DRAM Is So Profitable That HBM Is Getting Cold-Shouldered: What the Memory Giants' Capital Shif

When traditional DRAM margins start to dwarf those of AI darling HBM, the capital expenditure logic of memory giants quietly pivots.  On the surface, it's a capacity reallocation. What really matters: the AI storage premium story is being overtaken by a traditional DRAM supply-demand gap. The investment narrative shifts from "AI narrative" to "cyclical dividend." ## 1. HBM Expansion Enthusiasm Cools: Three Drivers Behind the Capital Turn A recent BofA Merrill Lynch report highlights three reasons memory makers are becoming cautious on HBM expansion: - **Traditional DRAM is more profitable**: Current margins on LPDDR5 and other legacy DRAM significantly beat HBM, offering better capital efficiency. - **HBM4 order growth is slowing**: Demand momentum is fading at the margin. - **NAND KV cache substitution**: KV cache memory based on NAND is growing fast in AI inference, partially replacing HBM demand. Result: Some capex originally earmarked for HBM expansion is being redirected to traditional DRAM and NAND. H2 growth drivers will shift from HBM4 to legacy storage products. ## 2. Long-Term Agreements Lose Appeal: Locking Prices in a Rising Cycle Hurts Chipmakers' stance on long-term agreements (LTAs) has changed markedly. With customers willing to pay higher quarterly contract prices, the incentive to sign LTAs weakens. Key concerns: - If supply shortages persist into 2027-2028, LTAs would cap price and profit upside. - If the market turns down, long-term contracts carry default risk. Expect most legacy storage sales to remain on quarterly price negotiations through H1 2027, with limited new LTA volume. ## 3. Price Trends: Traditional DRAM Hits 25-Year Highs, DDR4 Shows Divergence Current memory prices are at historically extreme levels, but with internal divergence: - **16Gb DDR5**: Spot ~$38.5; contract prices expected to rise >20% MoM in April (after ~200% and ~80% sequential gains in Q4 and Q1). - **16Gb DDR4**: Spot ~$63.9, but down 5-8% this week as spot prices are >50% above DDR5, triggering substitution. - **DDR5 outperforming DDR4** likely to persist through June. - South Korea's semiconductor exports in the first 20 days of April surged 182% YoY, driven by DDR5, HBM, and NAND; DDR4 now accounts for <1%. ## 4. 2028 Capacity Expansion Expectations Downgraded: Supply Constraints Bolster Long-Term Bull Case Earlier fears of massive DRAM capacity from new fabs in Korea and the US by 2028 may be overblown. The report notes: Most chipmakers expect each new fab to add only 50k-60k wafers per month in 2028. Although cleanroom space could theoretically support 100k wafers per year, high execution risks (e.g., 1d node) make firms cautious on equipment spending. Global DRAM industry wafer capacity growth is only in the single-digit percentages. Chipmakers' optimism about long-term supply shortages is based precisely on this capacity expansion constraint. Notably, TCB (thermal compression bonding) equipment demand is expected to ramp significantly in 2028, mainly for 16- and 20-layer HBM4e packaging to support Nvidia Rubin Ultra and advanced AI ASICs. ## 5. So What? What Investors Should Watch The core tension in this memory cycle has shifted from "AI demand explosion" to "traditional supply shortage." - **Short term**: Watch DDR5 contract price increases and DDR4 pullback magnitude to gauge when substitution ends. - **Medium term**: Track HBM4 order growth and NAND KV cache penetration to assess if AI storage demand is overestimated. - **Long term**: Monitor actual 2028 capacity additions. If supply constraints persist, the memory cycle could last longer than expected. **Bottom line**: Don't let the AI narrative lead you around. Traditional DRAM's earning power is the hardest logic right now.