HBM allocations become a board-level item as memory makers ration 2026 output

SEOUL — Memory executives used to brief product managers. This quarter they are briefing boards. High-bandwidth memory for AI accelerators is being allocated in tightly written 2026 contracts, with take-or-pay language that looks more like LNG than DRAM.

Buyers said the conversation has shifted from price per gigabyte to guaranteed stacked-die supply. A hyperscaler that under-orders now risks a 2027 training cluster that cannot be populated. Over-ordering ties up cash in a part that has few alternative uses.

That rationing is already showing up in accelerator road maps. Some chip designers are delaying wider memory buses not because the controller is late, but because the stacks they were promised have been reassigned to a larger customer.