Samsung Electronics (005930.KS)
The DRAM Supercycle. SOCAMM Demand. The HBM4 Redemption.
15 April 2026 · YK Research · Watchlist
Contents
The Setup
Samsung is printing the best quarterly numbers in its history and the stock is barely moving. Q1 2026 preliminary operating profit: ₩57.2T ($37.8B), up 8x YoY, blowing past the ₩40-42T consensus by 40%. That single quarter nearly matches FY2025’s entire operating profit of ₩43.6T.
The memory segment alone delivered ~₩54T in operating profit. The market yawns because it thinks this is a cyclical peak. The DRAM supercycle thesis says it’s not.
Company Snapshot
Segment Breakdown
Device Solutions (DS): Memory (DRAM + NAND) + Foundry. The profit engine. Memory is a machine; foundry lost ₩1.6T in Q1. Device eXperience (DX): Galaxy phones, TVs, appliances. ~₩4T OP in Q1, declining. Samsung Display (SDC): OLED panels. Stable. Harman: Car audio. Rounding error.
The DRAM Supercycle: Three Converging Forces
1. AI Inference Is Eating Conventional DRAM
Everyone fixates on HBM for AI training. But AI inference servers need enormous amounts of regular DDR5. As inference scales (faster than training), server DDR5 demand pulls capacity away from consumer applications. TrendForce projects DRAM contract prices up 50%+ in Q1, with another 30%+ expected in Q2.
2. SK Hynix HBM4 Shipment Cut: Margin Positive
SK Hynix reportedly cut its 2026 HBM4 shipment target by 20-30% (from ~6B Gb to ~4B Gb). The reallocated capacity goes to HBM3E and server DDR5. Because HBM4 has lower die yield (larger die area, finer TSV pitch), converting capacity back to DDR5 actually increases total bit output. Net: positive for Samsung’s highest-margin product (server DDR5) and gives Samsung more time to close the HBM gap.
SOCAMM: The Overlooked Demand Driver
NVIDIA’s AI server platforms don’t just need HBM. They need massive amounts of SOCAMM (System-on-Chip Attached Memory Module), which uses LPDDR5X, the same memory that goes into smartphones. Samsung is the world’s largest LPDDR5X producer.
Most investors are focused on VR300’s 1TB of HBM4E. The SOCAMM angle is the overlooked part. DRAM makers benefit across both HBM and LPDDR product lines, not just the HBM the market is pricing in.
The HBM4 Redemption Arc
Samsung lost DRAM market share leadership to SK Hynix for the first time in three decades because they couldn’t get HBM3 qualified with NVIDIA. The CEO publicly admitted to the competitive gap. But HBM4 looks different:
HBM was still <10% of Samsung’s DRAM revenue in Q1. That’s the opportunity. As HBM4 ramps through 2H 2026, Samsung gets a new high-ASP revenue stream on top of the already-exploding DDR5 business.
The Numbers
FY2025 Actuals
Q1 2026 Preliminary
Forward Estimates
Valuation
At ₩500T market cap with ₩57.2T quarterly OP, Samsung trades at ~2.2x annualized Q1 operating profit. Even backing out mobile/display/Harman (~₩10T annual OP), the memory business alone is valued at perhaps 2.5x run-rate OP.
Compare SK Hynix at 5-6x annualized OP. The “Korea discount” is real, but the gap is extreme given Samsung is now generating more absolute memory profit than at any point in its history.
What Breaks It
| Risk | Severity | Probability | Impact on Thesis | Mitigant |
|---|---|---|---|---|
| DRAM cycle turns | HIGH | 30-40% | Stock down 30-50%. Memory margins compress fast when cycle turns. Samsung has more commodity exposure than Hynix | AI structural demand shift may extend cycle. Watch TrendForce quarterly pricing. 2 consecutive flat/down quarters kills thesis |
| HBM4 mass production stumbles | MEDIUM | 25-35% | Kills the re-rating narrative. Samsung stays an HBM also-ran. Market is pricing zero HBM upside so limited downside | Initial qual passed. Vertical integration (own 4nm logic die) reduces supply chain risk vs SK Hynix/TSMC dependency |
| Foundry losses widen | MEDIUM | 40-50% | ₩1.6T loss in Q1. If foundry bleeds ₩2-3T/quarter, it drags consolidated numbers. Capex wasted competing with TSMC | HBM4 logic die provides internal demand for foundry. Samsung may rationalize non-HBM foundry capex |
| Korea discount is permanent | MEDIUM | 50%+ | Chaebol governance, cross-shareholdings. Discount may never close regardless of fundamentals | None. This is structural. Requires higher margin of safety on entry |
| Geopolitical disruption | HIGH | 15-25% | Middle East → helium shortage (critical for fabs). China tensions → Xi'an NAND fab at risk. Tariff uncertainty | Diversified fab footprint. Helium risk shared across all chipmakers, not Samsung-specific |
Kill level: DRAM contract prices flat or declining for two consecutive quarters. Watch TrendForce quarterly data.
Catalyst Stack
Analyst Commentary, Edgewater Research (Apr 14, 2026)
MU: $426.56, OUTPERFORM · SNDK: $952.50, NEUTRAL
Headline: DRAM Price Hikes Moderating in C2H26, NAND Forecasts Raised
Suppliers are messaging that DRAM supply is still incredibly tight, but they recognize how disruptive the price increases have been. They say they will choose to slow the pace of increases in C2H26. Samsung appeared to want to lock in prices early in C2Q to calm the market.
DRAM Supply/Demand
CY26 DRAM bit demand: mid-20s% YoY growth. Production: ~17-19% YoY. Shipments: +20-22% YoY.
Hyperscale continues to lead demand, with bit demand projections at 2x YoY for CY26 and CY27.
NVDA LPx bit demand roughly equal to all other US hyperscaler bit demand combined, projected to grow 3x in CY27.
US hyperscaler RDIMM addressing projected at ~20% of total DRAM supply in CY26. NVDA LPDDR demand for AI servers: ~25-33% of total DRAM.
MU projected to be the largest LPx/SOCAMM2 supplier into NVDA to support Vera ramp in CY26.
Samsung expected to serve all supply shortages on 255 96GB RDIMMs, while 256GB RDIMMs demand appears to have materially accelerated.
OEMs vocal that rising prices will negatively affect end demand beginning C2Q, worsening in C2H26.
HBM Update
HBM4 quals on NVDA Rubin: feedback remains mixed. All suppliers facing yield challenges on the raised NVDA spec, likely requiring additional work or even a respin.
Rubin production decisions have declined over past few months, largely or entirely offset by increases to Blackwell builds.
Despite feedback that NVDA Rubin HBM4 quals are not yet completed, Hynix is likely the largest HBM4 supplier into NVDA Rubin with share projected at 60-70%.
Samsung forecast at 20%+ share of Rubin with MU share potentially below 20%.
Samsung is expected to remain the lead supplier for HBM into AMD & CY26 including HBM into AMD MI455 series with MU expected to be the primary 2nd source.
HBM3e CY26 pricing not yet set but broadly higher. HBIM packing 40%+ above orig, expectations HBM3E1 pushing higher on NVDA.
Customers buying as much HBM as possible in CY26 to prep for much higher prices in CY27.
NAND Update
CY26 NAND demand forecasts: ~25-28% YoY growth. Supply: ~17-19%. Production: ~17-18%.
CMIT expected to have low PC LPx qual samples available in CY26 but unwilling to commit to any volumes.
CMIT datacenter supply was aimed to cut marketable 64GB DDR RDIMM to C4Q with 96GB volumes expected to grow QoQ throughout CY26 to potentially 20%+ of volumes by CY27.
Feedback suggests YMTC will also have mobile LPx product in the market in CY26, likely in very limited volumes.
Projected Memory Configs by NVDA and AMD Datacenter GPUs
| GPU | HBM Config | DRAM | Notes |
|---|---|---|---|
| H200 | 144GB HBM3E (24GB × 6 stacks) | Hynix primary, MU 2nd source | |
| B200 | 192GB HBM3E (24GB × 8 stacks) | Hynix lead, MU 2nd source | Samsung qual'd, low sids |
| B300 | 288GB HBM3E (36GB × 8 stacks) | Hynix lead, MU 2nd source | |
| V200 (Vera) | 288GB HBM3E (36GB × 8 stacks at 12hi) | Ramp mid C2H26 | Hynix and still expected CoQ, others potentially C2Q |
| V300 Ultra | TBD (64GB HBM4, 16 stacks at 16 Hi) | TBD | |
| RTX 6090 PRO | 96GB 6e (12 × 8) | Samsung lead | |
| AMD MI350x | 192GB HBM3E (24GB × 8 stacks) | Samsung lead | |
| AMD MI355x | 288GB HBM3E (36GB × 8 stacks) | Samsung lead, limited Hynix | |
| AMD MI400x | 432GB HBM4 (36GB × 12 stacks at 12hi) | Samsung led, ramp starting C2Q mid CY26 |
Key Signals for Samsung
Samsung expected to serve all supply shortages, positioned as swing supplier across RDIMM, HBM, and LPDDR.
Samsung capacity est at 12-14B with their actual output est on 10-11.6.
Samsung and MU have been noted in price actions as having less appetite for supporting additional HBM demand but the PC/server DRAM sales offering better margins.
Feedback suggests they also be known to limit their HBM production supply and margins: are viewed as near-likely inflexion for all three suppliers versus standard DRAM.
Samsung has been previously noted as having shifted some planned HBM capacity to standard DRAM starting in CY26, at least partly due to challenged HBM4 yields, and feedback suggests MU would prefer to service via range of HBM customers rather than lock in with the largest buyer at less attractive pricing.
China brokers are back full on Samsung. Spot price moved fast, they're in momentum play.
Analyst Commentary, Cleveland Research (Apr 14, 2026)
MU: OUTPERFORM · SNDK: $991.72, NEUTRAL
Headline: DRAM ASPs Strong in C2Q, Likely Decel in C2H. NAND Increases Higher.
Key Takeaways
DRAM ASPs +45-55% Q/Q in C1Q. Expected to moderate to +10-15% Q/Q in C2H but remain positive into C1H27.
Demand feedback: supply tight in CY26 and CY27, inventory low. CY27 demand sentiment positive.
HBM shipments up 10-12% higher in price in C1Q6. Demand even higher. HBM ASPs likely to see further increases in CY27.
Further upside to NAND ASPs in C2Q & C2H26 as eSSD demand strengthens tied to AI capex.
Server DRAM: 128/256GB products following suit. PC D5 ASPs expected to close below $2 in 2Q.
Customers buying as much HBM as possible, prepping for much higher CY27 prices.
MU Estimates (Cleveland vs Consensus)
| Period | Cleveland Rev | Street Rev | Cleveland EPS | Street EPS |
|---|---|---|---|---|
| FY26Q2 (Mar) | $34.58B | $33.63B | $20.26 | $12.25 |
| FY26Q3 (Jun) | $41.58B | $39.48B | $24.11 | $22.85 |
| FY26Q4 (Aug) | $112.83B | $107.18B | $61.98 | $67.80 |
| FY27 | $182B | $135.28B | $116+ | $66.21 |
Supply/Demand Dynamics
Hyperscale/Server driving ongoing demand strength. Conventional DRAM ASP feedback suggests healthy Q/Q growth through CY26 & CY27 but at a more modest pace vs. 60 days ago.
Inventory remaining lean. Fulfillment far below target across most industries/segments.
Large customers seeking 3-5 year LTAs with DRAM suppliers including prepayments and price floors.
CY27 HBM ASPs expected to grow meaningfully, supported by AI/HPC strength and higher wafer trade rates.
NAND feedback points to upside in 2H26 ASP forecasts driven by eSSD demand tied to AI capex.
NAND expected to grow 10-20% Q/Q in C2Q, moderating in 4Q as vendors try to limit demand destruction.
SNDK / NAND Outlook
NAND ASPs higher in C2Q & C2H26. Server demand expected to uptick in C2H on AI capex & UR ramp.
Jun-Q pricing: $3.25+/GB expected, following Mar-Q finishing at $2.25+/GB.
eSSD demand continues to outpace supply through C2H26. Near-term potential for NAND margins to reach 44-46% by C2 end.
HDD shortages expected to persist into 2028 and potentially beyond, drives eSSD substitution.
Cleveland FY26 SNDK estimate: revenue $32.48B (+96% Y/Y) vs street $26.58B. EPS $112.83 vs street $105.75.
PC & mobile orders appear largely intact near-term despite growing CY26 caution.
Position Sizing & Timing
⚠️ Watchlist. Not a position yet.
Memory cycles are hard to time. Buying at peak margins has historically been a trap. Need to assess whether the AI-driven structural demand shift genuinely extends the cycle or just makes the peak higher. Full Q1 earnings (April 30) will reveal the margin structure in detail.
Preferred entry: Pullback to ₩75-80K on any macro scare or tariff headline. At that level, you’re paying <2x annualized OP for a business generating $150B+ annual run-rate. Vehicle: KRX:005930 directly (Korean brokerage), SSNLF (OTC), or LSE:SMSN (GBP). This is a cyclical bet with a structural kicker, not a compounder.