JUST IN: Nvidia has informed some of its major customers that servers powered by its AI chips are getting ~17% pricier in many cases.

Memory inflation is pushing the prices of Nvidia’s Grace Blackwell and Vera Rubin systems sharply higher.

Memory is the pressure point because NVIDIA Vera Rubin NVL72 packs 20.7TB of HBM4 and 54TB of LPDDR5X in each rack.

TrendForce expects DRAM supply to remain tight through 2027, as AI servers keep pulling production toward HBM and server memory.

A 17% Nvidia server-price increase could add at least $5B to 1GW builds.

That extra $5B arrives before operators pay for the rest of the data center, including power, cooling, networking, buildings, and financing.

Cloud providers can absorb some of the increase, but passing it through would raise the cost of renting Nvidia compute for training and inference.

Nvidia's recent land-and-power investments look more consequential once each delayed rack is worth more. Nvidia has invested in Cloverleaf and SB Energy, both tied to securing sites and power for AI data centers.

A delivered GPU produces no revenue if the data center lacks power or a finished shell. When memory inflation raises the capital tied up in each deployment, customers have more money sitting idle during delays. Nvidia therefore has a financial reason to secure site readiness itself, reducing the chance that infrastructure delays push expensive hardware orders into later periods.