Samsung, SK Hynix shift DRAM to DDR5 as margins exceed HBM

Samsung Electronics and SK Hynix are reallocating flexible DRAM manufacturing capacity toward server DDR5 production, a shift driven by improving profitability in conventional memory compared to HBM. According to data cited in the week's reporting, 64GB DDR5 RDIMMs now generate more revenue per wafer, and conventional DRAM margins exceeded HBM by over 15% in the first quarter of 2026. Neither company is abandoning HBM—they remain committed to their HBM output volumes—but they are redirecting available flexible capacity into the higher-return DDR5 segment. The supply outlook remains tight: Meritz Securities estimates that memory suppliers met only 75–80% of total demand in the second half of 2026 and projects the fulfillment rate could fall further to the 60% range in 2027. Samsung has also signaled continued capital investment in memory testing infrastructure, awarding DRAM tester contracts worth KRW49.85 billion (approx. US$33.5 million（約54億円）) to Exicon and KRW8.6 billion to Neosem, moves that indicate preparation for higher production volumes ahead. The rebalancing underscores how customer preferences and pricing dynamics can override traditional margin hierarchies: HBM remains essential for GPU-centric AI workloads, but the sheer volume of conventional server DRAM demand—driven by CPU-based infrastructure, hyperscaler expansion, and the broader move to AI-ready datacenters—has shifted the economics in DDR5's favor.

Memory demand outpaced supply throughout 2026, with DRAM and HBM both commanding premium margins as cloud providers and AI chipmakers compete for capacity. The fact that DDR5 servers now return more per-wafer revenue than HBM—historically the highest-margin memory product—signals a structural shift in datacenter architecture demand. Rather than a shortage of memory appetite, the reallocation reflects where customers are buying: cloud platforms prioritizing conventional server DRAM over the specialist bandwidth memory that powers GPU clusters. Samsung and SK Hynix's decision to flex capacity toward DDR5 while "maintaining committed HBM output" suggests they are optimizing the mix rather than abandoning HBM entirely, but the margin signal is clear. If suppliers can only satisfy 60% of total demand in 2027, the question becomes whether that shortfall concentrates in DDR5, HBM, or both—and whether customers with longer lead times or stronger negotiating power (like hyperscalers with long-term contracts) secure their portion first.

No comments yet. Be the first to share your thoughts!