Two years is a lifetime in the memory business, long enough for a boom to become a bust and back again. Yet Samsung, SK Hynix and Micron have already sold their 2027 production capacity, and Microsoft and Google are separately negotiating long-term supply agreements with the Korean memory makers, according to JoongAng Ilbo and multiple industry reports.
The sellout covers both halves of the AI memory diet: HBM, the high-bandwidth memory that sits next to accelerators, and standard DDR5, the workhorse memory in the servers around them. AI data centers consume both in unprecedented volume, and the suppliers’ capacity is committed to the buyers who signed first and largest.
The contract dynamics have inverted, and the inversion is the story. In the last memory cycle, suppliers begged customers to sign long-term agreements, offering price concessions to lock in demand against an uncertain future. Now customers are begging suppliers for allocation, and the suppliers are the ones setting terms. The power in the relationship has moved decisively.
Why 2027? Capacity decisions made today take two to three years to come online: fabs must be built, cleanrooms qualified, equipment installed and yields improved. By selling 2027 now, suppliers lock in prices, and customers lock in supply, and both sides avoid the annual ritual of negotiating in a market where there is nothing to negotiate over.
Microsoft and Google’s long-term talks reflect a procurement shift that has been building for a year. The cloud giants have moved from buying memory on the open market to reserving it like power: contracted years in advance, with penalties for interruption. The shift is a response to the 2022 crash, when memory prices collapsed and customers found themselves locked into high-priced contracts, but the current shortage has made the risk calculus different: the cost of being short is now higher than the cost of being locked in.
The sold-out status is partly a function of conservatism. The memory makers could add capacity more aggressively, and their customers are pressing them to do so, but the industry has refused to repeat the overbuilding that destroyed it in 2022 and 2023. Every new fab announcement is deliberately modest, and the suppliers have told investors that they will prioritize profitability over market share, a message that markets have rewarded with higher valuations.
The risk of the strategy is on the customers’ side. If AI demand slows, 2027 contracts become a liability: locked-in prices above a falling spot market, and allocation commitments that cannot be cancelled without penalties. The cloud giants are placing a bet that demand will hold, and their willingness to sign long-term agreements is the strongest evidence available that they believe it will.
For the memory industry, the sellout is the clearest sign yet that the current cycle differs from its predecessors. Memory has moved from commodity to custom: HBM stacks are co-designed with individual customers, tailored to specific accelerators, and contracted years in advance. The industry’s executives say the change is permanent, that memory is no longer a simple standardized commodity, and the sold-out calendar supports the claim.
The winners are the three suppliers, who control a market with no meaningful new entrants on the horizon. The capital costs of memory manufacturing are prohibitive, the technology curve is steep, and the customers are already locked in. New challengers face wafer-thin odds and shorter memories of the last crash, which is why none has committed to a serious entry.
The sellout also explains the industry’s pricing power. Memory makers have been raising contract prices steadily through 2026, and the sold-out calendar gives them room to keep raising: customers cannot walk away from a supplier whose capacity is the only capacity available. The negotiation dynamic has reversed so completely that suppliers now dictate terms the way buyers did in the last downturn.
The response from customers has been to buy earlier and buy bigger. The cloud giants are not just negotiating for 2027; they are discussing agreements that extend into 2028 and 2029, according to people familiar with the talks. The suppliers’ answer so far has been to sell allocation in stages, keeping some capacity free for the premium prices they expect to command as the shortage persists.
For the broader economy, the sold-out memory market is a cost story: every server, every phone and every AI system built over the next two years carries memory priced in a seller’s market. The industry’s executives say the prices are justified by the capital they are deploying and the risk they are taking, and the market’s response, a rising stock price for every memory maker, suggests investors agree, at least for now.
The tension ahead is the negotiation over 2028. Customers who locked 2027 will push to extend their agreements, and suppliers will sell 2028 at a higher price, arguing that the shortage will persist and that allocation is the scarcest resource in the industry. The cycle, in other words, is being sold forward one year at a time, and right now, every year is sold out.
