Many readers have likely noticed that memory-equipped products—PCs, game consoles, and cars—have been getting more expensive over the past several quarters. The cause is usually dismissed with the single phrase "semiconductor shortage due to AI demand," with little explanation of the actual structure behind the price hikes. At its earnings call on July 29, 2026, SK hynix revealed that it had completed negotiations on long-term agreements (LTAs) with approximately 10 companies, including major customers. Behind this record-setting quarter—with contract periods of up to roughly five years and DRAM average selling prices up approximately 30% quarter-over-quarter—the agreements were officially described as "deals to curb sudden price volatility." But when you line up the conditions against contracts Micron signed around the same time, it becomes clear that what was actually smoothed out was only the downside, while the manufacturer captures the entire upside.
An "Approximately 10 Companies" Deal Struck Behind a Strong Earnings Report
SK hynix's Q2 FY2026 earnings showed revenue of KRW 79.3187 trillion (up 257% year-over-year, up 51% quarter-over-quarter), operating profit of KRW 60.5426 trillion (up 557% year-over-year, up 61% quarter-over-quarter), an operating margin of 76%, and net profit of KRW 93.9226 trillion (a net margin of 118%). Net profit exceeding revenue is not something that occurs in ordinary earnings reports, and the main cause here lies outside the core memory business's rapid expansion. SK hynix recorded KRW 63.27 trillion in non-operating profit (gains and losses related to investment assets), which is believed to largely reflect gains from the sale and valuation of shares in Kioxia, the NAND manufacturer in which SK hynix had invested alongside Bain Capital and others through a Japan-US-Korea consortium. According to reporting by Seoul Economic Daily, this one-time investment gain was the primary driver of the sharp increase in net profit. Note that these figures are preliminary and unaudited.
Converted to yen, operating profit amounts to roughly ¥6.8 trillion and net profit to roughly ¥10.56 trillion. DRAM average selling prices rose approximately 30% quarter-over-quarter, while NAND average selling prices climbed in the mid-50% range.
At the same earnings call, Song Hyeon-jong (President of the Corporate Center) stated that LTA negotiations with about 10 companies, including major customers, had been completed. Contract periods run up to roughly five years, and Song explained that "the pricing structure varies by customer and product characteristics, but is designed to respond to price fluctuations." He added that financial mechanisms such as deposits are also built in to ensure contract compliance.
SK hynix is also in discussions with major customers regarding 2027 HBM supply volumes and pricing, stating that "HBM prices are not determined solely by movements in commodity DRAM prices." Full-scale mass production of HBM4E is targeted for 2027. For Q3, the company projects DRAM shipment volume to rise 10% quarter-over-quarter, with NAND shipment volume growing in the low single digits percentage-wise. Whether this shipment plan is actually achieved will be the first indicator of whether the LTAs signed this time are backed by real demand.
Behind the "Price Smoothing" Label: The Removal of the Price Cap
What Song himself stated at the earnings call was only the abstract phrase "designed to respond to price fluctuations"; no more specific mechanism has been disclosed. Filling in this gap is reporting from market research firm TrendForce. Citing industry sources, TrendForce reports that SK hynix has removed the industry-standard price cap from its LTAs, shifting to a structure in which it fully captures the upside when spot prices surge.
Under conventional long-term contracts, a price cap protects customers by preventing the manufacturer from charging more than a set ceiling no matter how high market prices climb, while a price floor supports the manufacturer's revenue by obligating the customer to purchase at least at that minimum level no matter how far market prices fall. Having both a cap and a floor together has functioned as a bidirectional buffer, protecting both manufacturer and customer to some degree from sudden market swings.
The SK hynix proposal reported by TrendForce removes only this cap. Without a cap, the manufacturer can charge customers the full market price no matter how high it surges. As of what was disclosed in July, there is no confirmation of any protection against downside price movements, and the deposits and confirmed purchase volumes built into the contract are, according to The Register, mechanisms meant to strengthen contract compliance and demand visibility—not to protect customers in the event of a price decline. This results in an asymmetric design: the manufacturer secures the right to the upside, while it remains unclear how much protection, if any, customers are given on the downside.
Behind this asymmetric design lies SK hynix's own negotiating leverage. The negotiations took place immediately after the company posted a record quarter, with revenue of KRW 79.3187 trillion and an operating margin of 76%—a situation in which the manufacturer could afford to negotiate from a position of strength as long as the supply shortage continued. Conversely, this suggests that hyperscalers had no choice but to accept the removal of the cap because securing supply took priority over the price level itself. The judgment appears to have been that, in the context of sustaining AI investment, securing the necessary volume mattered more than holding down price increases.
Micron's Insurance vs. SK hynix's Bet: Two Contract Designs
This asymmetry stands out clearly when compared with the long-term contracts Micron has signed. TrendForce's headline already frames SK hynix's contract as "diverging" from Micron's, and the framework for contrasting the two companies originates from that report. Building on that comparison, this article interprets the nature of the contracts through an economic frame of "insurance" versus "a bet," going further to examine the distribution of winners and losers and the shift in the memory cost/capital expenditure ratio (from 8% to 30%).
Micron reportedly signed contracts running from 2026 through 2030 with 16 strategic customers, and for its major large-volume contracts among these, set the Q2 FY2026 market price as the ceiling for existing products, with a floor guaranteeing a margin above the company's own historical peak profit margin (roughly 62%) (some of the 16 contracts include fixed pricing or no cap/floor at all). The contracts cover roughly 20% of DRAM shipments and about one-third of NAND shipments, guaranteeing cumulative revenue of at least $100 billion through 2030. However, this $100 billion figure is the cumulative total of the remaining contract terms across 14 of the 16 companies, and is not simply revenue generated evenly over five years (automotive contracts generally run three years, shorter than the other agreements).
To put Micron's design in other words: it is an insurance-like contract that says, "We won't chase the current abnormally high prices any further, but in exchange, we will absolutely defend a profit margin above our own historical peak even if prices fall." It surrenders some of the fruits of rising prices at a certain level, and in exchange, locks in earnings for a downturn. What SK hynix chose is the mirror image of this. While it remains unclear from public disclosures how much protection against downside risk is built in, it has not given up the right to the full upside.
Ordinary insurance is a mechanism where one pays a premium in preparation for future losses, but here the manufacturer takes on almost none of the downside risk while keeping all of the upside gains for itself. Which approach is correct can only be determined by how future price cycles unfold, but the fact that both companies chose opposite contract structures in the face of the same supply crunch at least shows that judgments about how to place this bet diverge even within the industry. Within the same vessel called a "long-term contract," completely opposite bets have been placed.
Why Memory Costs Are Approaching 30% of Capital Expenditure
According to SemiAnalysis estimates, the share of memory-related spending within hyperscalers' total capital expenditure is expected to reach roughly 30% by 2026, up from about 8% around 2023-2024. This sharp rise in the ratio is the backdrop behind hyperscalers' demand for "price smoothing" through long-term contracts.
Since each company's breakdown is undisclosed, this figure remains an industry estimate. Even so, the shift from 8% to 30% amounts to roughly a 3.75x increase (30÷8) by simple calculation. The ratio has ballooned to nearly four times its original level in just a few years—a scale substantial enough to shake up budget allocation across AI investment as a whole.
The main protagonist of AI data center investment is supposed to be spending on GPUs and accelerators. Memory was merely one of the supporting components. But once this cost item comes to account for around 30% of total investment, the outlook for the entire business plan becomes harder to forecast.
GPU procurement budgets can be planned, but if memory prices are expected to swing by tens of percent every quarter, the return-on-investment calculations for entire data centers will keep shifting quarter by quarter. What hyperscalers were seeking was likely not an absolute reduction in price itself. What they truly wanted was predictability they could build into their budget plans.
There is a preceding chapter to this story. In April 2026, TrendForce reported that SK hynix and Microsoft had entered final negotiations on a three-year DDR5 supply contract worth tens of trillions of won, negotiating terms including an upfront payment of 10-30% of the total contract value and a minimum price floor in case of price declines. Around the same time, SK hynix was also reportedly negotiating with Google on a five-year contract for commodity DRAM (with a concept for a two-year extension contingent on supply of next-generation HBM), and Samsung Electronics was said to be separately negotiating contracts of three years or more with each of Google, Microsoft, and AMD. It is unclear whether these individual negotiations from April and the completion of LTAs with approximately 10 companies that Song disclosed on July 29 refer to the same set of contracts, but chronologically it appears that hyperscalers first raised the request to "please smooth out the pricing," and the manufacturers then proceeded with negotiations in response.
AI Demand and HBM Driving the Supply Crunch
Since 2025, demand for HBM and high-value-added DRAM for AI servers has expanded rapidly, and DRAM and NAND prices have come to record increases of tens of percent each quarter. In this latest earnings report as well, DRAM average selling prices rose approximately 30% quarter-over-quarter, while NAND average selling prices climbed in the mid-50% range. Samsung Electronics is also negotiating, around the same time, to raise Q3 DRAM average selling prices by up to 20%, with the possibility of exceeding 20% for LPDDR. However, this is the level of increase Samsung has proposed; TrendForce forecasts the actual increase that will be agreed upon will land between 13% and 18%.
There are multiple factors driving the price increases, but HBM is undoubtedly the symbolic one among them. The fact that SK hynix is negotiating 2027 supply volumes and pricing individually with major customers reflects the existence of a pricing formation mechanism for HBM separate from that of commodity DRAM. The very fact that negotiations are underway to lock in HBM pricing and volume in advance—even while HBM4E's mass production plans remain fluid—demonstrates just how strong the supply anxiety is on the demand side.
At a press conference at the Korea Chamber of Commerce and Industry's Summer Forum held in Jeju on July 15, SK Group Chairman Choi Tae-won described current memory prices as "abnormally high." This remark was made at a different venue by a different person than what Song stated at the earnings call, and the two should not be treated within the same context. Chairman Choi stated that the company is considering constructing a new factory in the United States in addition to Korea's Honam region, signaling willingness to avoid a chain of price increases through supply expansion. Historically, the memory market has repeated boom-bust cycles of roughly 2-4 years, with hyperscalers historically procuring favorably during sharp price downturns. The current shift toward long-term contracts represents a rare recent attempt by manufacturers to preemptively cut off downside risk from that cycle through contractual means, and whether it succeeds remains undetermined.
Who Profits and Who Bears the Burden
The primary beneficiaries of this contract structure are the three memory manufacturers—SK hynix, Samsung, and Micron. While mitigating price decline risk through contracts, they may be able to preserve record profit margins over a multi-year horizon. Samsung is also negotiating DRAM price increases around the same time, and the shift of pricing power to manufacturers is a pattern common to multiple companies.
Hyperscalers such as Microsoft and Google gain in terms of securing supply, but as the price for that, they take on sustained high procurement costs in the form of long-term contracts. The burden then flows downstream to PC and smartphone manufacturers and the consumers who buy their products. In February, IDC forecast that the 2026 PC market would decline 11.3% year-over-year and smartphone shipments would fall 12.9%; the smartphone forecast has since been revised down further to 13.9%.
This is not an unrelated matter for Japanese readers either. Game console makers such as Sony and Nintendo, as well as automakers that use automotive memory, are directly exposed to the increased procurement costs resulting from this sustained price surge. Some domestic PC manufacturers in Japan have also reportedly raised prices on DDR5-equipped products, and SK hynix's operating profit figure of ¥6.8 trillion reflects a structure in which the source of such domestic price hikes accumulates as profit for an overseas manufacturer. Since net profit of ¥10.56 trillion includes the gain from the sale of Kioxia shares, operating profit is a more accurate indicator of the actual earnings from the memory business itself.
SK hynix has not disclosed the specific names of individual hyperscalers or the total value of the contracts. It has only used the collective terms "approximately 10 companies" and "major customers," and it is also unclear whether the cap-removal terms reported by TrendForce apply uniformly to all of these roughly 10 contracts. The conditions under which the removal of the price cap actually takes effect—at what spot price level—also remain undisclosed.
This very lack of disclosure itself may help preserve the manufacturer's negotiating leverage. If the trigger conditions are not made public, customers cannot precisely estimate their future cost ceiling and must rely on the manufacturer's explanations at every price negotiation. This information asymmetry stands in exact mirror relation to the asymmetry built into the contract design itself.
The 2027 HBM Price Negotiations Will Test How Serious the Contracts Are
SK hynix states it is in discussions with major customers regarding 2027 HBM supply volume and pricing. The goal of full-scale HBM4E mass production in 2027 effectively sets the deadline for these negotiations. The first test of whether the LTAs signed with the roughly 10 companies this time are truly backed by real demand, however, lies closer at hand—in the current Q3 outlook.
If the Q3 shipment plan falls short, the stance of hyperscalers who accepted the removal of the price cap could harden. Having accepted the entire upside risk of price increases in exchange for supply certainty, if the crucial shipment volumes fail to materialize as planned, that would give customers grounds to demand the reinstatement of a price cap in the 2027 HBM negotiations. The value of a contract with no cap is truly tested precisely when spot prices surge beyond expectations—but once the price increases settle down, hyperscalers' assessment of the arrangement is likely to shift as well. Whether the bet of removing the cap from the contract succeeds can only be scored once the capacity-expansion investment Chairman Choi mentioned actually eases the supply shortage. Until then, the asymmetric contract structure remains fixed in the manufacturer's favor.