusion

TSMC remains the fundamental beneficiary of the AI semiconductor buildout, underpinned by its unmatched CoWoS/SoIC advanced packaging capacity and pricing power. Concurrently, a structurally growing AI TAM, robust cloud capex, and the parallel rise of a competitive China AI semiconductor ecosystem create multiple investment avenues across the supply chain.

Evidence Chain

TSMC’s earnings momentum and capacity leadership are set to accelerate. The company’s 2Q26 revenue is projected to grow 5-10% QoQ with gross margin of 64-65%, driven by unrelenting AI demand. CoWoS capacity is expected to expand to 165k wafers per month by 2027, with systems integration (SoIC) becoming a new growth vector. This capacity expansion is supported by strong leading-edge utilization and stable pricing, funding a sustained high capex cycle.

AI semiconductor demand is broadening beyond general-purpose GPUs. Global cloud capex is estimated at $632bn for 2026, driving the AI semi TAM toward $753bn by 2030. Within this, inference and custom ASIC chips are growing at a faster CAGR (~65-68% from 2023-30) than training chips. Major CSPs continue to develop custom silicon (e.g., AWS Trainium, Google TPU), which diversifies demand across foundry and design chains.

The China AI semiconductor ecosystem is achieving competitive scale and cost efficiency. Despite export controls, domestic AI GPU TAM is projected to reach $67bn by 2030, with 76% self-sufficiency. Local chips like Huawei’s offer materially lower pricing, resulting in competitive performance-per-dollar and lower TCO for inference versus NVIDIA’s China offerings. SMIC’s advanced node roadmap (N+2/N+3) is key to supporting this production.

Key Divergences and Risks

Supply-demand imbalances for advanced packaging (CoWoS) and legacy memory (DDR4) could persist, creating bottlenecks or shortages. The sustainability of cloud capex intensity, given rising capex-to-EBITDA ratios for CSPs, poses a demand risk. In China, the development timeline for competitive domestic GPUs and potential WFE/EDA tool bottlenecks are critical uncertainties. Geopolitical tensions and further export control tightening remain overarching risks.

Valuation or Trade Implications

Overweight TSMC as the primary conduit for AI semiconductor growth. Within memory, prefer niche DRAM (Macronix, AP Memory, GigaDevice) where DDR4 supply constraints support pricing into 2H26. In the equipment chain, beneficiaries of packaging complexity and longer test times, such as Hon Precision (handlers), MPI (probe cards), and Winway (sockets), offer leveraged exposure.

Appendix Data Summary