Samsung Electronics and SK Hynix have raised high-bandwidth memory (HBM3E) contract prices by approximately 20% for 2026 deliveries, as AI accelerator demand from NVIDIA, Google, Amazon, and other hyperscalers vastly outpaces available production. Both manufacturers have reportedly sold out their entire 2026 HBM production capacity to long-term agreements, locking customers into multi-billion-dollar supply contracts. HBM3E currently trades at roughly $300 per 36GB stack, while HBM4 (expected to enter mass production later in 2026) is estimated at $500 per 48GB stack.
The supply crunch stems from the resource-intensive nature of HBM manufacturing, which consumes approximately three times more wafer capacity per bit than standard DDR5. NVIDIA's B200 Blackwell GPU alone requires eight HBM3E stacks costing $2,400 per chip, with a fully configured DGX B300 system (eight GPUs) demanding 768 DRAM dies just for HBM modules. Google's TPU v7 integrates eight HBM3E stacks per processor, while Amazon's Trainium3 uses four. Meanwhile, manufacturers are systematically deprioritizing consumer DRAM and NAND production: HBM now accounts for approximately 23% of DRAM wafer production despite being only a small fraction of total DRAM bits produced.
Architects and hyperscalers see this as a structural constraint on AI data-center scaling. Memory represents 30-40% of AI accelerator manufacturing cost, up from under 20% two generations ago, and lead times remain at 20-26 weeks for HBM3E stacks. SK hynix and Samsung are accelerating fab construction in South Korea and the U.S. (SK hynix's $3.87B Indiana plant, Samsung's Pyeongtaek expansion), but new capacity won't reach volume until 2027-2028. Architects need to plan multi-year memory commitments and evaluate performance-per-memory-unit tradeoffs, as spot pricing is unavailable and negotiation power is concentrated in the hands of three suppliers.