Source Report

Samsung's HBM Expansion: 50% Capacity Surge Via 1c DRAM Conversion Creates Short-Term Bottleneck Before P5 Fab Relief

Samsung leverages its Pyeongtaek P4 fab expansions and 1c DRAM process upgrades to ramp HBM output 50% in 2026: starting from ~170,000 wafers/month in late 2025, it converts existing lines (adding 80,000 wafers Q2 2026 + 60,000 Q4) while prioritizing HBM4 shipments already validated at 11 Gbps for Nvidia Rubin GPUs, but full P5 fab (60 trillion KRW/$41.5B investment) won't contribute until 2028 mass production.[1][2][3]
- Current HBM wafers: 160k-170k/month (end-2025), ramping to 250k/month by end-2026 via 1c DRAM (60k end-2025 → 200k end-2026 total 1c capacity, ~1/3 HBM-dedicated).[1][4]
- P4 adds 60k DRAM wafers mid-2026; P5 construction (Nov 2025 start) targets 2028 for HBM4E/beyond, doubling Pyeongtaek investment.[5]
- HBM bit shipments triple YoY to 11.2B gigabits in 2026 (HBM4 ~50%), market share to 35% from 16%.[3]

New entrants face Samsung's data moat from Nvidia validation and scale—competing requires 2+ years of yield optimization on 1c nodes (currently ~50-60%) before matching this ramp, locking in AI GPU dependency.

SK Hynix's HBM Dominance: M15X Fab Wafer Starts in Feb 2026 Add 20-30% Capacity Amid 1c DRAM Pivot to 200k/Month

SK Hynix accelerates via Cheongju M15X (wafer deployment Feb 2026, full cleanroom May 2026, 50k wafers/month by mid-2027) and Yongin cluster (first fab Feb 2027, 3 months early), channeling 1c DRAM ramps (170-200k wafers/month early 2027, up to 270k by 2028 end) into HBM4 (world-first 16-layer/48GB at CES 2026, mass prod Q3 2026), securing ~70% Nvidia HBM4 share and all 2026 chips sold out.[6][7][8]
- HBM trajectory: 170k wafers/month end-2025; M15X boosts 20-30% (10k start → 50k Q4 2026); P&T7 packaging (April 2026 construction, end-2027 complete, $13B).[2][9]
- 1c DRAM: From 90k planned (April 2025) to 170-200k early 2027 via Icheon upgrades; HBM3E ~2/3 of 2026 shipments.[10]
- U.S. Indiana 2.5D packaging (H2 2028, $4B) for HBM4.[11]

Entrants must navigate SK Hynix's Nvidia lock-in (70% HBM4) and TSV/packaging head start—replicating requires $13B+ in parallel wafer/packaging fabs, with 18-24 month lags.

Micron's HBM Catch-Up: Packaging-Led Ramps in 2027 Delay Wafer Relief, 2026 Capacity Sold Out

Micron prioritizes Singapore HBM packaging ($7B, ops 2026, meaningful 2027) over wafer fabs (Idaho #1 mid-2027, #2 end-2028; NY groundbreak early 2026/2030 prod; Taiwan PSMC P5 for DRAM phases), enabling HBM4 high-volume (11+ Gbps, Q1 2026 shipments early) despite smaller base, with all 2026 HBM sold out and supply meeting only 50-67% demand.[12][13][14]
- No explicit wafer targets; relies on 1-gamma transitions + Singapore NAND fab (H2 2028 wafers); Hiroshima HBM fab ($9.6B, shipments 2028).[15]
- Capex $20B FY2026 (up from $18B) for HBM/1-gamma; forecasts $100B HBM TAM by 2028 (40% CAGR from $35B 2025).[16]
- DRAM constraints persist beyond 2026; HBM4 ramps Q2 2026.[12]

Micron's U.S.-heavy buildout ($200B total) lags Korean peers' speed—new players risk certification delays (2+ years) on HBM4 yields before accessing Nvidia/AMD chains.

3:1 Cannibalization Locks In Structural Deficit: HBM Wafer Use Triple DDR5 Per Bit Drives Broader DRAM Crunch

HBM's stacked architecture demands ~3x wafer area per gigabit vs. DDR5 (worsening to 4x+ for HBM4), forcing reallocations: e.g., one HBM stack wafer yields less bits after TSV/hybrid bonding losses, cannibalizing conventional DRAM even as total wafers grow ~20% YoY, amplifying shortages (DDR5 prices 4x since Sep 2025).[17][2]
- HBM takes 23% DRAM wafers 2026 (up from 19% 2025), demand +70% YoY; industry HBM shortfall widens 5%→6%→9% (2025-27).[18][19]
- All vendors sold out 2026; AI servers need 2-3x prior DRAM/NAND bits.[12]

Competitors can't sidestep: HBM priority starves DDR ecosystem, raising entry barriers via 3-year fab cycles and 18-month quals.

Supply-Demand Outlook: Deficit Persists To 2028, Rebalance Unlikely Before New Fabs (2027-28 Adds ~20-30%)

Analysts (SemiAnalysis, Micron, SK Hynix) forecast HBM shortages through 2028: 2026 capacity (~500-600k wafers/month total est. from ramps) sold out, demand outpaces 20% supply growth; rebalance possible late-2027/2028 if M15X/Yongin/P5/Idaho hit timelines, but AI TAM $54B→$100B (2026-28) + Rubin/Feynman GPUs suggest deficit widens first.[18][16][20]
- Shortfall: HBM 6% 2026→9% 2027; DRAM supply lags demand >20% to 2026; tight to 2028 per SK Hynix/Micron.[12][20]
- 2027-28 adds: SK Yongin/M15X (50k+), Samsung P5, Micron Idaho/Singapore (~20-30% industry boost), but 3-5 year fab timelines limit catch-up.[2]

Aggressors entering now bet against sustained AI (Nvidia Rubin 2026+), but Big 3's $100B+ capex moat favors incumbents through deficit.

Implications for Market Entrants: 3-Year Fab Lag + Nvidia Lock-In Cements Triopoly

Capacity ramps (total ~20% YoY to 2027) chase 40%+ HBM demand CAGR, but 3:1 cannibalization + packaging/TSV yields (SK> Samsung > Micron) sustain premiums (HBM 35% over HBM3E); rebalance 2028+ risks oversupply if AI moderates.[16]

New firms need $20B+ capex, 2-year Nvidia quals, and 1c yields >60%—realistically, triopoly (SK 50-60%, Samsung 30-35%, Micron 20-25%) captures 90%+ through 2028.

---

Recent Findings Supplement (February 2026)

SK Hynix HBM Capacity Leadership with Sold-Out 2026 Supply

SK Hynix dominates HBM market share at ~50-60% for NVIDIA's HBM4 allocations, enabling it to pre-sell its entire 2026 HBM output (including HBM4 ramp) months ahead; this locks in pricing power as AI hyperscalers compete for limited stacks, but new packaging fabs delay meaningful volume additions until 2027.[1][2]
- SK Hynix's Cheongju HBM/packaging fab (P&T7, ~$13B investment) completes in 2027; Indiana HBM facility targets end-2028 production.[3]
- HBM4 mass production delayed to Q1 2026 end due to NVIDIA pin speed revisions (>11 Gbps), but SK Hynix secures mid-50% of NVIDIA's 2026 HBM4 volume.[4]
For competitors entering HBM, SK Hynix's first-mover yields (e.g., 12-Hi HBM4) and NVIDIA qualification create a 12-18 month lead; new entrants face 3+ year fab timelines and must fund $10B+ capex without guaranteed AI contracts.

Samsung's Aggressive 50% HBM Capacity Ramp for 2026

Samsung plans a 50% HBM capacity surge in 2026 via Pyeongtaek P5 fab start, shipping industry-first commercial HBM4 (11.7 Gbps/pin, 3.3 TB/s stack, 24-36GB 12-Hi) to close the gap with SK Hynix; this triples 2026 HBM sales vs. 2025 by prioritizing HBM4E samples in H2 2026, but relies on existing lines amid yield recovery from prior HBM3E issues.[5][2]
- P5 HBM production begins 2028; mid-20% NVIDIA HBM4 share secured.[6]
- Overall DRAM/HBM wafer starts prioritized for AI, with 2026 production fully allocated.[7]
Entrants must match Samsung's vertical integration (wafer-to-stack) and secure NVIDIA quals early; 50% ramps signal capex discipline favors incumbents, risking oversupply if AI demand softens post-2027.

Micron's U.S.-Centric Expansion Lags Near-Term HBM but Targets 2027-2030

Micron accelerates Idaho Fab 1 (H2 2027 DRAM wafers, pulled from late 2027) and Fab 2 (end-2028), plus $24B Singapore NAND/HBM packaging (H2 2028 wafers, 2027 packaging contribution) and Taiwan Tongluo (H2 2027); HBM4 shipments started Q1 2026 (quarter ahead), with full 2026 capacity sold out (~20% NVIDIA share), but New York megafab groundbreaking (Jan 2026) yields first output only in 2030.[1][8]
- FY26 capex hits $20B for HBM/US ramps; HBM TAM forecast: $35B (2025) to $100B (2028).[9]
- Refutes HBM4 disqualification rumors, confirms >11 Gbps speeds.[1]
New U.S. players benefit from CHIPS Act subsidies like Micron's, but 2-4 year fab equip timelines mean 2026-2027 reliance on Asian supply; Micron's HBM lag (11-20% share) highlights yield moats for leaders.

3:1 HBM Cannibalization Locks in Structural DRAM Constraints Through 2027

HBM's mechanism—requiring ~3x wafer input per GB output vs. DDR5 due to stacking/TSV complexity—cannibalizes commodity DRAM capacity even as total wafers grow 20-30%; all three vendors confirm 2026 HBM sold out, diverting ~20% global DRAM wafers to HBM by 2026, tightening DDR5/NAND supply despite AI-driven demand.[10][11]
- TrendForce: DRAM demand +35% vs. supply +23% (2026); prices +90-95% Q1 2026.[12]
- UBS: Shortages into 2027; HBM TAM $100B by 2028 (40% CAGR).[13]
Competitors face "no quick fix"—HBM shift is irreversible for margins; entering requires AI quals and $20B+ capex, with rebalance unlikely before 2028 new fabs.

Analyst Consensus: No HBM Rebalance Before 2028 Amid Supercycle

TrendForce/UBS forecast HBM shortages persist through 2027 (supply +15-17% vs. demand +20-22%), with rebalance ~2028 as Micron/Samsung/SK Hynix fabs ramp; "supercycle" extends to 2028 ($100B HBM TAM), but disciplined capex (DRAM +14% to $61B in 2026) prevents glut if AI moderates.[14][12]
- All 2026 HBM pre-sold; prices +55-60% Q1 2026 (TrendForce).[7]
- Potential 2028-2029 oversupply if AI demand slows post-new capacity.[14]
For market entrants, 2026-2027 is "reservation-only" for HBM; stockpile via long-term contracts now, as post-2028 glut risks 50%+ price drops—focus on HBM4/HBM5 quals to capture supercycle tail. Confidence high on sold-out status (direct quotes); medium on exact rebalance (analyst variance).