Nomura: Memory Supply-Demand Gap Widens, Stock Price at Just 3x P/E Is Severely Undervalued
nashnova research
Nomura's latest report flags a stark disconnect: memory stocks are still ~37% below peak, trading at roughly 3× forward P/E for FY2027 — yet the supply shortage is deepening and long-term agreements are reshaping the entire business model.
How wide is the supply gap?
Nomura's checks with Samsung, SK Hynix, SanDisk, and Micron confirm: data-center customers may receive only 70%–80% of the memory they need; non-data-center customers, roughly 50%.
"Spec-downs" — downgrading orders from high-end configurations to whatever can actually be supplied — have become the industry default, not the exception.
This means → the surge in data-center memory demand starting Q3 2025 will stay elevated through FY2027. This is not a one-quarter squeeze — it is a two-year structural shortage.
How does the HBM standoff affect the entire price chain?
If Nvidia and Google insist on 16-layer / 12-layer HBM, FY2027 HBM demand would jump over 80% year-on-year. Nomura calls the industry-wide push for high-stack HBM from the outset an unrealistic assumption.
The endgame is a compromise: HBM demand and supply are both revised down, HBM prices go up; the freed-up wafer capacity protects commodity DRAM supply, supporting commodity DRAM prices.
In plain terms = demand is not being destroyed — buyers and suppliers are re-slicing a fixed wafer pie. Fewer, pricier HBM units mean commodity memory avoids a supply cliff.
Why are long-term agreements the key to re-rating?
LTAs — multi-year volume-and-price lock-in contracts between memory makers and major customers — are expected to cover 50%–70% of total DRAM, NAND, and SSD sales, spanning roughly 10–20 clients over contract terms of about 5 years, with prices more rigid in the first three.
Customers pay upfront deposits equal to 20%–30% of expected contract-period revenue; breaking the deal costs roughly one year's purchase value in penalties. This means → customers are deeply locked in, and order visibility far exceeds the old spot-cycle model.
SanDisk guided a three-year operating margin above 75%; Micron's minimum profitability guidance is well above its historical peak gross margin of roughly 60%. This reflects a shift from violent cyclical swings toward utility-like profit stability.
Nomura does acknowledge: LTA negotiations are not fully complete, investors cannot yet see all terms, and consensus estimates may be less reliable than usual for some time.
Can capacity expansion keep up with demand?
Nomura estimates that lifting memory supply growth by another 15 percentage points would require wafer capacity growth to jump from the recent ~5% pace to over 20% — doubling current DRAM and NAND capacity within 4 years and tripling it within 6.
In plain terms = that means adding roughly 3 Pyeongtaek-scale fabs per year — Pyeongtaek is Samsung's largest production complex — nearly impossible on any realistic capital and construction timeline.
Chinese memory makers pose limited near-term risk: even adding 100k wafers/month annually would not change the global picture. Nomura estimates US Big Tech data centers will account for ~75% of global memory demand in FY2027; the market actually reachable by Chinese players is only about 15%.
What are the macro-level risks?
Korean won appreciation: memory makers book 100% of revenue in US dollars with margins around 80%; a 10% won appreciation could cut profits by roughly 12%, likely dragging Korean memory makers' Q3 2026 operating profit below prior forecasts.
Interest-rate risk: if global rates fail to decline steadily and instead keep rising, the weakest links in the AI supply chain — which is in the middle of massive capex and financing — could come under pressure.
Nomura argues, however, that given the larger-than-expected upside in memory pricing, the FX headwind has limited impact on the FY2026–2028 earnings outlook overall.
What is Nomura's bottom line?
Nomura's core call: a soft landing in memory supply before FY2029 is very unlikely — the shortage is structural, not cyclical.
Whether the market fully prices in the LTA-driven business model will be the key test for a valuation re-rating of memory stocks.
This means → the current ~3× P/E is pricing in old-cycle earnings volatility, not the locked-in profit stability that LTAs deliver. If the market eventually accepts this shift, the re-rating room is enormous.
市场有风险，内容仅供研究参考，不构成投资建议。