Eighty percent of the memory chips Samsung Electronics plans to make in 2027 has already been claimed by buyers, and most of those commitments were signed before the company finished selling this year’s output. The estimate came from KB Securities in a note published Tuesday morning in Seoul, and it carried a conclusion unusual in any cycle: next year’s profit growth is essentially settled.
Kim Dong-won, who runs research at the brokerage, put the coverage at roughly 80 percent of planned production across DRAM, NAND and high-bandwidth memory. Five-year agreements with global technology companies account for about 70 percent of planned conventional DRAM and NAND output, according to the note, while the company’s entire projected HBM volume is spoken for. The uncontracted remainder, Kim wrote, is the subject of continuing competition among hyperscalers.
That order book differs from the ones memory makers carried into previous upcycles. In 2018 and again in 2021, the industry sold quarterly allocations at prices negotiated weeks before shipment, with the spot market setting the marginal dollar. A five-year contract with a named customer turns a commodity into something closer to a subscription, and it lets a manufacturer commit capital to a fab or a packaging line without guessing what the chip will fetch when the line starts.
The buyers have reasons of their own. Memory supply has been tight since the middle of 2025, when the wafer capacity that went into HBM came out of conventional DRAM and NAND, and the companies that build servers, phones and PCs have learned that a purchase order is not a supply agreement. Micron’s chief executive, Sanjay Mehrotra, told analysts on September 30 that more than 75 percent of its planned 2027 output had been committed and that customers were already negotiating allocations for 2028. SK hynix said in July that it had completed long-term agreements with about 10 major customers.
Korean trade press has reported that PC makers including Lenovo, HP and Dell have signed 2027 agreements, which leaves the companies that assemble smartphones with the least coverage and the most exposure to whatever price the contract holders leave behind. Handset makers buy on annual terms because their bills of materials are set by retail prices they cannot raise at will, and a sold-out year at the fabs leaves them with the weakest claim on supply.
The pricing inside those contracts is where the argument gets thinner. KB expects average selling prices for Samsung’s HBM to rise 100 to 150 percent next year, driven by HBM4, the generation built for Nvidia’s newest accelerators, and it expects HBM4 to double to about 80 percent of the company’s HBM revenue. A 100 percent increase inside a multiyear agreement implies either a renegotiation clause or a contract written with annual price steps, and the note does not say which. Long-term agreements in memory have historically included volume bands, price reviews and, in weak markets, quiet reopening.
What Samsung gets for accepting those terms is a schedule it can build to. The company has been the slowest of the three large makers to qualify its HBM with the most demanding customers, and a sold-out year removes the risk that new capacity arrives into a market that has stopped buying. It also gives management something to show while it spends: preliminary third-quarter results published on October 8 showed operating profit rising nearly ninefold from a year earlier, with detailed figures due on October 29.
What it gives up is the upside if spot prices run past the contract. That is a familiar trade in commodity industries, and it usually looks wise for about eighteen months. Memory has a documented habit of teaching the wrong lesson at the top of each cycle.
In 2018, makers signed multiyear deals with hyperscalers and data center builders, then watched prices halve within four quarters as new capacity came online and inventories backed up. The downturn that began in 2022 opened with NAND shipments down only 5 percent and prices down 20 to 30 percent, which was enough to erase operating profit at two of the three large makers.
The difference this time is who holds the inventory. If the contracts are take-or-pay, the customer carries the risk of a demand shortfall. If they are not, the manufacturer has swapped price risk for volume risk and booked it five years forward, and memory contracts rarely disclose which of the two they are.
A supply response is also building that no order book captures. All three large makers are converting conventional capacity to HBM and adding wafers, and equipment arriving next year will produce chips that were not in anyone’s 2026 plan. Coverage of 80 percent describes the shape of 2027, not the shape of 2028, and the remaining fifth of output is where the price of the marginal chip gets discovered.
For the customers, the calculation is a hedge. A hyperscaler signing a five-year agreement at today’s prices buys certainty for a buildout whose electricity bills and construction schedules are already uncertain. If demand compounds at the rate the largest cloud companies forecast, the contract will look cheap in hindsight. If it plateaus, it becomes a liability that finance departments will try to renegotiate.
Samsung’s shares have fallen with the rest of the sector, and Kim attributed the decline to macroeconomics rather than fundamentals, a defense that will be tested by what the company says on October 29 about 2027 prices, capital spending and HBM4 yields. The useful question is not whether Samsung can fill its fabs next year. It plainly can. It is what price the company agreed to in the months before anyone knew how long the boom would run, and whether the contracts it signed leave it any way to find out.
