AI memory may balance after mid-2028 while compute and supporting chips stay short, SMIC says

SMIC's forecast exposes a structural mismatch in the semiconductor industry: a persistent divergence between cyclical memory oversupply and structural compute scarcity. This bifurcation stems from AI inference's distinct reliance on high-bandwidth memory versus logic chips. With equipment lead times stretching to two years, the bottleneck has shifted upstream to lithography and etching tools, creating a 'time-lag' arbitrage opportunity in global supply chains. Competitors like TSMC are leveraging advanced packaging to bridge process gaps, while Samsung accelerates HBM4 mass production to secure storage dominance. Over the next 12-24 months, expect a market characterized by stabilizing memory prices and sustained compute premiums. Companies must mitigate single-technology dependency risks by adopting heterogeneous computing architectures to buffer supply volatility. Geopolitical fragmentation further complicates this landscape, forcing firms to balance compliance costs against R&D velocity. Ultimately, the winners will be vertically integrated players capable of efficiently orchestrating global resources and rapidly adapting to demand shifts, rather than those relying solely on process node leadership.