Goldman sees memory chip LTAs gaining traction amid tight supply and AI demand

Long-term memory contracts are starting to look less like a late-cycle warning and more like a fight for scarce supply. Goldman Sachs said the deals now getting traction reflect a market where customers want to lock in AI-era capacity, and suppliers are demanding better terms, including prepayments, not just longer commitments.

That matters because Goldman has been unusually blunt about the backdrop. In a February 8 report, the bank said the global memory market may face one of its most severe supply shortages in history in 2026 and 2027, with DRAM undersupply forecast at 4.9% in 2026 and 2.5% in 2027. NAND and high-bandwidth memory are also tightening as chipmakers shift capacity toward AI-optimized products.

The latest deal chatter suggests that pressure is no longer theoretical. Reuters reported on March 18 that Samsung Electronics and AMD signed a memorandum of understanding to deepen memory-chip supply ties for AI infrastructure, including Samsung’s HBM4 for AMD’s Instinct MI455X accelerators and optimized DDR5 for AMD’s EPYC processors. The agreement also left room for foundry cooperation, a sign that advanced memory and compute are being planned together, not separately.

Other market reports point in the same direction. CNBC reported on March 11 that hyperscalers are increasingly favoring long-term contracts over one-year agreements, and Micron said it had been in talks on a five-year strategic customer agreement. Bits&Chips reported on March 19 that Samsung was considering extending memory contracts from quarterly or annual terms to three to five years, while Micron was discussing five-year deals with several customers. Separate reports said Samsung was talking with Google and Microsoft about long-term memory supply arrangements, including more than $10 billion in prepayments from Microsoft and, in some cases, price-floor protections and upfront deposits equal to 10% to 30% of contract value.

For Goldman’s clients, the key question is whether this is a real structural turn or just the same old peak-cycle behavior in a different wrapper. The answer may hinge on supply. Samsung, SK Hynix and Micron are redirecting capacity toward AI memory, and new fabs are not expected to ease the market until late 2027 at the earliest. If that timeline holds, LTAs may end up looking less like exuberance and more like insurance in a market where access to memory, not just price, now decides who can ship AI infrastructure on schedule.