Prices forecast to double, then shares fell by a tenth in a day
On September 29, 2026, the research firm TrendForce projected that the average selling price of AI memory chips, blended across products, would rise 121% in 2027, more than doubling in a single year. Two weeks later, on October 15, shares of SK hynix, the largest maker of those chips, closed down 11.06% in Seoul. Samsung Electronics fell 9.48%. The KOSPI, Korea's main stock index, closed 7.57% lower at 5,554.22, after a slide steep enough to halt all trading for 20 minutes.
Within the same few weeks, prices were expected to double and a share price fell by a tenth in a day. Both can be true at once, and seeing why is the clearest way to understand how this business works.
What an AI memory chip actually is
The chip in question is called HBM, short for high-bandwidth memory. An AI processor can only work as fast as data reaches it. HBM stacks a dozen or so thin memory chips on top of one another and places the stack right beside the processor, linked by thousands of tiny vertical connections. Ordinary memory, the kind in a laptop or a regular server, sits farther away and moves data far more slowly.
Only three companies make HBM in volume : SK hynix and Samsung in Korea, and Micron in the United States.
AI memory uses about three times the silicon wafers for the same storage
Memory is made on round silicon wafers in very expensive, ultra-clean factories. Micron has said since 2024 that HBM3E, the current generation, needs about three times as many wafers as ordinary DDR5 server memory to make the same amount of storage, when both are made with the same manufacturing technology. The chips are larger, more of each chip goes to the wiring that links the layers, and stacking adds losses. If one layer in a stack is faulty, the whole stack is worth less.
So every wafer turned into HBM is a wafer not turned into ordinary memory, and building new factory space takes years. TrendForce gives three reasons for its 2027 projection. Supply stays tight, a growing share of sales will be the newer and pricier HBM4, and the next generation, HBM4e, starts up in the second half of 2027.
Micron says most of its 2027 AI memory is already sold
HBM is not bought off a shelf. Prices are set roughly once a year in negotiations between the three makers and their buyers, and much of the volume sits under long contracts. On September 30, 2026, Micron told investors it had completed agreements for "the vast majority" of its own calendar 2027 HBM supply, "with significant price increases year over year." That shows a tight market at one of the three makers. It does not show that most of the industry's 2027 supply is sold.
Micron also described 26 multi-year "take-or-pay" agreements, under which the buyer pays for the agreed volume whether or not it takes delivery. About three-quarters of the revenue Micron expects from these agreements has a defined pricing framework, and in a majority of that portion the price moves inside a band with a floor and a ceiling. The ceiling matters here, because it means part of the gain from higher prices has already been given up in exchange for certainty.
The 121% is one firm's projection of an average
TrendForce's figure is a projection, not a price anyone has paid yet. It is also a blended average across products. When buyers shift toward newer, more expensive generations, the average rises even if no single product's price changes. So 121% does not mean every chip will cost 121% more.
Tight supply and booked contracts explain why TrendForce expects prices to rise, but they do not by themselves produce a figure of 121%. That number is TrendForce's own estimate, and the shift toward newer chips is part of it. The contracts give firmer evidence of the direction, at least for one maker : Micron says its 2027 HBM prices are significantly higher than its 2026 prices.
Why the shares swing anyway : they are priced beyond the contracts
Contracts make next year's sales easier to see, and some of Micron's run for several years. But they do not fix future profit margins, how much customers will want, how much business competitors win, or what new factories do to prices. A share price reflects what investors expect a company to earn for many years, and memory has a long record of booms that ended in gluts. In past cycles, when prices were high, every maker built new capacity at once. The new factories opened around the same time, supply overshot demand, and prices fell faster than costs.
SK hynix's own results show how fast that can happen. In 2021 it made an operating profit of 12.41 trillion won, a 29% margin. By the first quarter of 2023 it was losing money : an operating loss of 3.402 trillion won on revenue of 5.088 trillion won, a margin of negative 67%. The company blamed weak demand and falling memory prices.
More capacity is on the way. Micron alone expects its Singapore HBM packaging plant to start output in early 2027, a new US factory to start in mid-2027, and more cleanroom space from late 2028. Micron's own view, and it is the company's outlook, is that supply will stay tight through 2028 and that it cannot yet see when supply and demand will balance.
The boom-and-bust history is a risk investors weigh, not proof that another bust is coming. Multi-year take-or-pay contracts, in which buyers commit to volumes in advance, may make this cycle less like the busts of the past, though they cannot guarantee profits. Whether this cycle will differ from earlier ones is exactly what investors disagree about.
Reports tied October 15 to interest rates and AI funding
Korean and international reports tied the drop that day to US inflation and to AI funding. US consumer prices rose 0.6% in September from August, according to the US Bureau of Labor Statistics, a hot reading that fed worries the Federal Reserve could raise interest rates. Reports also pointed to doubts about how AI projects will keep being funded. These are reporters' explanations of a market move, not a proven single cause.
Memory makers sell into AI spending, so doubts about that spending reach their shares directly.
SK hynix still leads AI memory, by a narrower margin
Share prices also move on who wins the next contracts. By Counterpoint Research's count of HBM revenue, SK hynix held 64% of the market in the second quarter of 2025. It held 56%, 57% and 58% over the next three quarters, and 50% in the second quarter of 2026. Over the same year, Samsung rose from 15% to 33%. These are shares of HBM revenue, not of all memory.
The latest drop is a single quarter, not a trend, and other research firms measure market share differently and get different numbers. What the figures do show is that the lead can change, which is one more thing a share price has to account for.
Contracts show next year more clearly than the years after
The thing to carry away is this. TrendForce expects AI memory to cost far more on average next year because supply stays tight and buyers are moving to pricier generations. Separately, Micron says this memory takes about three times the wafers of ordinary memory for the same amount of storage, and that it has already sold most of its own 2027 supply. The shares swing because they are priced on many years of profit that contracts do not settle, in a business that, not long ago, went from a large profit to a large loss within two years.
Nothing here says where prices or shares will go. What will shape the next turn is public : the terms of new contracts, how much new factory space opens and when, and how much of the newer HBM4 Samsung wins. The next quarterly results from SK hynix and Samsung, and what both say about 2027 supply, will be the next piece of evidence. The projections here belong to the firms named, and past results promise nothing about the future.
