Nomura Projects 110% Semiconductor Upside as AI Memory Demand Enters Structural Supercycle
Multi-year HBM expansion and geopolitical supply fragmentation position memory suppliers for sustained margin expansion through 2027.
Nomura expects memory suppliers to achieve 30% annual revenue and earnings growth over the next three to five years, following a 7-8x profit increase in 2026, marking what the firm calls a structural shift in semiconductor economics distinct from cyclical commodity pricing.
The thesis centers on high-bandwidth Memory (HBM) transitioning from supply-constrained scarcity to permanent architectural necessity across AI inference and training workloads. CNBC reports Nomura has identified a “triple memory super-cycle spanning DRAM, HBM and SSD memory” beginning Q3 2025, with margins expected to remain 200-300 basis points above historical averages through 2026.
HBM demand grew 130% year-on-year in 2025 and is projected to expand 70% in 2026, according to PatSnap, with AI and machine learning accounting for over 55% of total demand. The market itself has expanded from $7.27 billion in 2025 to a projected $9.18 billion in 2026, on track toward $69.75 billion by 2035—a 25.37% compound annual growth rate, per Precedence Research.
+58–63% QoQ
+70–75% QoQ
25.37%
>50%
Pricing Power Returns to Memory Oligopoly
DRAM prices roughly doubled since early 2025, while NAND Flash contract prices jumped 60% in Q1 2026 alone. Tom’s Hardware, citing TrendForce forecasts, projects conventional DRAM contract prices to rise another 58–63% quarter-on-quarter in Q2 2026, with NAND Flash climbing 70–75%.
The dynamic reflects deliberate capacity reallocation rather than transient supply shocks. Samsung and SK Hynix, controlling roughly 70% of the global DRAM market, have signaled prioritization of profitability over volume expansion, according to Sourceability. This strategic pivot has produced memory price surges reaching 80–90% quarter-on-quarter from Q4 2025 into Q1 2026 across most segments.
Supply constraints have structural underpinnings. IDC projects global DRAM supply growth at just 16% year-on-year in 2026 versus historical norms of 20–30%, with NAND supply growth at 17%. The constrained output reflects reallocation of fab capacity from commodity consumer DRAM toward high-margin HBM production, where SK Hynix alone maintains over 50% market share and has sold out capacity through 2026.
“Against this structural backdrop, we believe memory vendors have entered an unprecedented phase of rapid revenue growth and margin expansion in a short period of time.”
— Nomura
Hyperscaler Capex Anchors Multi-Year Demand Visibility
Amazon, Alphabet, Microsoft, Meta, and Oracle collectively plan $600–690 billion in capital expenditures for 2026, with approximately 75%—roughly $450 billion—directed to AI infrastructure, per CreditSights. This spending sustains elevated data center build-out cycles that drive memory intensity per rack beyond historical norms.
DRAM cost per gigabyte is forecast at $9.71 in 2026 versus $3.76 in 2025—a 2.58x increase—reflecting both pricing power and architectural shifts toward memory-intensive AI workloads. The economic reality has shifted procurement dynamics: Micron stated in recent guidance that it is “more than sold out” on HBM capacity through 2026, while Phison CEO K.S. Pua noted that “every NAND manufacturer told us 2026 is sold out.”
HBM differs from conventional DRAM by stacking multiple memory dies vertically and connecting them through silicon interposers, achieving bandwidth exceeding 2TB/s in next-generation HBM4 designs. This architecture is essential for GPU-accelerated AI workloads where memory bandwidth—not compute—often constrains model training and inference throughput. The transition from GDDR to HBM across AI accelerators has made high-bandwidth memory non-substitutable, creating pricing power absent in commodity DRAM markets.
Geopolitical Bifurcation Adds Supply Scarcity Premium
U.S. export controls effective January 15, 2026 revised licensing policy for advanced AI Semiconductors, shifting H200 and MI325X-class chips from presumption of denial to case-by-case review, according to the Federal Register. The same regulatory package imposed 25% tariffs on advanced AI semiconductors, excluding U.S. data center and R&D applications.
These controls have fractured global semiconductor supply chains along geopolitical lines. SupplyICs, citing Boston Consulting Group analysis, estimates maintaining parallel supply chains for advanced semiconductors adds 25–35% to landed costs for controlled markets and 10–15% for legacy nodes above 28nm. The fragmentation creates a supply scarcity premium favoring mature-node players operating outside restricted zones.
U.S. advanced semiconductor manufacturing capacity grew from 12% of global share in 2020 to 22% in Q1 2026 via $52 billion in CHIPS Act disbursements, per the Semiconductor Industry Association. Meanwhile, China’s CXMT is planning HBM3E mass production with 18–36 month supply lead times, while SMIC expands advanced node capacity toward 60,000 wafers per month in 2026.
Architectural Roadmap Extends Supercycle Visibility
Micron’s HBM4 samples feature 36GB, 12-high stacks delivering approximately 2TB/s bandwidth—60% higher than HBM3E—with production targeted for 2026. SK Hynix projects the global semiconductor industry to approach $1 trillion in annual sales in 2026, with the memory segment growing at 30%.
The architectural progression from HBM3 to HBM3E to HBM4 creates successive demand waves as each generation unlocks higher model parameter counts and inference throughput. This roadmap visibility underpins capital deployment confidence among memory suppliers, who historically faced boom-bust cycles driven by commodity oversupply.
Nomura’s thesis positions this expansion as categorically different: sustained data center intensity, oligopolistic supply discipline, and geopolitical fragmentation combine to support margin structures that break from historical mean reversion. The firm’s 110% upside projection for semiconductor equities rests on memory suppliers maintaining pricing power through architectural necessity rather than artificial scarcity.
- HBM demand growing 70% YoY in 2026 after 130% expansion in 2025, with AI/ML workloads accounting for 55%+ of total consumption
- Memory margins expected to hold 200–300bps above historical averages through 2026 on supply discipline and architectural necessity
- Hyperscaler AI infrastructure capex of $450B in 2026 provides multi-year demand visibility absent in prior cycles
- U.S. export controls and geopolitical supply chain bifurcation add 25–35% cost premiums for advanced nodes in controlled markets
- Major suppliers (Samsung, SK Hynix, Micron) sold out on HBM capacity through 2026; NAND manufacturers similarly constrained
What to Watch
Monitor hyperscaler Q2 2026 earnings for capex guidance revisions—any pullback from the $450 billion AI infrastructure baseline would pressure Nomura’s growth assumptions. Track BIS entity list updates and congressional semiconductor policy developments, particularly around reshoring incentives that could accelerate U.S. capacity expansion and erode supply scarcity premiums. Samsung’s Q2 earnings (late July) will provide the first major test of whether memory suppliers can sustain 200–300bps margin premiums amid rising fab utilization—historical patterns suggest margin compression as capacity comes online, but structural HBM demand may break precedent. Finally, watch for Micron’s HBM4 production timeline updates; any acceleration into late 2026 could intensify competition and pressure SK Hynix’s market share, potentially resetting pricing dynamics sooner than 2027–2028 projections anticipate.