In its September 30 report titled "Korea Technology: Memory Price Tracker," Goldman Sachs put forward a judgment: TrendForce's forecasts for memory contract prices in the fourth quarter of 2026 are broadly higher than Goldman's own estimates, with PC DRAM, server DRAM, and NAND all running above, while only mobile DRAM is roughly in line with Goldman. The bank reiterated its "Buy" ratings on Samsung Electronics Co., Ltd. (which is also on its Conviction Buy list) and SK hynix.

Three pricing lines in the fourth quarter

Citing TrendForce's forecasts, Goldman noted that in the fourth quarter, average selling prices for PC DRAM, server DRAM, and mobile DRAM are expected to rise 13%-18%, 10%-15%, and 0%-5% quarter on quarter, respectively; overall NAND is seen up 15%-20% QoQ, with enterprise SSD alone rising 23%-28%. Against Goldman's own model: on an average basis for Samsung and SK hynix, fourth-quarter PC DRAM rises 9% QoQ, server DRAM rises 10%, mobile DRAM rises 3%-4%, and NAND rises 9%-11%. In other words, apart from mobile, which is roughly in line, TrendForce is more optimistic on the other three pockets. Attention must be paid to the level of the metric—according to TrendForce's official press release, conventional DRAM contract prices in the fourth quarter rise 10%-15% QoQ, while overall DRAM including HBM rises 15%-20%; the 13%-18% in Goldman's report refers to the PC DRAM sub-segment, and the two cannot be used interchangeably.

Looking at the first quarter of 2027: TrendForce expects PC DRAM and server DRAM to rise 3%-8% QoQ respectively, with the server side roughly in line with Goldman's 3%-7%; the mobile side will regain momentum, with LPDDR5X and LPDDR4X forecast to rise 13%-18% and 5%-10% QoQ, respectively.

Spot is tighter than contracts: DDR4 premium already at 44%

Looking at September prices, the increases are not only written into forecasts. Citing TrendForce data, Goldman noted that the DDR4 8GB contract price rose 4% QoQ to $148, while DDR5 8GB was flat at $133, widening DDR5's discount to DDR4 to 10% (from 6% in August); on the server side, the DDR4 64GB module was flat at $1,295, while the DDR5 64GB module edged up 1% QoQ to $1,515, with the DDR5 premium rising from 16% to 17%. The real signal is in spot: DDR5 16Gb spot carries a 20% premium to the latest contract price, while DDR4 8Gb spot carries a premium as high as 44%—older-generation products are especially tight as manufacturers gradually exit. Goldman pointed out that PC OEMs, in response to a possible supply contraction in 2027, will maintain aggressive procurement throughout the fourth quarter, which is the main support for prices. The driving mechanism is not complicated: manufacturers continue to prioritize advanced-node capacity for high-performance server products, and the overall supply-demand imbalance remains; on the server side, as CPU supply improves, cloud service providers and server OEMs are adding bit procurement to meet demand for RDIMMs in general-purpose servers under agentic AI scenarios, but constrained by component supply, packaging and testing capacity, and the scheduling flexibility of DRAM front-end processes, the supply mix still cannot match demand. TrendForce also noted that some sellers, due to price-range mechanisms in long-term agreements (LTAs), will see shipment increases lag the market average—a layer easily overlooked when judging manufacturers' actual results.

Enterprise SSD: the only category accelerating upward

Clear divergence has emerged within NAND. According to TrendForce, as cloud service providers continue to ramp up AI inference infrastructure, enterprise SSD bit demand in 2026 is expected to grow more than 80% year on year, with the procurement driver shifting from large-model training to actual AI deployment; the large-scale rollout of agentic AI has multiplied the data volumes for real-time retrieval and caching, and QLC penetration in vector databases has risen in tandem; although manufacturers are expanding supply of QLC and high-capacity models, most of the new capacity has already been locked up by customers in advance, with limited volumes flowing into the open market. Enterprise SSD has therefore become the only memory category whose price gains accelerate in the fourth quarter. Client SSD presents a different picture. According to TrendForce, PC OEMs can rely on finished-goods inventory and channel stock built in the first half to support fourth-quarter shipments, needing only to replenish sporadic shortage items; brand vendors have also lowered SSD capacities in mainstream models to reduce overall bill-of-materials costs, with procurement volumes and average capacity per unit contracting in tandem. With buyers holding inventory and procurement limited, manufacturers' pricing stance has turned more flexible, curbing gains. eMMC/UFS is even weaker—phone brands mostly rely on existing inventory to support fourth-quarter production, with new procurement from manufacturers quite limited.

The mobile side is the only exception, but it will need to catch up in the first quarter of next year. Fourth-quarter mobile DRAM rises only 0%-5%, the weakest category in the table. TrendForce's explanation: high LPDDR costs suppress smartphone output, and manufacturers shifting capacity to server products has tightened handset allocations, depressing bit demand overall; although a price-up pattern is established, the fourth-quarter gains are narrower than the prior quarter because earlier pricing had already front-loaded much of the increase. Goldman attributes this to a structural squeeze of "supply being reallocated from mobile to server." By the first quarter of 2027, this suppression is expected to reverse, with LPDDR5X and LPDDR4X rising 13%-18% and 5%-10% QoQ, respectively.

Tighter in 2027: HBM eats capacity, and cutting specs won't save costs

What truly determines the length of this cycle is 2027. According to TrendForce research published on September 29, continued expansion of AI server demand keeps HBM and conventional DRAM competing for limited advanced-node and wafer capacity, and new-generation products need time to ramp yields and output; as a result, memory supply will remain tight in 2027; the firm has raised its 2027 HBM price outlook, estimating blended ASP to grow 121% year on year, reflecting a higher mix of high-priced HBM4 and the gradual ramp of HBM4e in the second half of 2027. The capacity-side squeeze has clear numbers. According to disclosures by Samsung Electronics Executive Vice President Kim Taewoo on September 29, HBM's share of total global DRAM wafer capacity is expected to rise from about 20% currently to nearly 30% by 2027; HBM and conventional DRAM share the same wafer capacity resources, so expanding HBM output will squeeze conventional DRAM supply. Each HBM unit occupies about three times the wafer area of conventional DRAM, meaning conventional DRAM supply could be squeezed out at two to three times the pace at which HBM's share rises. Notably, the cost-cutting route does not work. GPU and ASIC vendors are discussing reducing HBM from 12-Hi to 8-Hi to lower costs, but TrendForce points out that each HBM unit must carry a Base Die, and that cost does not decline proportionally with fewer layers; 8-Hi has less DRAM capacity over which to spread the Base Die cost, so the converted cost per Gb is actually higher, and in 2027 the per-Gb price of 8-Hi may carry a 10%-20% premium to 12-Hi—reducing specs can partially ease GPU cost pressure but cannot reverse the overall sharp HBM price increase.

Goldman's valuation and the risks it lists

Goldman's 12-month target price for Samsung Electronics is KRW 490,000 for common shares, using 2026-2027 EV/EBITDA sum-of-the-parts (SOTP) valuation; the preferred-share target price is KRW 360,000, corresponding to a 27% target preferred discount (the average of the two-factor model discount and the past month's average discount), with both common and preferred rated "Buy." It lists three downside risks: a sharp deterioration in memory supply-demand, a sharp contraction in smartphone margins, and loss of mobile OLED market share. For SK hynix, the 12-month target price is KRW 3.5 million, using a 2026-2027 average P/E method with a target P/E of 9.0x. It lists four risks: a sharp deterioration in memory supply-demand and delays in technology migration; weakening demand for smartphones, PCs, and servers dragging on conventional memory demand; Samsung's positive HBM progress hurting its HBM revenue and profit; and a decline in AI-related capital expenditure weakening HBM demand.

Why the market is still selling

Interestingly, the more prices rise, the more Korean semiconductor stocks come under pressure. According to Korean media reports, global funds aggressively sold easily liquidated large-cap chip stocks in emerging markets to reduce risk, with foreign investors at one point net selling a combined roughly RMB 49.6 billion of Samsung Electronics and SK hynix over six trading sessions; the sharp short-term strengthening of the won gave foreign investors a chance to sell chip stocks with unrealized gains and lock in foreign-exchange gains, while also triggering concerns that major exporters' third-quarter results would miss expectations. Samsung Electronics previously announced a KRW 15 trillion buyback plan, and SK hynix plans to buy back and cancel KRW 40 trillion worth of shares. On fundamentals, third-quarter results still point to historic highs. According to brokerage consensus compiled by FnGuide, Samsung Electronics' third-quarter operating profit is expected at about KRW 110.3 trillion, up more than eightfold year on year, which would make it the first Korean company to top KRW 100 trillion in quarterly operating profit; SK hynix is expected at about KRW 78.1 trillion, up about 29% QoQ; the two combined total about KRW 188 trillion. Yet disagreement exists—DS Investment & Securities gives more conservative forecasts of KRW 104 trillion for Samsung and KRW 70 trillion for SK hynix, mainly because the won strengthened sharply versus the end of the second quarter (the USD/KRW rate fell about 11%): both companies' revenue is almost entirely USD-denominated, so a stronger won means revenue shrinks when converted back into won, offsetting part of the price-increase benefit.

The latest footnote on the earnings side comes from Micron. According to its fourth-quarter fiscal 2026 report released on September 30, quarterly revenue was $54.23 billion, up about 379% year on year, adjusted gross margin was 87.0%, and adjusted earnings per share were $33.42, all above market expectations; first-quarter fiscal 2027 revenue guidance is about $61.5 billion, also above analysts' expectations of $56.8 billion. After the report, Micron fluctuated slightly in after-hours trading, while the stock has risen about 273% year to date, making it the best performer among Philadelphia Semiconductor Index constituents. The core variable of this memory rally has always been the same: whether AI capital expenditure can continue. As long as hyperscale cloud providers' buildout pace does not stop, price increases can be passed through; but once AI investment cools, the high-priced inventory now locked into long-term agreements could backfire quickly.