On the sales floor of Samsung Electronics’ memory business, the conversation has shifted. Instead of courting spot buyers through brokers, the team now divides a fixed pool of output among a handful of giant customers, and apologizes for what it cannot deliver. According to a report in South Korean media, the division has committed roughly 70 percent of its production capacity through 2031 to long-term agreements, or LTAs, with clients that include Nvidia, Microsoft and Google.
The commitments reflect the force reshaping the entire memory industry: the AI building boom. High-bandwidth memory, the specialized chips that sit beside AI accelerators, has become the most sought-after commodity in semiconductors, and the rush to produce it is crowding out conventional DRAM. Suppliers have little capacity left for the commodity memory that powers phones, PCs and servers, and buyers are feeling the squeeze. Spot prices for standard DRAM have climbed for months, and memory suppliers have told customers that allocations will stay tight through next year.
What makes the Samsung arrangement notable is its scale and duration. LTAs have existed in the memory industry for years, but they typically covered a slice of output and a horizon of a few quarters. Committing most of a fab’s capacity for five years locks in relationships that used to be renegotiated every cycle. “This is the memory industry operating like a foundry business,” said an analyst at a Seoul-based brokerage. “Capacity is reserved years in advance, and the spot market is becoming the residual.”
The structure also reveals how far the AI boom has stretched even the most powerful buyers. According to people familiar with the matter, the scramble for HBM has been so intense that Nvidia, Microsoft and Google, customers with nearly unlimited budgets, cannot secure the full contracted volumes from their suppliers. When a company as large as Microsoft cannot get what it signed for, it bids for whatever is available, and spot prices respond. The same dynamic has pushed memory makers to raise prices on existing contracts, and some are said to be renegotiating terms signed earlier this year.
For Samsung, the strategy is a hedge with two edges. If AI demand keeps growing, the LTAs provide predictable revenue and shield the company from the violent price swings that have historically defined memory cycles. If demand disappoints, Samsung is bound to deliver volumes at prices set in a stronger market, a risk its competitors have been quicker to flag. SK hynix, the market leader in HBM, has pursued a similar strategy but kept a larger share of capacity flexible, according to analysts. Micron, the third major supplier, has signaled it will remain more exposed to spot pricing.
The consequences reach beyond the three memory makers. Every company that buys DRAM, from phone makers to cloud providers to automakers, now faces a market where supply is pre-committed years into the future and shortages are priced in advance. PC makers have begun passing higher memory costs to consumers, and analysts expect the pass-through to continue. Some industry executives have warned that the shift could make the memory business more stable but also more expensive, with price discovery moving out of the open market and inside private contracts.
For the broader semiconductor economy, the memory crunch has become a bottleneck on AI deployment. Server makers say the lead time for high-capacity modules has stretched to months, and some AI projects are being delayed not by accelerator supply but by memory supply. The LTAs do not solve that problem; they concentrate it, giving a small group of suppliers and their largest customers control over most of the world’s memory production for the rest of the decade.
The economics of the AI build-out explain why customers accept the arrangement. Memory has become a larger share of the cost of a server than it was two years ago, with HBM stacks selling for several times the price of the DRAM they replaced, according to industry estimates. Hyperscalers, desperate to secure supply for training runs and inference fleets, have signed contracts that lock in volume at prices that would have seemed extraordinary in earlier cycles. The result is a memory market that behaves less like its old boom-and-bust self and more like the market for advanced foundry capacity, where customers pay years ahead for access.
Analysts said the real test will come in 2027 and 2028, when new HBM capacity comes online and the first LTA tranches mature. If demand stays strong, Samsung’s early commitments will look prescient. If the AI cycle cools, the company will carry contracts written in a boom into a downturn. For now, the market is pricing the boom. “Nobody is modeling the downside scenario in memory right now,” the analyst said. “That is precisely when the downside arrives.”
