Every story about the memory crunch this year has been filed as a market story: prices are up, memory stocks are up, AI demand is enormous. SanDisk climbed again on Tuesday on news that hyperscale customers had locked in long-term capacity, and the Korean market went with it, with the KOSPI up 4.61% as SK Hynix gained 8.26% and Samsung Electronics added 5.68%. Filed that way, it reads like the top of a semiconductor cycle, and semiconductor cycles end.
That framing is missing the mechanism. What the memory makers signed this year are not purchase orders. They are multi-year supply agreements secured by customer collateral and default penalties, and the terms are what turn this from a cycle into something closer to a permanent allocation regime.
Read the Contract Structure, Not the Price
SanDisk’s executives have been unusually specific. Five signed long-term deals will cover more than a third of the company’s NAND shipments for fiscal 2027, and the chief executive has said the proportion could eventually pass half. The minimum revenue committed under just three of those contracts, signed in a single quarter, reached $42 billion. The agreements carry billions of dollars of customer collateral and explicit default penalties, alongside roughly $11 billion in financial guarantees across eight data-center and edge customers.
Collateral is the detail that matters. Buyers do not post collateral in a functioning spot market, because in a functioning spot market you can simply go and buy the part. They post collateral when supply is scarce enough that a place in the queue is itself the asset worth paying to protect.
SK Hynix has done the same thing from the other side. It has wrapped up long-term agreement talks with roughly ten customers and, according to reporting on its contract terms, scrapped the price ceilings that used to cap those agreements so it can capture spot pricing in full. Samsung is targeting a further increase of up to 20% in average DRAM selling prices in third-quarter negotiations, on top of a roughly 90% jump in the first quarter and another 50% to 60% in the second.
Buyers do not post collateral in a functioning spot market. They post it when a place in the queue is the asset.
The Supply That Disappeared Was Never Freed Up
The conventional read is that AI demand is temporarily crowding out consumer devices, and that capacity will rebalance. The allocation data says otherwise. Data-center demand reached roughly half of global DRAM consumption in 2025, up from 32% five years earlier, and manufacturers have been reallocating capacity toward high-bandwidth memory because it carries better margins than commodity DRAM does.
The result shows up as rationing rather than pricing. CNBC reported in March that major OEM customers were receiving only 50% to 66% of the memory volumes they had actually ordered. Tesla, Apple and a dozen other large manufacturers have signalled that memory availability is constraining what they can build. SK Hynix said its HBM, DRAM and NAND capacity was essentially sold out for 2026, and its chief executive told Bloomberg in July that the crunch will probably persist beyond 2030, possibly into the next decade.
Now put the contracts back on top of that. If a third to a half of a major supplier’s output is pre-committed under binding multi-year agreements, and the remaining suppliers are signing comparable deals, the volume left to clear in the spot market shrinks structurally rather than cyclically. A cycle resolves when capacity catches up with demand. This does not, because the new capacity is being sold forward before it exists.
What It Costs the People Who Did Not Sign
J.P. Morgan Global Research estimates DRAM prices will have risen more than 400% between the start of 2024 and the end of 2026. Gartner expects memory prices to climb roughly 130% by year end, feeding through to a 17% rise in PC prices and 13% on smartphones against 2025. Lenovo, Dell, HP, Acer and ASUS have already confirmed hikes of 15% to 20% and contract resets to their channel partners.
Those numbers are usually presented as a forecast. They are better understood as a floor. Every one of them assumes the buyer can get parts at the quoted price, and the whole point of the contract structure described above is that a growing share of buyers cannot. A laptop maker without a long-term agreement is not choosing between paying 20% more and paying last year’s price. It is choosing between paying whatever the residual spot market asks and not shipping the product.
BusinessTech.News has been tracking this pass-through as it moved up the stack, from Nvidia raising AI server prices more than 15% on memory costs to Dell guiding its AI server line lower because it could not buy enough memory. The consumer leg is the last one to land, and it lands hardest on buyers with no negotiating position at all.
Our Read
The memory industry is not experiencing a shortage in the way that word is normally used. It is executing a deliberate and highly profitable shift from selling a commodity into renting access to a rationed one, and the language of natural scarcity is doing a lot of work to make that sound like weather.
To be clear about what is and is not defensible: selling capacity forward is ordinary commercial behaviour, and no one is obliged to keep supplying low-margin consumer DRAM when data centers will pay more. Taking collateral from customers is aggressive but legitimate. What deserves scrutiny is the combination. Three suppliers control most of the world’s memory output. All three are simultaneously reallocating capacity toward the same higher-margin segment, removing price caps from long-term contracts, and pre-selling output years forward, while the residual market that everyone else depends on gets thinner. That pattern may be nothing more than three firms independently reaching the same obvious conclusion. It is also exactly what a tightening oligopoly looks like from the outside, and competition authorities in Washington, Brussels and Seoul have not asked a single public question about it.
They should. Not because the price increases are illegitimate, but because the market structure that produces them has changed materially in eighteen months and no regulator has established whether the reallocation is genuinely independent. Micron passing a trillion dollars in market value on the back of this ought to be enough to draw attention on its own.
The number to watch is not the DRAM contract price. It is the share of each supplier’s output committed under long-term agreements. SanDisk has told investors it wants that above 50%. If the rest of the industry follows, the spot market that sets consumer electronics pricing stops being where memory is actually sold, and the published price stops describing anything most buyers can get.