Samsung's AI memory boom drives record profit as device division stays in loss

Samsung's earnings reveal a structural reallocation, not a cyclical peak. The HBM ramp is cannibalizing consumer DRAM wafer capacity—explaining persistent device division losses as a deliberate capacity trade-off, not a demand failure. The bottleneck is shifting upstream. HBM4's 16-layer TSV stacking makes CoWoS advanced packaging, not DRAM die yield, the binding constraint. Samsung's race to secure NVIDIA and AMD HBM4 allocations in H2 2025 forces direct competition over TSMC's packaging roadmap, making CoWoS-L capacity the hidden ceiling on AI compute delivery. Competitive dynamics are being rewritten. SK Hynix's entrenched position as NVIDIA's preferred HBM supplier is under pressure, compelling accelerated HBM4 timelines. Micron is pivoting toward custom ASIC memory for differentiation. The oligopoly's pricing power is transitioning from cyclical bargaining to structural premium—AI capex growth outpaces memory capacity expansion, implying a sustained upward shift in the 2025-2026 price floor rather than a traditional supercycle reversion. Risk vectors: tightening US export controls on advanced memory to China raise compliance costs for HBM lines involving sub-14nm logic dies or advanced packaging. CXMT's rapid DDR5 iteration poses a 3-5 year structural substitution threat. Core judgment for the next 12-24 months: HBM is no longer a memory category—it is the pricing anchor of AI infrastructure. Samsung's device losses are a strategic cost, not an operational failure.