Nvidia has notified some of its largest customers that prices for AI servers built around its flagship accelerators will rise by more than 15%, moving the pressure from silicon shortages to the broader infrastructure bill.
The increases apply to rack-scale systems using Grace Blackwell and Vera Rubin configurations that are scheduled for shipment in early 2027. According to reports, the message was delivered to major hyperscalers and cloud providers through contract manufacturers, meaning the cost shock is expected to ripple across planned data-center expansions.
The immediate trigger is memory. High-bandwidth memory and server DRAM needed for large AI training and inference clusters have grown sharply more expensive as demand outpaces supply from Samsung, SK hynix, and Micron. Rather than absorbing the full shock, Nvidia is passing through a meaningful share of those increases at a time when customers are already locked into multi-year infrastructure programs.
The financial effect is not trivial. A Vera Rubin NVL72 rack, previously estimated near $7.8 million to $9 million, could move higher once memory and system costs are fully repriced. For cloud providers planning to add hundreds or thousands of racks, the cumulative effect can add hundreds of millions to capex plans before a single server is installed.
That prospect has revived an old anxiety in the AI buildout: whether the economics can hold if hardware, power, real estate, and memory all climb together. Some customers have already pushed back, while others are accelerating in-house chip work designed to reduce dependence on Nvidia's pricing power.
For now, Nvidia retains enviable leverage because its architecture remains the default for large training workloads and because the memory bottleneck is industry-wide, not company-specific. But the notice signals that the cheapest days of the AI infrastructure boom may already be behind buyers who locked in older assumptions.
Image source: i.ibb.co