Samsung Electronics released its second-quarter results on July 30, completing the latest earnings set for the world’s three largest memory chip makers: Samsung, SK hynix and Micron Technology. All three reported the best results in their histories.
That has not translated into record share prices. Over the past month or so, all three stocks have pulled back sharply as investors weigh a familiar risk: whether memory is heading back into the old pattern of rising prices, capacity expansion, oversupply and a steep downturn.
Based on public disclosures, capital spending is increasing across the industry, but the approach looks more restrained than in prior upcycles. The focus is on AI-related products rather than broad capacity additions across every segment.
SK hynix, for example, is expected to spend close to KRW 50 trillion, or about $35 billion, in 2026. Nearly all new advanced-node capacity is being directed to HBM, or high-bandwidth memory, and DRAM used in AI servers. Supply available to consumer electronics is tightening as capacity shifts toward AI.
An investor who has followed the memory industry for years told the Economic Observer that the three suppliers are doing two things at once: expanding in AI-related categories and signing three- to five-year supply contracts with downstream customers. Those contracts include floor prices and take-or-pay clauses, meaning buyers must pay the contracted amount whether or not they take delivery. Buyers back the commitments with prepayments and deposits.
Micron management said on its June 25 earnings call that the gross margins implied by its 16 signed five-year strategic customer agreements, measured at floor prices, are above the peak levels seen in any previous cycle.
The companies have also continued to raise guidance for the next quarter. Micron gave revenue guidance with a midpoint of $50 billion, well above the prior market expectation of $43.2 billion, and said the supply-demand outlook through 2027 had not weakened. Samsung management said on its July 30 call that meaningful new supply is unlikely before 2028 because a new fab takes more than three years to move from groundbreaking to volume production, and capacity from projects already under construction cannot be realized quickly. On July 29, SK hynix management said AI-related demand will keep accelerating in the second half, with third-quarter DRAM shipments expected to rise by a mid-single-digit percentage sequentially, and it also said the supply-demand outlook through 2027 had not softened.
The supply gap remains the foundation of supplier leverage
The reason the three companies can secure floor prices, take-or-pay terms and large prepayments is straightforward: buyers have little room to bargain in the current market.
Micron management said on its second-quarter earnings call that market tightness should extend beyond 2027. Kim Jae-joon, executive vice president of Samsung’s memory business, said on July 30 that the supply gap in 2027 will be more severe than in 2026 and that meaningful new supply is not expected before 2028.
Goldman Sachs, in an August 4 report titled Answering the Eight Core Questions on the Memory Industry, estimated that the global DRAM supply-demand gap in 2026, defined as supply minus demand as a share of demand, will be -4.9%, the worst imbalance in the past 15 years. It expects the market to tighten further in 2027.
A supply-chain source close to SK hynix told the Economic Observer that suppliers have clearly accelerated expansion over the past month and that long-term contracts with U.S. customers have improved visibility into future demand. Even so, capacity plans through the first half of 2028 were already locked in, and output tied to newly added projects will not begin to come online until the second half of 2028. The main bottleneck is EUV lithography equipment, the only tool used to manufacture advanced memory chips, and delivery times have stretched beyond two and a half years.
The same source said SK hynix has publicly targeted a doubling of capacity by 2030 from the level expected at the end of 2025, implying annual growth of roughly 16% to 17%.
Tool delivery is one constraint. Allocation of existing capacity is another. Lin Meibing, chief analyst at ICTIME, told the Economic Observer that HBM requires about three times as many process steps in lithography and etching as conventional DRAM. Each additional 10,000 wafers of monthly HBM capacity consumes the equivalent of roughly 30,000 wafers of monthly conventional DRAM capacity through conversion. Industry-wide HBM capacity currently totals about 326,000 12-inch equivalent wafers per month, with Samsung at about 140,000, SK hynix around 150,000 and Micron about 36,000. That is still not enough.
Wang Xudong, a semiconductor analyst at Sigmaintell, said AI servers already account for about 50% of global DRAM wafer capacity in 2026, and that share is expected to reach about 60% in 2027.
Each extra wafer put into HBM effectively reduces supply available for conventional DRAM by three wafers, reinforcing shortages on both sides. In prior profit peaks, the major suppliers tended to expand across multiple directions at once. This time, expansion is concentrated on AI, which makes the consumer-side supply gap more durable.
The profit picture is also less intuitive than many expected. The supply-chain source close to SK hynix said HBM3E is currently priced at about $12 to $13 per GB, while HBM4 scheduled for shipment starting in the second half is priced at about $16 to $19 per GB. Both products, however, are priced below DDR5 in 2026.
Gross margins for standard DDR products are generally above 80% across the industry. HBM margins are lower, around 50% to 70%, because of yield losses tied to packaging and TSV, or through-silicon via, processing.
Goldman Sachs expects average conventional DRAM pricing to reach about $2/Gb by the end of 2026, up from $0.5 to $0.6/Gb at the end of 2025. As of the second quarter, conventional DRAM was already priced above HBM on a per-GB basis. The bank said HBM blended average selling prices would need to rise 87% to 100% year over year in 2027 to regain a premium over standard DRAM.
The source close to SK hynix said HBM could see a sharp catch-up in pricing in 2027 because the current annual negotiation model has not fully reflected market conditions.
Large buyers are already changing product design as a result. TrendForce said on August 4 that Nvidia has, starting in the third quarter this year, shifted the HBM configuration for its next-generation Rubin Ultra GPU from HBM4E 12-high stacks to parallel evaluations of reduced-spec alternatives including HBM4E 8-high, HBM4 12-high and HBM4 8-high. TrendForce also said Nvidia decided to halve the memory capacity of its next-generation superchip module because LPDDR5X remains in short supply.
TrendForce expects HBM bit shipments to rise 50% to 60% year over year in 2027, but still fall short of demand. If even Nvidia is cutting specifications to cope with limited memory supply, other buyers facing a supply gap that may last at least until 2028 have one practical option: sign early and lock in supply.
Three- to five-year contracts are replacing quarterly bargaining
For years, memory chips were mainly sold through quarterly pricing negotiations.
At the start of each quarter, suppliers set a temporary shipment price. The final settlement price was then fixed in the last month of the quarter. Pricing moved with real-time supply and demand, and both sides optimized around a short negotiation window. That structure is one reason memory prices have historically been much more volatile than those of other semiconductor products, and why investors long treated memory stocks as classic cyclicals.
The three suppliers are now moving away from that system and toward contracts that run three to five years.
According to reporting in the article, these contracts typically set supply volumes for three to five years, revisit unit prices once a year, and define a pricing corridor with a ceiling and a floor. Buyers support the commitments with take-or-pay clauses, prepayments or deposits, and some agreements also include minimum revenue provisions.
Several people interviewed by the Economic Observer said annual supply agreements had appeared occasionally in the past, but they usually ran for no more than one year, covered only a small portion of demand, and did not include floor prices or take-or-pay provisions. The new contracts are materially different in duration, scale and enforceability.
Micron said it signed 16 strategic customer agreements in the third quarter of fiscal 2026. They run for five years and use take-or-pay terms. Fourteen include floor prices, and the value calculated at those floor prices totals about $100 billion. Micron expects about $22 billion in cash deposits and letters of credit as performance support, including about $18 billion in cash and $4 billion in letters of credit. Its target is for these agreements to cover more than 40% of revenue over time.
Samsung’s intended coverage is even broader. Kim Jae-joon said on July 30 that Samsung plans to allocate 60% to 70% of capacity to multi-year contracts, and that the ratio could rise as more customers seek to sign. The structure uses a five-year base with annual rolling renewals. Samsung said the top five global data center customers have all signed, another five large AI customers are in final negotiations, and the company has already received about one quarter of the agreed prepayment total.
On July 29, Park Joon-deok, head of DRAM marketing at SK hynix, also confirmed that negotiations with about 10 core customers had been completed, though he did not disclose the share of total sales that will be covered.
Goldman Sachs said in a July 29 report that more than half of server DRAM is now covered by long-term agreements.
The change is not limited to DRAM. NAND supplier Sandisk said on August 5 that it had signed multi-year contracts with eight customers under what it called a “new business model.” The weighted average term is more than four years, revenue calculated at floor prices totals $93.9 billion, and associated financial guarantees amount to $16.5 billion. Chief executive David Goeckeler said these agreements are expected to cover more than 50% of Sandisk bit capacity in fiscal 2027 and about two-thirds in fiscal 2028.
Why would buyers accept terms like these? The article points to the same answer again: the supply gap is expected to last long enough that waiting for prices to fall is no longer a viable strategy.
A product manager at a large memory module maker in Shenzhen said server memory is in severe shortage across both CPU and GPU systems, to the point that some servers must ship with reduced specifications. The priority for customers is obtaining supply, not negotiating a better price.
The pricing structure itself also shows how much leverage has shifted to suppliers. BofA Securities said in an August 1 report that Samsung has limited price cuts under long-term contracts to no more than 5% sequentially, while allowing upward revisions of 10% to 20% or more with no cap. That gives the suppliers upside when the market rises and income protection when the market weakens.
Lin Meibing said several large U.S. cloud customers that have signed long-term deals have already largely locked in 2027 supply prices, with contract durations typically extending to 2030 and in some cases 2035. If shortages continue through 2027, he said, suppliers have little reason to offer lower prices. He sees the chance of a meaningful pullback only after 2028, when more industry capacity starts to come online.
After reviewing these contracts, Goldman Sachs argued that duration, coverage, pricing mechanisms and performance obligations are all moving in suppliers’ favor. In the bank’s view, that business model shift could extend the period of high profitability and lift valuation multiples for memory suppliers from roughly 5 to 6 times to about 8 to 10 times.
On an August 5 earnings call, David Goeckeler said the market was still operating on quarterly pricing three or four quarters ago. Sandisk now has more than four years of demand visibility, and some customers returned one quarter after signing to add purchases for future years. “Before, this was just a quarterly supply-chain price negotiation,” he said.
These contracts mainly cover conventional DRAM and NAND flash. HBM is sold differently. Because each generation must be custom-validated for specific AI chips, suppliers usually negotiate annually with core customers such as Nvidia, locking in the following year’s price and volume in advance.
Micron management confirmed on its earnings call that all of the company’s HBM supply for full-year 2026 has already sold out and both pricing and volume have been set.
Prices are still rising, but the slope is easing
DRAM contract prices are still increasing, though the pace has slowed from quarter to quarter. TrendForce data cited in the article show that DRAM contract prices rose about 90% sequentially in the first quarter of 2026 and about 60% in the second quarter. Several industry people said the third-quarter increase is likely to narrow again to 13% to 18%.
The rate of increase is cooling. The debate is over how far the industry still is from a peak.
Lin Meibing said memory price gains should narrow materially after the middle of 2027, with room for pullbacks in 2028 as new capacity comes online.
The supply-chain source close to SK hynix said supplier inventories currently stand at about two to four weeks, below the normal level of four to five weeks and far below the 10-plus weeks that usually appeared before earlier downturns. In that view, there is still no sign of a supply-demand reversal, and any substantial price correction is more likely after the second half of 2028.
Even if prices begin to ease after 2028, industry participants quoted in the article expect the decline to unfold very differently from prior cycles.
The Shenzhen-based memory module manager said consumer electronics would feel the first signs of softer pricing, followed by niche products such as smart speakers and set-top boxes, then servers. Automotive electronics would adjust last because contract durations are longest there and the market is smaller.
That would be a break from past downturns, when prices across the memory market could collapse within months. This time, the three large suppliers may have staggered adjustment cycles across customers and product lines through the use of long-term contracts.
CXMT and domestic Chinese suppliers are the next major variable
The article says the global memory market will face another structural variable after 2028: ChangXin Memory Technologies, or CXMT.
CXMT listed on Shanghai’s STAR Market on July 27 and raised about RMB 57.9 billion. Its market capitalization at one point approached RMB 4 trillion after listing. Based on Omdia shipment data, the company held a 7.67% share of global DRAM sales revenue in the fourth quarter of 2025, ranking fourth worldwide.
According to the report, CXMT’s production cost and end-market quotations are both about 10% below comparable overseas products. As the three global leaders redirect advanced-node capacity toward AI, the consumer electronics share they leave behind is being filled by CXMT.
Lin Meibing said Chinese memory suppliers are now broadly on par with overseas leaders in mainstream DDR and LPDDR products, with no meaningful generational gap. Revenue from server products has risen quickly for Chinese DRAM suppliers in recent years, and DDR5 shipments have driven that structural shift.
HBM remains the most difficult technology challenge. Lin said the three overseas leaders can already mass-produce HBM3, and some have completed volume-production verification for HBM4. Chinese suppliers, by contrast, are still ramping and validating HBM3, leaving a gap of about two generations. At the same time, he said Chinese companies do not carry the burden of legacy production lines, and their inability to buy EUV tools has pushed them to invest in new routes such as VC-T and wafer bonding, leaving open the possibility of narrowing the gap in next-generation architectures.
The article also said CXMT is working with a domestic foundry focused on mature nodes on CBA, a hybrid bonding packaging approach in which memory and logic chips are manufactured separately and later bonded together. CXMT would make the memory wafers while the partner fabricates the logic wafers, forming a domestic HBM manufacturing chain.
A person close to CXMT said its LPDDR6 product is near the end of R&D verification. The first design targets 12800Mbps using 16Gb dies and could enter mass-production introduction in the second half of 2026.
Wang Xudong said Chinese device brands and AI server makers are increasing the share of domestic DRAM in their procurement mix to secure supply chains. Domestic substitution in general-purpose DRAM and consumer NAND should keep expanding, though high-end HBM and enterprise AI memory remain hard to replace at scale in the near term.
Consumer electronics are under pressure, but new demand is forming
Pressure on consumer devices is already changing the industry’s demand structure.
The memory module manager said full-year 2026 smartphone memory demand is expected to be revised down by 15% to 20% year over year. In the second quarter, some handset makers sharply reduced purchases because they could not accept Samsung’s pricing. Consumer memory, in that person’s view, may stop rising by the fourth quarter.
On July 30, Apple chief executive Tim Cook described the current memory price surge on the company’s earnings call as a “once-in-a-century flood.” Apple’s product gross margins are under pressure from higher memory costs. Chief financial officer Kevan Parekh said more than 100% of the sequential decline in gross margin over the past two quarters could be explained by changes in memory costs, and Apple lowered its gross margin guidance for the next quarter to 47% to 48%.
Wang Xudong estimated that in a budget smartphone with a 4GB+128GB configuration, storage cost as a share of the bill of materials has climbed from 22% in the third quarter of 2025 to 64% in the third quarter of 2026. He expects smartphones priced below RMB 1,500 to become hard to find in 2027.
But while consumer demand is tightening, fresh sources of demand are also building.
Lin Meibing said automotive-grade memory is growing strongly in China. Faster penetration of ADAS is driving much higher memory-capacity needs in smart cockpits and autonomous-driving domain controllers, and year-over-year growth in automotive memory demand in China is expected to reach 70% to 80%.
CXL, the interconnect technology that supports pooled and shared memory across servers, is expected to begin scaling around 2028 as Nvidia and Google adopt it. Samples already exist, but mass production has not started. The first industry standard for HBF, a high-bandwidth flash layer positioned between HBM and SSDs, was jointly released on August 4 by SK hynix and Sandisk, with shipments expected by the end of 2027.
If those new demand drivers gain scale after 2028, they could offset part of the supply loosening created by new memory capacity.
The Shenzhen-based product manager said that even though the slope of price increases is narrowing quarter by quarter, the three big suppliers could still post larger profits in 2027 than in 2026. The market debate is centered on when the cycle peaks, but for the suppliers, a more important variable may be how much revenue can ultimately be put under long-term contracts.
Micron is targeting more than 40%. Samsung is already at 60% to 70%. If the industry’s average long-term contract coverage settles above 50%, the manager said, then more than half of supplier revenue would remain locked in even if the spot market turns volatile, and earnings swings would be much smaller than in any previous cycle.
The report was originally published by the Economic Observer and written by Zheng Chenye.