Infrastructure

Seventy Percent Of Memory Is Already Committed

Samsung told its Q2 call it will put 60 to 70 percent of memory capacity into multi-year contracts. That reprices everyone still buying on an annual cycle.

InfrastructureBy Servola Tech Desk2026-07-316 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Seventy Percent Of Memory Is Already Committed

Key takeaways

Samsung told its second-quarter call it will allocate approximately 60 to 70 percent of total memory capacity to long-term supply agreements, which means an annual purchasing cycle is now bidding for the third of the market that is left over.
The qualification for a contract is a forecast. Samsung said these customers are sharing mid to long-term demand forecasting and requesting multi-year supply agreements, and that is why they hold capacity while a buyer with a twelve-month horizon does not.
A contracted book changes what a published price describes. Once most volume clears at terms agreed earlier, the spot number tracks only the residual, so it stops telling you what the industry pays.
Server SSD is expected to exceed 60 percent of Samsung's NAND sales mix this year, up more than 20 percentage points in twelve months. Plan your 2027 refresh against a customer substitution, not a price cycle.

The sentence that was not in the coverage

A hardware buyer at a European logistics group spent the morning of 31 July reading the same headline everywhere: Samsung expects the memory shortage to worsen through 2027 and last until 2028. It is a true headline and it is not the one that changes a purchase order. The line that does was said further down the same earnings call, by Jaejune Kim, executive vice president and head of global sales and marketing for memory.

Samsung will place roughly two thirds of its memory capacity outside the open market. Kim said the company plans to allocate approximately 60 to 70 percent of its total capacity to long-term supply agreements while preserving flexibility for other customers. Read that as a structural statement rather than a commercial one. A shortage is a condition that passes. An allocation model is a decision about who is served first, and it outlives the condition that produced it.

Samsung also said who is signing. Frontier model developers are approaching the company directly, and in Kim's words these customers are sharing their mid to long-term demand forecasting and requesting multi-year supply agreements. The company put the duration plainly as well: supply constraints are expected to become even more severe in 2027 than 2026, reinforcing its view that the shortage will persist through 2028, with limited visibility beyond 2029.

What a contracted book does to a shortage

A market that is 60 to 70 percent contracted stops behaving like one market. Contracted volume does not clear at today's price. It clears at a price agreed months or years earlier, on terms that already assumed scarcity. What remains on the open market is the residual, and a residual absorbs the entire mismatch between demand and capacity because it is the only part of the supply that can move.

This matters for a specific and unglamorous reason: the published DRAM price stops being a description of the market and becomes a description of the leftovers. If you benchmark next year's hardware budget against spot memory prices, you are reading an index of the segment you are exposed to, not an index of what the industry pays. Two buyers can face the same component in the same quarter at prices that no longer converge, and neither number is wrong.

The corollary is harder. When the shortage does ease, it eases in the residual first, because that is where the flexible capacity sits. A buyer on an annual cycle therefore sees more of the downside now and less of the upside later, which is a worse position than the average of the two. Volatility is not distributed evenly across a market that has been split this way; it is concentrated in the part that did not sign.

The same reallocation, told in NAND

If the capacity split sounds abstract, the NAND mix makes it concrete. Samsung said server SSD is expected to exceed 60 percent of its NAND sales mix in 2026, up more than 20 percentage points year on year. A supplier's sales mix does not move by twenty points in a year because of pricing. It moves because the customer changed.

That is a substitution, not a cycle. A price cycle raises what everyone pays and then lowers it. A substitution reassigns the output to a different buyer and leaves the previous one to compete for what is behind. Samsung's own consumer businesses are the evidence: the company raised Galaxy phone and tablet prices and demand fell, while its semiconductor unit posted record sales in the same quarter. The group is absorbing on one side of the house what it earns on the other.

The capital line points the same way. Samsung reported second-quarter capital expenditure of 16.8 trillion Korean won, up 5.5 trillion sequentially, with 15.4 trillion of it going to Device Solutions. Capacity is being built. It is being built against forecasts that specific customers have already handed over, and equipment ordered in 2026 does not produce wafers in time to change a 2027 quote.

Forward visibility is a procurement instrument

The useful inversion is this: the scarce asset in Samsung's account is not memory. It is a credible multi-year demand number. The AI labs did not receive priority because they pay the most per bit. They received it because they were willing to state what they will need in 2028 and sign for it, which converts a supplier's capital risk into someone else's commitment.

Most mid-sized European buyers can produce that number and never do. A three-year committed refresh volume is not a harder forecast than the one already sitting in a depreciation schedule. It is the same number, stated to a different party, with a signature under it. Firms treat the annual tender as prudence, but in a contracted market an annual tender is a declaration that you will not commit, and it is priced accordingly.

There is a real cost to being wrong, and it should be named rather than waved past. A multi-year commitment on falling prices is a loss, and memory has fallen hard before. The judgement is not that committing is always right; it is that the choice is now between two priced positions rather than between a commitment and a neutral default. Staying on the spot market is a position, and after this call it is a more concentrated one than it was.

What a European buyer settles before 2027

Three questions belong in the next hardware review, and none of them is about price. First, ask your server or device supplier whether the memory in a 2027 quote is drawn from its contracted book or from the open market. Vendors know the answer and are rarely asked. It determines whether your quote is a forecast or a promise.

Second, ask for term rather than discount. In a market this tight a supplier can more easily give you 2028 certainty than a better 2027 number, because certainty costs it capacity it has already secured while a discount costs it margin it has not. The negotiation that is available is not the one procurement teams are trained to run.

Third, decide who inside your organisation owns a three-year volume commitment, because today the answer is usually nobody. Finance owns the budget, IT owns the refresh, and the forward number that would qualify you for a contract sits between them. In the United Kingdom the same question arrives with a currency exposure attached, since memory is dollar-priced and a sterling budget carries the swing on top of the shortage. Settle the ownership before the pricing conversation, not during it.

Frequently asked questions

Does a long-term memory contract actually protect against price rises?

It protects against exposure to the residual market, which is where the volatility concentrates once most capacity is committed. It does not protect against a price agreed too high, and it removes the benefit if prices fall. The point is that both staying out and going in are now priced positions.

We buy servers, not memory. Does this reach us?

It reaches you through your server vendor, which is itself either a contract holder or a residual buyer. That is why the question to ask is where the memory in your quote comes from. A vendor drawing on a contracted book can hold a price for you; one buying on the open market can only forecast it.

How firm is the 2028 date?

It is Samsung's own stated view from its second-quarter call, not an independent projection, and the company said visibility beyond 2029 is limited. Treat it as the largest supplier's planning assumption. That is worth more than a forecast and less than a fact.

Every outlet took the duration from this call and almost none took the allocation. The duration tells you how long conditions stay uncomfortable. The allocation tells you which half of the market the discomfort is assigned to, and that assignment was made in a sentence about supply agreements, not in the headline about 2028.

SamsungDRAMNANDMemory supplyProcurementData centres

Your 2027 Memory Is Already Spoken For

SK hynix customers want 60 to 100 percent more AI memory in 2027. The chairman wants prices lower. That combination tells a European operator what to contract now.

5 min read

Consumer Tech

Memory Prices Did What the Note 7 Could Not

Samsung's mobile division lost 0.7 trillion won, its first quarterly loss ever. During the Note 7 recall it still made money. Memory prices did this.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles