Nvidia Corp is raising server prices by more than 15% for customers using its artificial intelligence accelerators, according to confidential communications reviewed by Bloomberg News. The increases, which apply to systems scheduled for delivery in early 2025, reflect sustained upward pressure on memory chip costs and Nvidia’s dominant pricing position in the AI chip market.
The price adjustments affect servers equipped with Nvidia’s flagship AI chips, including the Vera Rubin and Grace Blackwell models, which command gross margins of about 75%. Industry participants note that the magnitude of the increase varies depending on the chip generation and memory configuration, with some configurations seeing steeper cost escalations tied to rising DRAM prices.
Major technology companies—including Microsoft Corp, Alphabet Inc’s Google, Oracle Financial Software, Amazon.com Inc, and Meta Platforms Inc—are among the customers notified of the higher costs. While these firms are developing proprietary AI chips to reduce reliance on external suppliers, they remain dependent on Nvidia for immediate data center expansion due to limited alternative sources of high-performance accelerators.
The price hikes compound existing challenges in the AI infrastructure buildout, which include project delays, labor shortages, tighter capital markets, and local opposition to data center construction. Memory chip suppliers such as Samsung Electronics Co, SK Hynix Inc, and Micron Technology Inc—responsible for the majority of global DRAM production—have also raised prices, further inflating server costs. Nvidia has separately increased prices for its PC graphics cards aimed at gaming, underscoring broader sector-wide cost pressures.
Analysts highlight that Nvidia’s pricing power stems from its near-monopoly in high-end AI accelerators, with customers paying tens of thousands of dollars per chip. The firm’s ability to pass through higher component costs reflects both its strong market position and the critical role its processors play in AI workloads.
