Nvidia warns customers of 15% price hikes on AI servers as memory costs soar

Nvidia has notified its biggest customers that prices on AI servers built around its Grace Blackwell and next-generation Vera Rubin chips will rise more than 15%, Bloomberg reported on August 22, 2026, with the exact increase varying by chip generation and memory configuration. The hikes apply to systems shipped starting in early 2027.

The increases are being driven not by the GPUs themselves but by the memory surrounding them. Server DRAM prices roughly doubled in the first quarter of 2026, and Counterpoint Research has tracked 80-90% quarter-over-quarter increases across DRAM, NAND, and HBM during that period. Deloitte projects AI-server DRAM prices will quadruple over the course of the year. Memory alone now accounts for roughly a quarter of the bill of materials on a high-end AI server rack.

A memory shortage with no easy fix

The server builders that assemble systems on contract for hyperscale data-center operators — including firms building for Microsoft, Google, and Oracle — are the ones passing the increases down the chain, since Nvidia itself doesn't sell finished servers directly to most customers. The shortage stems from memory manufacturers reallocating fab capacity toward the high-bandwidth memory used in AI accelerators, squeezing supply of conventional server DRAM at the same time demand for both is climbing.

Nvidia isn't alone in passing memory costs on to customers. Apple has raised prices on some products by as much as 20%, and Amazon increased the price of its Echo Dot smart speaker from $49.99 to $79.99 — a 60% jump — both citing memory cost increases tied to AI data-center demand pulling supply away from consumer electronics.

What it means for AI infrastructure spending

For hyperscalers and enterprises already committed to massive AI capital expenditure — much of it financed through debt, including Broadcom's reported $70 billion-plus financing package to support Anthropic's compute buildout — a 15%+ increase on server costs compounds an already steep spending trajectory. It also raises the question of how long consumer electronics and enterprise IT budgets can keep absorbing memory costs that are climbing faster than the underlying AI hardware itself.

Analysts see no near-term resolution: with fab capacity for advanced memory largely locked into multi-year contracts with AI chipmakers, the current supply squeeze is expected to persist well into 2027.