The memory industry has spent 2026 telling investors it finally solved its oldest problem. High-bandwidth memory, the stacked DRAM that feeds Nvidia’s accelerators, is sold out. SK Hynix has sold its DRAM, NAND and HBM output through 2026, most of it to Nvidia, and its executives now warn the shortage could run past 2030. Samsung and SK Hynix stretched their supply deals from the traditional one-year term to three and five years. The pitch, repeated across sell-side notes, is that long contracts and locked-in demand have broken memory’s boom-and-bust cycle.
They have not. In July, SK Hynix quietly removed the piece of the contract that would have proved the claim.
What actually changed
Memory long-term agreements traditionally carried a price cap. The buyer committed to volume, the seller committed to supply, and a ceiling protected the customer from a spike while a floor protected the seller from a crash. It was a shock absorber built into the paperwork.
SK Hynix took the cap out. Per TrendForce, its new agreements let spot-market increases flow fully into contract pricing during a shortage, and the company “may now be the only major memory supplier not applying a price cap in its long-term agreements.” SK Hynix later confirmed to Korean press that its long-term deals use varied pricing rather than a single fixed number.
The logic is obvious while the shortage lasts. Uncapped prices let the seller capture every dollar of an AI-driven spike, and the two Korean makers are now running 40 to 50 percent operating margins. Customers accepted it because in 2026 the scarce thing is supply, not price. Guaranteed allocation is worth more than a cheap ceiling when you cannot get the parts at all.
Volume is locked. Price is not.
Here is the part the “sold out” headline hides. A sold-out order book locks volume. It does not lock price. Those are different forms of safety, and the industry has been quietly trading the second for upside on the first.
When prices ran fixed, an order book was a genuine hedge: even in a glut, the seller kept selling at the agreed number. Strip the cap out and the same order book becomes a volume commitment priced at whatever the market does. On the way up, that is a windfall. On the way down, it is a direct pipe from the spot market into your revenue line. The contract no longer absorbs the cycle. It transmits it.
Memory has never been cyclical because contracts were short. It has been cyclical because supply and demand swing violently and someone always ends up holding price risk when they cross. Lengthening the contract to five years changes who you sell to. Removing the cap changes what you are exposed to. Only the second one touches the cycle, and it points the wrong way.
The capacity is already being poured
The reason this matters now is that the supply side is moving. SK Hynix is scaling its 1c-node DRAM capacity from roughly 20,000 to between 160,000 and 190,000 wafers a month by the end of 2026, close to a ninefold increase. Samsung has recovered from its HBM4 yield problems, began mass production in February, and reached one billion dollars in HBM revenue within four months.
That recovery is already visible in the market structure. Counterpoint Research puts Samsung’s HBM revenue share at 33 percent in the second quarter of 2026, up from 21 percent in the first, while SK Hynix slipped from 58 to 50 percent and Micron eased to 18.
A three-way race with a resurgent number two is exactly the setup that ends shortages. The revenue at stake is large enough to pull every dollar of that capacity forward: Bank of America estimates the HBM market reaches 54.6 billion dollars in 2026, up 58 percent year on year, and HBM4 runs around 550 dollars a stack against roughly 300 for HBM3E. Those margins are the signal that invites the flood. HBM4 volume only becomes meaningful in the second half of 2026, so the fresh capacity and the newest, priciest generation ramp into the same window.
Three makers, three bets
The most telling detail is that the three suppliers have not agreed on any of this. Their contract structures encode three different views of when the shortage ends.
Micron kept its caps. It sets ceilings at second-quarter 2026 levels, with volume commitments and price floors across the term, and prices next-generation parts separately. That is a hedge: give up some upside, keep the shock absorber.
Samsung is locking in 60 to 70 percent of its capacity under deals that fix both price and volume. That is a bet on stability, and a rational one for the maker trying to win back share.
SK Hynix went the other way. It removed the cap and, unable to meet all demand anyway, keeps some volume outside contracts entirely to stay maximally exposed to the spot market. That is the most aggressive posture available, and it is the market leader taking it, at the moment its lead is narrowing.
What to watch
None of this says the correction is imminent. Demand may well outrun the new fabs into 2027, and if it does, SK Hynix will have made the correct call and printed the highest margins doing it. The point is narrower and harder to argue away: the thing being sold as proof the cycle is broken is proof of nothing. Long contracts moved the counterparty. Uncapped pricing kept the risk.
So the tell is not the length of the order book or the number of years on the deals. It is the point where HBM4 supply catches its demand, sometime after the second-half ramp, with Samsung back to a third of the market and Micron holding a hedge. On that day the fixed-price makers keep selling at their agreed number and the uncapped leader repriced its entire book to a market that just turned. The industry did not break the cycle. It kept the volumes, took the price protection off, and decided that this time it will win. The interesting quarter is the first one where it does not.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.