Over the past month, shares of Samsung Electronics and SK Hynix have fallen by 23% and 35% respectively. Yet, Goldman Sachs has reiterated its Buy ratings for both companies. The firm's report argues that HBM price re-negotiations in 2027, the expansion of long-term supply agreements (LTAs), and low supplier inventories are expected to reduce profit volatility in this memory cycle and enhance the visibility of future revenues and earnings.

The market has not fully embraced this logic. Investors are concerned about weakened memory price expectations, insufficient transparency in LTA terms, rising inventories at module makers, increasing supply from China, and the possibility that capacity expansion will once again lead to oversupply. SK Hynix's Q2 operating profit missing expectations has also heightened market doubts about the sustainability of earnings.

Goldman Sachs's response can be summarized in three points: HBM will re-establish its price premium over conventional DRAM in 2027; customers are pre-booking capacity through multi-year agreements; and low inventories coupled with enterprise SSD demand reduce the near-term risk of the industry swinging back into oversupply.

Behind the $2.9/Gb: HBM Set to Regain Its Premium

HBM is a key memory component for AI accelerators, and its price directly impacts the earnings leverage of Samsung and SK Hynix over the next two years.

Goldman Sachs forecasts that the blended average selling price (ASP) for HBM at both Samsung and Hynix will approach $2.9 per Gb in 2027, representing year-over-year growth of 87% and 100%, respectively. Within this, price increases for main products on a like-for-like basis are estimated at around 60%, with the remaining growth coming from product mix improvements, including a higher share of newer generation and higher-stack HBM.

The basis for the 2027 price hike stems first from supply and demand. AI server demand is still likely to outpace HBM supply, while the manufacturing difficulty of the latest-generation HBM is increasing. Core dies and base dies are adopting more advanced processes, and higher-layer stacking also increases process complexity and depresses yields. The newest HBM consumes more wafer capacity, further constraining effective supply.

Another reason is the changed price relationship between HBM and conventional DRAM. HBM prices are typically negotiated annually and remain relatively fixed within a year, whereas conventional DRAM is priced monthly or quarterly, reflecting market changes faster. By Q2 2026, conventional DRAM prices had already surpassed HBM. Goldman Sachs expects the conventional DRAM ASP to rise to around $2 per Gb by the end of 2026, far above the $0.5-$0.6 per Gb at the end of 2025.

For HBM to maintain the profit margins and price premium expected of a premium product, a re-pricing in 2027 appears necessary. The Visible Alpha consensus for SK Hynix's 2027 HBM ASP is approximately $2.3 per Gb; Goldman's forecast of $2.9 is about 24% higher.

SK Hynix's blended HBM ASP is projected to rise to $2.9 per Gb in 2027, re-widening the price gap with conventional DRAM.

The revenue contribution from HBM will also increase accordingly. Goldman Sachs expects HBM's share of DRAM revenue at Samsung and Hynix to rise from 8% and 14% in 2026 to 16% and 22% in 2027, reaching 18% and 25% in 2028. If the price and volume materialize, AI-related memory will become a more significant and higher-visibility part of the profit structure for both companies.

3 to 5-Year LTAs Locking Volume: Customers Pre-Book Capacity

The annual pricing of HBM explains the room for price increases in 2027, while multi-year LTAs address why this memory cycle could be more stable.

In traditional memory cycles, customers often place concentrated orders during tight supply and rapidly cut purchases when demand weakens. Suppliers, expanding capacity based on high-demand expectations, are left to shoulder the risks of price declines and inventory build-ups alone. The new round of long-term agreements is beginning to share some of this risk upfront with customers through longer contract durations, higher coverage, price protection, and prepayments.

Goldman Sachs states that current LTA durations are typically 3 to 5 years, with most agreements based on 5 years and some clients opting for 3 years. Samsung employs a rolling contract structure, extending the contract by one year annually through negotiations, so the actual cooperation period may exceed 5 years.

Coverage is also expanding. Samsung has already signed contracts with the world's top five data center clients and is in final negotiations with another five large clients with AI-related demand. Upon completion, multi-year agreements are expected to cover 60% to 70% of its planned capacity.

Hynix indicated it has completed LTA negotiations with about 10 clients, including key customers. Micron disclosed it has signed 16 strategic customer agreements, covering approximately 20% of its DRAM shipments and about one-third of NAND shipments during the contract periods, and expects such agreements to eventually cover over 50% of its revenue.

The binding nature of LTAs is also strengthening. Current pricing mechanisms mainly include fixed prices, price bands with ceilings and floors, and floor prices. Price bands can reduce severe volatility, while floor prices limit suppliers' downside risk while retaining upside from market price increases.

Some agreements also incorporate take-or-pay clauses, prepayments, and penalty terms for breach. Micron expects to receive $22 billion in cash deposits and related financial commitments; SanDisk disclosed financial guarantees and prepayments exceeding $11 billion. Samsung also stated it has received roughly one-quarter of the total contract prepayments.

Summary of Long-Term Agreement (LTA) terms. Highlights include 3 to 5-year duration, 60% to 70% coverage target, price bands, floor prices, prepayments, and financial guarantees.

For memory makers, securing more customer commitments before expanding capacity can reduce their vulnerability to future demand shifts. Increased capital expenditure may still bring supply pressure, but with a higher proportion of future capacity already covered by contracts, the buffer during price downturns would be thicker.

Low Inventories Temporarily Suppress Cycle Reversal Fears

Recent market concerns about rising inventories at module makers are not unfounded. Relatively weak smartphone and PC demand has indeed prompted some module makers to build inventory.

Goldman Sachs believes this change has a greater impact on market sentiment than on industry fundamentals. The market where module makers operate accounts for only a single-digit percentage of the total memory market, and inventories at more critical suppliers and major customers remain at healthy levels.

As of the end of Q2 2026, Goldman estimates DRAM and NAND inventories at suppliers were at 2 to 4 weeks, below the normal 4 to 5 weeks and far below the 10+ weeks often seen before previous memory downturns. Given that capacity expansion and supply growth over the next 12 to 18 months are likely to remain below demand growth, the low inventory state could persist.

Client-side inventory is also considered close to normal levels. Despite strong procurement over the past few quarters, server memory supply has been largely used for immediate production and has not led to significant hoarding. Low inventories combined with increased LTA coverage make a sudden short-term reversal in memory prices relatively less probable.

Enterprise SSD Catches Demand, NAND Not Yet Headed for Oversupply

HBM's strength does not mean the entire memory industry is risk-free. NAND and conventional DRAM will still be impacted by weak consumer electronics demand from smartphones and PCs.

Goldman Sachs judges that the primary incremental demand for NAND is shifting toward AI servers and enterprise SSDs. The report forecasts enterprise SSD demand for 2026-2028 at 474EB, 619EB, and 755EB, representing year-over-year growth of 66%, 31%, and 22%, respectively. Even if consumer demand remains soft, server demand could offset part of that pressure.

Supply-side expansion is also relatively restrained. Major players' capital expenditure focus is more on DRAM, while NAND investment leans toward technology migration rather than large-scale wafer capacity additions. Goldman therefore expects mid-term NAND supply growth may still lag demand growth.

Recent weakness in NAND spot prices is also concentrated in specific products like TLC 512Gb, while other products like TLC 1Tb have shown relative stability. The price of TLC 512Gb had risen nearly 600% over the past year, and the recent correction occurred after it significantly outperformed other products, so it cannot be directly interpreted as the entire NAND market turning toward oversupply.

Enterprise SSD demand and year-over-year growth rate. Demand for 2026E, 2027E, and 2028E is 474EB, 619EB, and 755EB, with growth of 66%, 31%, and 22% respectively.

Goldman's supply-demand model also indicates tightness. For 2027, the DRAM, NAND, and HBM supply-demand gaps are projected at approximately 5.9%, 4.6%, and 6.0%, respectively, with HBM being the tightest. A supply gap does not guarantee prices will keep rising, but it indicates that NAND has not directly entered oversupply due to soft consumer demand, supported by low inventories and AI server demand.

Supply gaps exist for DRAM, NAND, and HBM in 2027, with HBM supply-demand being the tightest.

CXMT Unlikely to Alter Short-Term Tightness, Remains a Long-Term Supply Variable

The market is also watching the impact of CXMT's capacity expansion on global DRAM supply and demand. Goldman Sachs expects CXMT to primarily expand its market share relying on domestic Chinese demand, while its overseas sales are subject to both commercial and geopolitical factors. Considering the technology gap, CXMT is unlikely to materially alter the tight global memory supply-demand landscape in the short to medium term.

CXMT's current growth in mobile DRAM mainly comes from procurement by Chinese smartphone makers. According to TrendForce data cited in the report, approximately 70% of CXMT's mobile DRAM shipments this year are still LPDDR4(X); Samsung and Hynix are already predominantly on LPDDR5(X), with such products expected to account for 75% to 85% of their mobile DRAM shipments.

CXMT still lags behind Korean players in process technology. CXMT's current main process is still about 2-3 generations behind global leaders, making its near-term expansion unlikely to directly change high-end DRAM supply-demand.

There is also a gap in technology nodes. CXMT's current main process is roughly equivalent to the 1z node, while global leaders are transitioning from 1a/1b to 1c nodes, a gap of about 2-3 generations. Historically, each process node migration typically takes several years, and catching up won't happen immediately with capacity expansion.

Equipment restrictions will also affect its upgrade pace. More advanced DRAM processes require wafer fabrication equipment like EUV, which remains restricted for Chinese firms. Domestic lithography equipment also requires a long R&D cycle to reach EUV-level performance.

The challenge is even greater for HBM. Beyond conventional DRAM processes, HBM involves advanced packaging, data transfer speeds, power consumption, yield, and base die foundry capabilities. For CXMT to build competitiveness in HBM, it would require simultaneous upgrades in China's domestic advanced process and packaging supply chain.

Risks therefore haven't disappeared. If AI server demand underperforms, HBM's price increase potential would narrow; if final LTA terms on price protection, prepayments, or purchase obligations are weaker than current assumptions, earnings visibility would also be discounted; if mature-node supply continues to expand, prices for conventional DRAM and consumer products could still face pressure.

The low valuations of Samsung Electronics and SK Hynix following their pullback reflect investor skepticism about whether this memory cycle can escape the severe volatility of the past. The core of Goldman's bullish stance lies in the simultaneous presence of HBM price hikes, long-term volume locks, and low inventories. What truly needs verification next is whether AI server orders will continue to grow, whether LTA terms will be fully realized, and whether new capacity additions can remain restrained ahead of demand expansion.