Sold Out Is Not a Price

"Sold out" is not supposed to happen to a commodity. Commodities clear. Demand rises, price rises, the marginal buyer walks, and everyone still standing gets a quantity at a number. That is the entire mechanism. When a supplier tells you there is no quantity available at any price, the mechanism is gone — and if your capacity model still has a unit cost in the memory cell, it is modeling a market that no longer exists.

That is where DRAM sits as of this week.

The reporting stacked up fast over the last ten days. Samsung, SK hynix, and Micron have effectively booked their entire 2027 DRAM and HBM output. Buyers are being filled at roughly 60 to 70 percent of requested volume — not quoted higher, filled short. SK hynix's chief executive has called 2027 the worst supply year the industry has ever seen, and that's the guy whose company benefits from the shortage saying it. TrendForce has server DRAM contract prices up 13 to 18 percent quarter over quarter in 3Q26, with increases running through the back half of 2027. Server plus HBM is tracking toward roughly 70 percent of all DRAM bits shipped; every PC, phone, console, camera, and car on earth divides the other 30. Microsoft moved the Series X to $799 on August 1. It launched at $499.

The price moves are the part getting written about. They're the least interesting part.

You cannot populate the sockets you already own

Here's the number I keep going back to, buried in a TrendForce note in July and quoted almost nowhere: total RDIMM bit supply grows 15 to 20 percent year over year, and server CPU shipments are growing faster than that.

Read what that actually implies. Not "memory is expensive." There is less memory per server than there was. And it's showing up in procurement behavior already — cloud operators have been stepping down from 96GB and 128GB modules to 32GB and 64GB since early in the year. Not because they discovered they didn't need it. Because that's what they can get.

Memory per socket is falling in the middle of the largest compute buildout in industrial history. That is a genuinely strange sentence and it should stop you.

And it lands on exactly the wrong workload. Everybody watches the accelerator, because the accelerator has a ticker symbol attached to it. But serving a model at scale is a memory problem long before it's a FLOPs problem. KV cache scales with context length times batch size. Longer contexts, more concurrent sessions, bigger working sets — all of it cashes out as capacity and bandwidth, not arithmetic. The constraint moved to the part nobody was tracking. You can rent more compute. You cannot rent bits that were never fabbed.

The contract is the product now

Second-order effect, and this is the one with teeth.

Several of the large US cloud providers signed multi-year long-term agreements that cap what suppliers can charge them. TrendForce is explicit about the consequence: starting in the third quarter, the increases land on customers without LTAs, and on any volume bought outside an LTA. The hyperscalers aren't getting a discount. They're getting a different market.

So the market split in two, and which side you're on has nothing to do with how much you're willing to pay. It has to do with whether someone in your procurement org signed a five-year commitment in 2025. That's not price discovery. That's a queue with a date on it, and the date was set before most people knew there was a line.

Which inverts an argument a lot of engineering orgs have been having for three years. The case for pulling workloads out of the cloud and back onto owned iron was always arithmetic: rent forever versus amortize over five years, and the spreadsheet said amortize. That spreadsheet had a hardware cost in it. It did not have a procurement power term, because it never needed one — anyone could buy a server. The past six months added that term, and it's large enough to swamp everything else in the model. Cloud pricing didn't get better. Your alternative got worse, and it got worse in a way no unit-cost comparison will show you.

The reasonable objection

Memory is the most reliably cyclical business in technology. Shortage, panic, capex, glut, price collapse, tears. It has run that loop for forty years and there is no serious argument it has stopped. SK hynix is putting something like $38 billion into capacity. The relief is coming.

True, and mostly beside the point on the timeline that matters. Fab capacity is a three-year instrument and the first thing new capacity gets pointed at is HBM, because that's where the margin is — Samsung's HBM4 yields reportedly hit 80 percent last week, up from under 60 at February's ramp. Every wafer that improvement wins goes to accelerators, not to the RDIMMs in your refresh cycle. Commodity relief arrives after HBM relief, and HBM relief arrives after the buyers who prepaid.

The sharper version of the objection cuts the other way, anyway. If you believe the cycle, then the parties who "won" this round are the ones now holding three-to-five-year fixed commitments taken at the top. That's the trade they made for supply, and it may not look clever in 2029. Notice that SK hynix reportedly pulled price caps out of its long-term agreements while Micron kept them. The suppliers are not confused about which side of the cycle they're standing on.

Stop treating memory as a line item with a price attached. It's a queue position with a date attached, and until the fabs come online those are two different kinds of object. Plan against the one you're actually holding.

Every market runs on price until the day it runs on who you know. Find out which one you're in before you write the budget.

— Dustin