With the release of Samsung Electronics' second-quarter financial report on July 30, the latest financial quarter reports for the world's three major memory chip manufacturers—Samsung Electronics, SK Hynix, and Micron Technology—are now complete. Each company has set new records for their highest-ever quarterly performance.
However, the record-breaking performance did not lead to record-breaking stock prices. Over the past month or so, the share prices of these three companies have all experienced significant pullbacks. Market views suggest that the investors' selling logic stems from concerns that this round of industry prosperity might repeat the "price hike—capacity expansion—oversupply—steep drop" pattern—the "death cycle" that has recurred over the past several decades.
Based on publicly available information, although the capital expenditures of the big three are indeed increasing significantly, having experienced multiple cyclical ups and downs, they appear much more restrained in this round of expansion—primarily focused on the AI direction, not a comprehensive capacity increase. For instance, SK Hynix's capital expenditure for 2026 is projected to be nearly 50 trillion KRW (approximately $350 billion), with almost all of the new advanced node capacity directed towards HBM (High Bandwidth Memory) and AI server DRAM (Dynamic Random-Access Memory). Supply for the consumer electronics end has, in fact, tightened further as capacity shifts towards AI.
A professional investor who has long followed the memory industry told a reporter from the Economic Observer that while the big three are concentrating on expanding AI-related capacity, they are also doing another thing—each signing three-to-five-year long-term supply contracts with downstream customers. These contracts stipulate floor prices and "take-or-pay" terms (meaning the buyer must pay the agreed amount regardless of whether they take delivery), with buyers providing advance payments and performance bonds as guarantees.
Micron's management also confirmed in their June 25 earnings call that the gross margins calculated based on the floor prices in the 16 five-year strategic customer agreements they have signed exceed the peak values of any previous cycle in history.
Furthermore, the next quarter's performance guidance provided by the big three in their earnings calls continues to be revised upwards. For example, Micron's revenue guidance for the next quarter has a median of $500 billion, far exceeding the market's previous expectation of $432 billion, with no signs of a weakening supply-demand outlook through 2027. Samsung's management stated in the July 30 call that significant new supply is unlikely to appear before 2028, as building a new fab from ground-breaking to volume production takes over three years, and the capacity increments from projects under construction cannot materialize in the short term. SK Hynix management confirmed in the July 29 call that AI-related demand would continue to accelerate in the second half, with the company's Q3 DRAM bit shipments expected to grow at a mid-single-digit percentage sequentially, and the supply-demand outlook through 2027 shows no signs of weakening.
Will the memory industry break the "death cycle" this time?
01
Supply Shortage
The prerequisite for the big three to secure favorable terms like floor prices, take-or-pay, and large advance payments in long-term supply agreements (LSAs) is that buyers have "no room for negotiation" under the current supply-demand dynamics.
Micron's management judged in the Q2 earnings call that the tight supply-demand situation will persist beyond 2027. Jaejune Kim, Executive Vice President of Samsung's Memory Business, stated in the July 30 earnings call that the supply-demand gap in 2027 will be more severe than in 2026, and no significant new supply will emerge before 2028.
Goldman Sachs calculated in a research report titled "Answering 8 Key Questions on the Memory Sector" released on August 4 that the global DRAM supply-demand gap (supply minus demand as a percentage of demand) in 2026 will be -4.9%, the most severe in the past 15 years, and is expected to tighten further in 2027.
A supply chain source close to SK Hynix told the Economic Observer reporter that the expansion pace of the original manufacturers (OEMs) has noticeably accelerated over the past month. Long-term contracts signed with U.S. customers have significantly improved the visibility of future demand. However, capacity plans up to the first half of 2028 are already locked in, and capacity from new projects will only begin to be released in the second half of 2028. The key bottleneck limiting the expansion speed is the EUV (Extreme Ultraviolet) lithography machines (currently the only lithography equipment capable of producing advanced node memory chips). The delivery cycle for such equipment has exceeded two and a half years.
The source explained that SK Hynix's publicly stated capacity target is "to double capacity by 2030 compared to the end of 2025," corresponding to an average annual growth rate of about 16%–17%.
Equipment delivery cycles constrain overall supply growth, and there are also constraints within the allocation of existing capacity. Lin Meibing, Chief Analyst at ChipSaying ICTIME, told the Economic Observer reporter that HBM requires about three times the number of process steps (such as lithography, etching) compared to conventional DRAM. Converting to produce 10,000 wafers per month of HBM capacity consumes approximately the equivalent capacity needed for 30,000 wafers per month of conventional DRAM. The industry's current HBM monthly capacity is about 326,000 12-inch equivalent wafers (Samsung ~140,000, SK Hynix ~150,000, Micron ~36,000), which is still insufficient.
Wang Xudong, a semiconductor analyst at Sigmaintell, said that AI servers already account for about 50% of global DRAM wafer capacity in 2026, and this proportion is expected to rise to about 60% in 2027.
Every extra wafer allocated to HBM reduces conventional DRAM supply by three wafers, mutually reinforcing the tightness of both product categories. In the past, during the high-profit periods of each memory cycle, the big three expanded capacity across all directions. This time, however, expansion is highly concentrated on AI, making the supply gap on the consumer end more persistent than before.
Even more counter-intuitive is the profit comparison between the two product categories.
The aforementioned source close to SK Hynix revealed that the current price per GB for HBM3E (the third-generation enhanced HBM, the mainstream memory product currently used in AI accelerators) is about $12–$13. The price per GB for HBM4, which starts arriving in the second half of the year, ranges from about $16 to $19. The 2026 pricing for both products is lower than that of contemporary DDR5.
Currently, the industry-average gross margin for conventional DDR products is generally above 80%. HBM, due to yield losses in packaging and TSV (Through-Silicon Via, a technology for drilling holes to connect data lines between multi-layer chips) processes, has a gross margin of about 50%–70%.
Goldman Sachs projected in the aforementioned report that by the end of 2026, the average price of conventional DRAM will rise to about $2/Gb (compared to $0.5–$0.6/Gb at the end of 2025). As of Q2, the price per GB for conventional DRAM has already surpassed that of HBM. The institution expects that HBM's blended average selling price will need to increase by 87%–100% year-over-year in 2027 to regain its premium relative to conventional DRAM.
The source close to SK Hynix believes that HBM will see a significant catch-up price increase in 2027, as HBM currently uses annual pricing negotiations, which have not kept up with market conditions.
Major buyers are even modifying product designs as a result. A research report released by market research firm TrendForce on August 4 shows that Nvidia has, starting from Q3 this year, changed the HBM configuration for its next-generation GPU Rubin Ultra from HBM4E 12-layer stacking to parallel evaluation of various downgraded specifications including HBM4E 8-layer, HBM4 12-layer, and HBM4 8-layer. Nvidia has also decided to halve the memory capacity of its next-generation superchip modules due to persistent shortages of LPDDR5X (a low-power memory widely used in servers and mobile devices).
TrendForce estimates that HBM bit shipments will grow 50%–60% year-over-year in 2027 but will still be insufficient to meet demand.
Even Nvidia has begun lowering product specifications because it cannot get enough memory. Faced with a supply-demand gap lasting at least until 2028, what other buyers can do seems limited to "signing contracts early to lock in supply."
02
Five-Year Long-Term Agreements
For many years, the trading model for memory chips has been quarterly pricing.
At the beginning of each quarter, OEMs first set a temporary price for shipments. The final settlement price is finalized in the last month of the quarter. Prices move with market supply and demand, with increases or decreases entirely determined by the current quarter's conditions. Both buyers and sellers seek their optimal solutions in a quarterly-unit game.
It is precisely this quarterly pricing model that causes memory chip prices to fluctuate far more than other semiconductor products. OEMs' profits fluctuate wildly with prices, and the capital market has long priced memory stocks as typical cyclical stocks.
Now, the big three's approach is to shift from the long-standing quarterly transactions to three-to-five-year long-term contracts.
The reporter learned from interviews that such contracts mainly include several core clauses: buyer and seller agree on supply quantities for three to five years; unit prices are renegotiated annually; prices have both upper and lower limits, forming a "price corridor"; buyers provide take-or-pay commitments, advance payments, or performance bonds as guarantees; some contracts also stipulate guaranteed minimum revenue amounts.
Multiple interviewees told the reporter that annual supply agreements have sporadically appeared in the memory industry in the past, but the duration typically did not exceed one year, coverage was small, and clauses lacked floor prices and take-or-pay terms, making their binding force limited.
The LSAs being promoted by the big three in this round are not on the same scale as before in terms of duration, scale, and binding force.
Micron signed 16 Strategic Customer Agreements (SCAs) in its fiscal Q3 2026, with a five-year term under take-or-pay terms. Fourteen of these include floor prices, with the total amount calculated based on the floor prices being approximately $100 billion. Micron expects to receive about $22 billion in cash deposits and letters of credit as performance guarantees (~$18 billion cash, $4 billion L/Cs). The goal is for such agreements to eventually cover over 40% of revenue.
Samsung's coverage is even larger. Jaejune Kim stated on July 30 that Samsung plans to allocate 60%–70% of its capacity to multi-year LSAs. Considering the number of customers seeking contracts is still increasing, this ratio may rise further; the agreements are based on a five-year rolling renewal basis. All of the world's top five data center customers have signed, and another five large AI customers are in final negotiations. Samsung has already received about one-quarter of the total agreed advance payments.
Additionally, SK Hynix's DRAM marketing head, Park Joon-duk, confirmed on July 29 that negotiations with about 10 key customers were completed but did not specify the proportion of total sales to be covered by LSAs.
According to a related Goldman Sachs report released on July 29, over half of server DRAM is now covered by LSAs.
Of course, the coverage of LSAs is not limited to DRAM. NAND flash manufacturer SanDisk confirmed on August 5 that it signed multi-year LSAs (SanDisk calls it a "new business model") with eight customers, with a weighted average term exceeding four years. The total revenue calculated based on floor prices is $93.9 billion, with accompanying financial guarantees of $16.5 billion. SanDisk CEO David Goeckeler stated that these LSAs are expected to cover over 50% of SanDisk's bit capacity in fiscal 2027, reaching about two-thirds in fiscal 2028.
The willingness of buyers to accept these terms stems from the core reason that the duration of the supply gap is long enough, leaving no room for "waiting for prices to fall."
A product manager at a large memory module manufacturer in Shenzhen stated that server memory is currently in severe shortage. Both CPU server and GPU server memory face supply gaps, even to the point where server configurations need to be downgraded. The primary need for customers is to secure supply, not to negotiate a better price.
The details of the price terms further reflect the strong position of the sellers.
Bank of America Securities analyzed in a report on August 1 that Samsung limits the price decrease in LSAs to no more than 5% sequentially, while the increase space is 10%–20% or more with no upper limit. The big three benefit from following market prices when they rise, while the lower limits agreed in the clauses support revenue when prices fall.
Lin Meibing said that the supply prices for 2027 are basically locked for several large U.S. cloud computing customers who have signed LSAs. The terms often extend to 2030 or even 2035. Given that supply will still be tight in 2027, sellers have no reason to offer lower prices. He judges that the possibility of price adjustments lies beyond 2028, when industry capacity begins to increase, and pricing will have room for adjustment.
After comparing the relevant LSA terms, Goldman Sachs stated in the aforementioned report that all four dimensions—duration, coverage, pricing mechanism, and enforcement constraints—are evolving in favor of suppliers. This shift in business model is expected to extend the duration of high profitability for memory manufacturers and support their valuation multiples rising from the previous 5–6x to 8–10x.
David Goeckeler said in the August 5 call that three or four quarters ago, the market was still pricing quarterly, but now SanDisk has visibility into demand for over four years. Some customers return to add procurement volumes for future years just one quarter after signing. "Previously, this was just a supply chain price negotiation conducted every quarter."
It is worth noting that the above LSAs mainly cover conventional DRAM and NAND flash. HBM sales models differ. As each generation of HBM requires custom validation for specific AI chips, OEMs typically negotiate pricing and volume separately with core customers like Nvidia on an annual basis, locking in next year's price and quantity ahead of time.
Micron management confirmed in the earnings call that their HBM supply for the entirety of 2026 is already sold out, with prices and quantities finalized.
03
Beyond 2028
Currently, the sequential quarterly increases in DRAM contract prices are already narrowing. According to TrendForce statistics, DRAM contract prices rose by about 90% sequentially in Q1 2026, dropping to about 60% in Q2. Multiple industry insiders stated that Q3 growth is expected to further narrow to 13%–18%.
The slope of price increases is slowing, but how far is it from the peak?
In Lin Meibing's view, price increases for memory chips will noticeably converge after mid-2027. With new capacity coming online in 2028, there will be room for price adjustments downward.
The aforementioned source close to SK Hynix stated that current OEM inventory is about 2–4 weeks, lower than the normal level of 4–5 weeks and far below the 10+ weeks typically seen before previous downturns begin. Signals of a supply-demand reversal have not yet appeared. The possibility of substantial price adjustments exists after the second half of 2028.
Even if prices begin to adjust downward after 2028, the industry consensus is that the pace of decline will be "significantly different" from before.
The product manager from the memory module company told the reporter that consumer electronics will be the first to feel price softening, followed by niche products (smart speakers, set-top boxes, etc.), then servers. Automotive electronics, having the longest LSA terms and the smallest market volume, will be the last segment to adjust.
In other words, in previous memory cycle downturns, prices would collapse across the board within months. This time, however, the big three are "staggering" the adjustment cycles for different customers and product lines through LSAs.
Furthermore, the competitive landscape of the global memory market will see new variables after 2028, with the keyword being ChangXin Memory Technologies (CXMT) (688825.SH).
CXMT was listed on the STAR Market on July 27, raising approximately 57.9 billion yuan. Its market capitalization once approached the 4 trillion yuan mark after listing. According to shipment statistics from market research firm Omdia, the company's global DRAM sales share was 7.67% in Q4 2025, ranking fourth globally.
The reporter learned that CXMT's production costs and end-product quotes are about 10% lower than comparable overseas products. As the big three actively shift advanced node capacity to AI, the market space vacated on the consumer electronics end is being filled by CXMT.
Lin Meibing told the reporter that on the two mainstream product lines, DDR and LPDDR, there is basically no generation gap between domestic memory manufacturers and overseas leaders. The share of server product revenue for Chinese DRAM companies has grown rapidly in recent years, driven by DDR5 shipments.
Lin Meibing believes the biggest technical challenge for domestic memory manufacturers lies in the HBM field. Overseas big three are already mass-producing HBM3, with some completing validation for HBM4 mass production. Domestic companies' HBM3 is still in the capacity ramp-up and verification stage, with a technology gap of about two generations. He also noted that domestic companies lack the burden of historical production lines. The inability to purchase EUV lithography machines has instead prompted them to invest in two entirely new technology paths—VC-T (Vertical Channel Transistor) and wafer bonding—potentially allowing them to narrow the gap in next-generation architectures.
The reporter learned that CXMT is currently advancing cooperation with a domestic mature-node foundry on CBA (Chiplet-Based Architecture, a packaging technology that manufactures memory and logic chips separately and then hybrid bonds them). CXMT manufactures the memory wafers, and the partner foundries the logic wafers, to build a complete domestic HBM manufacturing chain.
A source close to CXMT told the reporter that CXMT's LPDDR6 (the next-generation low-power memory for mobile and mobile devices) product is currently nearing the end of R&D verification. The first product design specification has a speed of 12,800 Mbps, using 16Gb dies, and is expected to achieve mass production introduction in the second half of 2026.
Wang Xudong believes that to ensure supply chain stability, domestic terminal brands and AI server manufacturers are significantly increasing the procurement and adoption ratio of domestic DRAM. Substitution with domestic products will continue to penetrate for general-purpose DRAM and consumer-grade NAND, but large-scale substitution for high-end HBM and AI enterprise storage is difficult to achieve in the short term.
Pressure on the consumer end is also changing the demand structure of the memory industry.
The product manager from the memory module company told the reporter that the year-over-year downward revision for smartphone memory demand for the full year 2026 is estimated at 15%–20%. In Q2, some smartphone manufacturers could not accept Samsung's quotes and significantly cut procurement. Consumer memory prices might "stop rising" by Q4.
On July 30, Apple CEO Tim Cook described the current round of memory price increases in his final earnings call as a "once-in-a-century flood." Apple's product gross margins are under continuous pressure due to rising memory costs. Apple CFO Kevin Parekh confirmed that the sequential decline in gross margin over the last two quarters "can be more than 100% explained by memory cost changes." The company further lowered its gross margin guidance for the next quarter to 47%–48%.
Wang Xudong calculated that for a $150-range smartphone with a "4GB+128GB" configuration, the share of memory costs in the total BoM has risen from 22% in Q3 2025 to 64% in Q3 2026. He predicts that smartphones priced below 1,500 yuan will be difficult to see in 2027.
However, while the consumer end contracts, new sources of demand are also accumulating.
Lin Meibing told the reporter that automotive-grade memory growth in the Chinese market is very strong. The rapid penetration of ADAS (Advanced Driver-Assistance Systems) is driving a significant increase in memory capacity demand for intelligent cockpits and autonomous driving domain controllers. Year-over-year growth for automotive memory in China is expected to reach 70%–80%.
Additionally, CXL (Compute Express Link, an interconnect technology supporting memory pooling and sharing between servers) is expected to begin volume adoption around 2028 following adoption by Nvidia and Google. Currently, samples are available from various manufacturers but not yet mass-produced. Also, the first industry standard specification for High Bandwidth Flash (HBF) (a new memory tier between HBM and SSDs) was jointly released by SK Hynix and SanDisk on August 4, with shipments expected by the end of 2027.
After 2028, if these new demand sources gain traction, they could partially offset the supply easing brought by new memory capacity additions.
The product manager from the memory module company believes that although the current price increase slope for memory is narrowing quarter by quarter, the profit scale for the big three in 2027 could still possibly exceed that of 2026. He emphasized that the market's current focus of debate is "when the cycle peaks." But for the big three, a more important variable might be how high the LSA coverage ratio can go. "Micron's target is over 40%, Samsung is already at 60%–70%. If the industry's average LSA coverage eventually stabilizes above 50%, then even if spot market fluctuations occur, OEMs would have over half their revenue locked in, and the volatility of profits would be much smaller than in any previous cycle."
This article is from the WeChat public account "Economic Observer", author: Zheng Chenye
