Goldman keeps KOSPI 12,000 target as DRAM shortage runs to 2030

Goldman Sachs is holding its 12-month KOSPI target at 12,000, arguing the memory selloff priced in a gloomier cycle than the data supports, with DRAM supply shortages set to run to 2030 on AI compute demand.

"The market's current pricing embeds a degree of fundamental pessimism that exceeds reality," Justin Park, a trader at Goldman Sachs Asia in Seoul, said in an Aug. 7 report.

KOSPI fell 39 percent from its June 22 peak before surging 17.9 percent on July 31, the index's biggest single-day gain on record. The rebound followed a shakeout in leveraged exchange-traded funds, falling margin exposure, and a pullback in hedge-fund positioning that left the market's positioning cleaner.

The call rests on a structural supply-demand imbalance: AI compute demand should keep DRAM tight through 2030, underpinning pricing power. Samsung Electronics reclaimed the top DRAM spot with 39 percent share in the second quarter, while SK Hynix slipped to 26 percent and Micron narrowed the gap to one percentage point.

Goldman pushes back on three bearish worries

Goldman's bullish stance runs against three concerns that drove the selloff. First, Nvidia is evaluating lower-spec HBM configurations for its "Rubin Ultra" chip — an 8-layer HBM4e, 12-layer HBM4, or 8-layer HBM4 replacing the planned 12-layer HBM4e baseline, per TrendForce. Goldman reads the shortage itself as proof of structural imbalance: if Rubin Ultra ships with lower-spec memory, customers running large-language-model workloads will need more GPUs, pushing overall compute demand higher.

Second, SK Hynix's early long-term supply agreements locked a chunk of capacity into older HBM3E lines, limiting its flexibility to shift toward conventional DRAM and next-generation HBM4 for AMD, Meta, and Google. That capacity lock already shows in market share: SK Hynix's DRAM share fell to 26 percent in the second quarter, while Samsung rose to 39 percent. Goldman said SK Hynix's next phase of competitiveness depends on how quickly it can switch production lines.

Third, SanDisk's earnings beat expectations but disappointed the most optimistic forecasts. The company reported fourth-quarter revenue of $89.7 billion, above the $83 billion to $85 billion consensus, but guided first-quarter revenue of $10.3 billion to $10.8 billion, slightly below the $108.2 billion some expected. Consumer and edge-computing demand — smartphones and PCs — fell 32 percent quarter over quarter, with management seeing no meaningful recovery until 2027.

Demand signals turn constructive

Two demand-side signals support the bullish case. ChangXin Memory Technologies, China's DRAM leader, has refused Apple's price-cut requests because demand from Huawei and Xiaomi remains strong, with CXMT pricing at parity with Samsung and SK Hynix. Separately, AI startup DeepSeek plans a "significant" increase in API prices, marking an end to the era of heavily subsidized AI inference — a shift Goldman ties to DeepSeek's capital needs ahead of a potential IPO in late 2026 to 2027 and a 1-gigawatt data center in Inner Mongolia.

Goldman also argues DRAM's structural constraints are tightening. NAND can expand capacity by adding stacking layers — potentially reaching 500 to 600 layers in two to three years — but DRAM's process-shrink headroom is nearly exhausted. The 1c node is effectively at 11 nanometers, and 10 nanometers may be the last achievable node, meaning yield declines and rising capital expenditure for EUV equipment and cleanroom expansion structurally support the memory cycle.

The overweight call on Korea carries risk: the 39 percent drawdown and record single-day rebound show how violently positioning can swing, and SK Hynix's share loss plus Nvidia's HBM spec review are live threats to the thesis. Goldman's 12,000 target implies meaningful upside from current levels if the DRAM shortage holds through the decade.

This article is for informational purposes only and does not constitute investment advice.