The memory crunch that already lifted PC prices is now reaching the most expensive hardware in the AI stack. Nvidia has alerted its largest customers that server systems will cost more than 15% extra when they ship early next year, with the notices flowing through contract assemblers that build machines for Microsoft, Google and Oracle.
Both the Grace Blackwell generation and the newer Vera Rubin line are caught in the increase, and the exact size of the rise tracks the memory configuration of each system.
The squeeze traces back to DRAM supply. The three dominant memory suppliers, Samsung, SK hynix and Micron, have been unable to keep output in step with AI demand, and analysts projected conventional DRAM contract prices up 58% to 63% quarter over quarter in Q2. The appetite is enormous: Nvidia’s Rubin GPU carries up to 288GB of HBM4 per package, and an NVL72 rack holds more than 20TB of HBM before counting the LPDDR on its Vera CPUs.
Nvidia runs a gross margin near 75% and still chose to pass the cost along. It raised GeForce prices earlier this month, AMD matched within days, and Amazon Web Services lifted GPU prices 20%.
The burden now lands on hyperscalers and, eventually, their customers. Some buyers will speed up plans to lean on AMD accelerators or in-house silicon, though every alternative draws HBM from the same three constrained makers. Nvidia’s quarterly earnings arrive next week, when the market will see how far the memory bill travels.
