The S&P 500 closed above 7,800 for the first time on Tuesday. Memory went the other way. SK Hynix lost 3.7% in Seoul and Samsung Electronics 1.5%, foreign investors were net sellers of 1.75 trillion won of KOSPI shares, and Seoul extended the slide on Wednesday ahead of Samsung’s preliminary third-quarter numbers on Thursday. Micron closed at $1,045.56, about 17% under its 52-week high of $1,255; SanDisk sits more than a quarter below its June record. The steepest drops belonged to Seagate and Western Digital, which are trading their own story (Toshiba’s plan to double hard-drive capacity by fiscal 2027). For the chipmakers the market is asking one question: does memory pricing hold into 2027 and 2028?
It does, with room to spare. Memory prices peak when supply catches up with demand, and the producers place that point well past next year. Micron sees supply tighter in 2027 and 2028 than in 2026 and says it has no line of sight to balance. SK Hynix expects demand to exceed its capacity beyond 2030. What the market is selling is a slower rate of increase in one product, conventional DRAM, read as a top for the whole complex.
Start with the number the bears are using. TrendForce expects conventional DRAM contract prices to rise 10% to 15% in the fourth quarter, after 13% to 18% in the third. That deceleration is the bearish case on pricing. The same forecast carries a second line that has drawn far less attention: total DRAM including HBM is projected to rise 15% to 20% in the fourth quarter, up from 8% to 13% in the third. The blended number is speeding up. NAND contracts are forecast another 15% to 20% higher, with enterprise SSD the one category where price growth accelerates. Server DRAM stays undersupplied through year-end. Every category still points up.
The acceleration comes from HBM, the part of the market that has yet to reprice. HBM is contracted a year ahead, so through 2026 it shipped at prices set before the shortage reached conventional DRAM, and its profitability fell behind the commodity parts. Micron said on its fourth-quarter call that its calendar 2027 HBM prices are significantly higher and reset in January. TrendForce has raised its 2027 forecast for blended HBM prices to a 121% increase; industry estimates put a 12-high 36GB HBM4 stack at about $1,300 next year against roughly $600 this year. The volume side tightens it further. Samsung expects HBM to take nearly 30% of industry DRAM capacity in 2027, up from about 20% now, and a wafer turned into HBM yields far fewer bits than the same wafer sold as DDR5. The most expensive memory in the data center has its largest price reset still ahead.
New supply cannot close that gap on the timeline the selloff implies. The binding constraint is cleanroom space, and cleanrooms take years. Micron’s first Idaho fab produces its first wafers in mid-2027, with meaningful output a few quarters later; its next wave (a second Idaho fab plus Singapore and Japan expansions) arrives in the second half of 2028. Samsung’s P5 is targeted for 2028 and its first Yongin fab for 2029. Micron told analysts that most of the increase in its fiscal 2027 construction budget is for cleanrooms that open in late 2028 and beyond, and that industry DRAM supply growth slows next year as HBM4 and HBM4E consume more wafer per bit and each new node delivers less than the one before. SK Hynix’s chief executive has called 2027 the worst year for supply in the industry’s history.
Buyers are acting on that view. Micron has more than 75% of its fiscal 2027 output committed and has moved allocation talks to 2028; its 26 strategic customer agreements cover about 35% of bit volume through 2030, some now run past 2030, and the newest were negotiated at current, higher prices. On the call, management described DRAM as the principal constraint on AI data centers, ahead of logic and power. Accelerator designers are cutting memory per chip because they cannot get enough: TrendForce says GPU and ASIC vendors are weighing 8-high HBM stacks in place of 12-high, at a 10% to 20% premium per gigabit, and it ties the shift to scarce supply and rising system costs. Micron describes the same behavior from the seller’s side; customers ship as many processors as their memory allocation allows, and any memory that frees up would be absorbed. Amazon lifted its 2026 capital budget to about $220 billion largely because memory cost more, and expects to be short of capacity in 2027 as well. A peak needs buyers to stop paying. These buyers are paying up for less than they want.
The trade data agrees. South Korea’s semiconductor exports reached $60.3 billion in September, up 262.8% from a year earlier and the first month above $60 billion, in a month with three and a half fewer working days because of Chuseok. August had been up 209%. Growth accelerated. DRAM spot quotes, which turn first at a genuine top, were higher across the board in Taipei on Wednesday while Seoul was selling the producers.
Equity pricing assumes the reverse. Micron trades at about 6.9 times its first-quarter earnings guidance annualized ($38.15 a share at the midpoint), before the margin expansion management has indicated for the rest of fiscal 2027 on continued, slower price increases. That multiple treats fiscal 2027 as the last good year. With HBM repricing in January and Micron’s own next block of DRAM capacity arriving in late 2028, after its fiscal 2028 closes, the stronger year is more likely the one after. The average of 49 analyst targets on Micron is $1,535.57, about 47% above Tuesday’s close. SanDisk holds floor-priced contracts worth at least $93.9 billion over their life, with management targeting coverage of more than half its fiscal 2027 volume and about two-thirds of fiscal 2028.
Samsung’s figures on Thursday will be read through the same lens. Consensus sits between 106 trillion and 108 trillion won of third-quarter operating profit, around nine times the year-ago level, and SK Securities sees the memory margin flat near 76%. A flat margin will be called a top. It is a margin earned before the 2027 HBM contracts take effect.
The signal that would mark the peak is specific: the first quarterly contract forecast with a minus sign in it. There is none for 2026. The next one covers the quarter in which HBM reprices.