Goldman Sachs Reiterates Buy Ratings on Samsung and SK Hynix, Raises Q4 Memory ASP Forecasts Across the Board
nashnova research
Goldman Sachs on September 30 reaffirmed its Buy ratings on Samsung Electronics (Conviction List) and SK Hynix, noting TrendForce's Q4 memory contract-price forecasts run above Goldman's own estimates across every category — enterprise SSDs lead at +23%–28% — while recent foreign-fund selling and the AI capex trajectory remain the key swing factors.
How much are Q4 memory prices rising — and who disagrees?
TrendForce forecasts Q4 contract prices up 13%–18% for PC DRAM, 10%–15% for server DRAM, and just 0%–5% for mobile DRAM quarter-on-quarter. NAND overall is up 15%–20%, with enterprise SSDs alone at +23%–28%.
Goldman's own model is more conservative: PC DRAM +9%, server DRAM +10%, NAND +9%–11%. This means → except for mobile, TrendForce is meaningfully more bullish on every category, and the market may be underpricing the upcycle.
A methodology note matters: TrendForce's headline "general DRAM +10%–15%" and Goldman's "PC DRAM +13%–18%" refer to different sub-segments — the two numbers are not interchangeable.
Why are enterprise SSDs the sole accelerating category?
Cloud providers are doubling down on AI inference infrastructure. Enterprise SSD bit demand is expected to grow over 80% year-on-year in 2026, with procurement shifting from model training to live AI deployment.
Manufacturers have ramped QLC — a flash-memory technology that packs more data into each storage cell — and higher-capacity drives, but most new capacity is pre-committed to customers. In plain terms = the cake is sold before it leaves the oven; remaining buyers must bid up.
Client SSDs tell the opposite story: PC brands hold first-half inventory, and some have cut SSD capacity in mainstream models to save on materials. Buyers are in no rush, so vendors are pricing flexibly and gains stay muted.
Why is mobile DRAM the weakest link — and when does it bounce?
Mobile DRAM rises just 0%–5% in Q4 — the weakest line in the entire table. This means → handsets are nearly sitting out the memory upcycle.
Two forces are at work: high LPDDR — low-power memory used in smartphones — costs are suppressing handset production, while manufacturers are redirecting wafer capacity toward more profitable server products, squeezing mobile allocations.
TrendForce expects a Q1 2027 rebound: LPDDR5X +13%–18% and LPDDR4X +5%–10% quarter-on-quarter. In plain terms = mobile memory is being neglected now, but the price rise is delayed, not cancelled.
What does the spot premium signal?
September contract prices: DDR4 8GB up 4% QoQ to $148; DDR5 8GB flat at $133. The DDR5-to-DDR4 discount widened from 6% in August to 10%.
Spot markets flash a louder signal: DDR5 16Gb spot trades at a 20% premium to the latest contract, and DDR4 8Gb spot commands a 44% premium. This reflects legacy products tightening fast as fabs phase them out.
PC OEMs are expected to keep buying aggressively through Q4, bracing for a possible 2027 supply crunch. On the server side, improved CPU availability lets cloud and server OEMs add RDIMM purchases for agentic-AI workloads, but back-end packaging capacity and front-end process flexibility still cannot match demand.
How will HBM squeeze conventional DRAM in 2027?
Samsung EVP Kim Taewoo disclosed that HBM — high-bandwidth memory, a high-speed stacked memory built for AI chips — is expected to rise from roughly 20% of global DRAM wafer capacity today to nearly 30% by 2027.
Each HBM die occupies about three times the wafer area of a conventional DRAM die, so conventional supply gets crowded out two to three times faster than HBM's share rises. This means → every bite HBM takes out of capacity hits traditional DRAM with a multiplied squeeze.
GPU and ASIC makers are discussing cutting HBM from 12 layers to 8 to save money, but TrendForce notes each HBM package requires a base die whose cost does not shrink proportionally with fewer layers. Result: 8-layer HBM costs 10%–20% more per Gb than 12-layer. In plain terms = stacking fewer layers looks like a saving, but the fixed base-die cost spreads over less capacity — unit price actually goes up. Downgrading specs won't reverse the HBM price surge.
What are the valuations — and where is the risk?
Goldman's 12-month target for Samsung Electronics ordinary shares is ₩490,000 (sum-of-parts EV/EBITDA), preferred shares ₩360,000, both Buy. For SK Hynix, the target is ₩3,500,000 (9.0× target P/E).
Yet strong fundamentals coexist with share-price pressure: global funds recently dumped emerging-market mega-cap chip stocks, with net selling of Samsung and SK Hynix totaling roughly ¥49.6 billion over six trading days; near-term won strength also raises concern about Q3 earnings translation losses.
This reflects the single variable that has defined this memory cycle throughout: whether AI capex can sustain its pace. If hyperscaler buildout slows, the high-priced inventory locked into long-term agreements could snap back fast.
市场有风险，内容仅供研究参考，不构成投资建议。