Long-term contracts, traditionally a buyer's insurance policy, have become a binding rope for both sides in the HBM market.
Producing a single HBM wafer consumes roughly the same capacity as three standard memory wafers. Logically, the most in-demand and expensive product should command the highest margins. Q1 2026 data tells the opposite story: HBM's per-wafer value and margins were both overtaken by a standard server memory module—DDR5 64GB RDIMM.
How did a long-term contract designed to lock in prices end up locking the most scarce product into lower margins? The answer lies in how HBM pricing is negotiated.
Long-term contracts lock in time, not price
HBM trading operates on a different rhythm than standard memory. Samsung, SK hynix, and Micron all use annual negotiations for HBM: one round of talks per year, locking in prices for the following twelve months. This mechanism works fine in years of stable supply and demand. The problem emerged after the second half of 2025, when standard DRAM prices began jumping quarterly while HBM contract prices remained fixed at the previous negotiation level.
Market research firm TrendForce explained this timing gap plainly in a report released on June 2: while the spot market had risen for six months, HBM contract prices were still locked to the previous year. The long-term contract does not just fix the price; it fixes whether a price adjustment can occur at all.
The scarcest product earns the least
The margin inversion tracked by TrendForce appeared in Q1 2026: DDR5 64GB RDIMM's per-wafer value exceeded that of HBM, and HBM's margins fell below those of DDR5 starting that quarter. The awkwardness of this conclusion is that HBM requires far more stacking layers, consumes more capacity, and involves greater packaging complexity than standard memory, yet its price is the one fixed by contract.
3:1
The amount of standard DDR capacity displaced by one HBM wafer
HBM's share of total DRAM wafer starts in 2027
100 billion
Cumulative minimum price (USD) corresponding to Micron's long-term contracts
These three figures represent HBM's capacity displacement multiplier, TrendForce's estimated HBM wafer start share for 2027 (approximately 18% in 2025 and 22% in 2026), and the cumulative minimum price scale of Micron's signed long-term contracts extending to 2030. The first explains why standard memory prices are rising; the third explains why Micron dares to retain price caps.
The 2027 price hikes are different in nature
TrendForce projects that the three major memory manufacturers will significantly raise HBM contract prices in 2027, describing the increase as 'several times' higher, while some market reports have used the term 'doubled.' It is crucial to clarify the nature of this shift: demand has not suddenly spiked again; rather, this adjustment corrects the inherent lag in pricing mechanisms. The manufacturers' confidence stems from three factors: overall DRAM supply shortages, the manufacturing complexity of HBM4, and rising costs driven by increased stacking layers and larger die sizes.
The demand-side dynamics also follow two distinct phases. In 2026, HBM growth is primarily driven by capacity upgrades in AI ASICs, with HBM capacity per AI chip increasing from 96GB / 192GB to 216GB / 288GB. In 2027, the momentum will shift to NVIDIA Rubin Ultra, where HBM capacity per GPU is expected to rise to 384GB, further compounded by increased adoption across platforms such as Google TPU.
Supply-side timelines align with this demand trajectory. Micron's earliest new capacity—its first fab in Idaho—is not expected to come online until mid-2027, with the second fab following in late 2028. On July 9, the company raised its US manufacturing investment plan from $200 billion to over $250 billion, anchored by the planned new facility in Clay, New York. Why HBM is squeezing conventional memory
For every wafer of HBM3E produced, approximately three wafers of conventional DDR capacity are foregone; by HBM4E, this ratio will approach one-to-four. The reason is straightforward: HBM involves stacking multiple layers of DRAM dies and connecting them through silicon vias, resulting in higher silicon area and process time per finished unit. Consequently, the tighter HBM supply becomes, the less conventional memory prices can be contained; these two forces are two ends of the same rope.
A benchmark illustrates how high conventional memory prices have climbed. In late September, analytics firm Kernel Insight estimated that, based on market value per unit area, 1b process DRAM is worth approximately $0.654 per square millimeter, compared to about $0.424 per square millimeter for TSMC 2nm wafers, representing a premium of roughly 54%. This comparison reflects market value versus market value, not manufacturing cost versus manufacturing cost, but it demonstrates that the market has assigned an unprecedented price to memory area.
The same seller's market, two different contracts
The real divergence lies between two of the major manufacturers. TrendForce's mid-year report noted that SK hynix has removed price caps from some long-term supply contracts, while Micron has chosen to retain them—specifically, the retained cap is pegged to market price levels in Q2 2026. Despite HBM selling out in both cases, one company is willing to let prices float, while the other has capped its upside.
The divergence between the two approaches stems from market share and the necessity of customer relationships. Industry tracking data indicates SK hynix holds approximately 50% to 55% of the HBM market, Samsung holds about 35% to 40%, and Micron holds a lower share. The more limited the share, the greater the need to secure long-term contracts, where a price cap effectively serves as a promise not to raise prices during periods of peak scarcity.
Micron’s other pillar of support comes from the other side of the contract. The company classifies a series of long-term agreements as Strategic Customer Agreements, which include customer prepayments, take-or-pay volume guarantees, and price floors that protect the seller. According to company disclosures, these agreements have accumulated approximately $22 billion in commitments, corresponding to a cumulative minimum price scale of roughly $100 billion by 2030. With both price floors and caps written into the contracts, both ends of the cycle are locked in. On August 10, Executive Vice President Sumit Sadana stated more directly at the KeyBanc technology forum that supply in 2027 will be tighter than in 2026, and the company “cannot see when supply will catch up with demand.”
Putting these factors together, my assessment is that the significant increase in HBM contract prices for 2027 is closer to a mechanism repair, making up in one go for the price increases suppressed by annual negotiations over the past two years. Whether this trend can sustain itself will be answered by next year’s contract price negotiations—if the year-over-year increase in 2028 remains close to that of 2027, it indicates the price center has genuinely shifted upward, invalidating this judgment.
See you in the comments. If you are a procurement officer for an AI accelerator manufacturer, in the 2027 negotiation round, would you sign a three-year fixed-price contract or accept an annual contract with no price cap?