AI Compute Demand Diversification Underpins TSMC's Leadership, Reshapes MediaTek and Memory Dynamics

Core Thesis

Sustained AI infrastructure investment, now diversifying across cloud CSPs and sovereign China initiatives, solidifies TSMC’s technological and pricing power in advanced nodes and packaging. This shift concurrently catalyzes a viable, cost-competitive domestic AI chip ecosystem in China, creating divergent winners in foundry (SMIC), design (MediaTek in ASICs), and creates structural tightness in legacy memory segments like DDR4 and NOR Flash.

TSMC: Capturing the AI Premium Through Unmatched Integration

TSMC’s dominance is less about discrete process nodes and more about its integrated leadership in advanced packaging (CoWoS, SoIC) which is critical for AI accelerators. Evidence shows CoWoS capacity is projected to reach 165k wafers per month by 2027, driven by demand from NVIDIA, Google, AWS, and custom chip projects. Furthermore, TSMC’s 2nm node demand is forecast to be overwhelmingly AI-driven, primarily from Apple and NVIDIA, supporting stable premium pricing. The investment implication is direct: TSMC’s revenue mix and gross margins are structurally lifted by this high-value AI demand, insulating it from cyclical swings in broader semi markets and justifying its capacity expansion capex.

The Rise of China's Indigenous AI Compute Ecosystem

Export controls have accelerated the decoupling of AI compute, fostering a domestic supply chain capable of meeting a significant portion of local demand. China's AI chip self-sufficiency is projected to reach 76% by 2030, supported by SMIC’s advancing nodes (N+2/7nm, N+3/5nm) and a cohort of "10 Dragons" domestic GPU/ASIC designers. Performance analysis indicates these local chips offer a compelling performance-per-dollar proposition for inference workloads, as seen with DeepSeek models. For investors, this signifies SMIC as a critical bottleneck and beneficiary, while fabless players like MediaTek gain in custom AI ASIC design (e.g., TPU v8 for Google). The risk is execution, dependent on overcoming WFE and EDA tool bottlenecks.

Structural Tightness in Niche Memory Driven by AI and Legacy Dynamics

AI is not only consuming leading-edge HBM but also creating indirect shortages in legacy memory. The AI-driven build-out necessitates supporting infrastructure, straining supply for DDR4 and NOR Flash. Supply-demand models indicate a persistent DDR4 shortage into 2H26, while NOR Flash faces undersupply due to robust demand from automotive, IoT, and display drivers. This creates a selective opportunity within memory. Companies like Macronix (NOR Flash) are positioned to benefit from this prolonged tightness and pricing power, distinct from the more cyclical dynamics of commodity DRAM and NAND.

Key Risks and Valuation Implications

Key risks include “tech inflation” where rising wafer and packaging costs pressure chip designer margins, AI demand potentially cannibalizing non-AI semiconductor budgets, and geopolitical disruptions to the China supply chain. Execution on China's advanced node capacity expansion remains uncertain. For valuation, TSMC trades at 20.8x 2026e P/E, with a 15% upside to our target, reflecting its premium but sustainable position. SMIC offers higher growth potential (22% upside) albeit with higher execution risk. In memory, we prefer selective exposure to names like Macronix (37% upside) benefiting from NOR Flash tightness. MediaTek’s role in custom ASICs provides a growth vector beyond smartphones, though near-term earnings face headwinds.

Appendix: Selected Company Valuation Snapshot