Nvidia tells customers AI server prices to rise over 15%

What happened: Bloomberg reported on 22 August that some of Nvidia's biggest customers have been told server prices will rise more than 15% in many configurations. The increases hit systems shipping in early 2027.

The details: Affected products include the flagship Vera Rubin and Grace Blackwell platforms. The size of the increase depends on chip generation and memory configuration. Contract server builders working for Microsoft, Google, Oracle and other large data-centre operators have relayed the news to customers.

The numbers: Blackwell NVL72 GB200 racks currently cost $2.8-3.4 million. Vera Rubin system inquiries run $5-7 million per unit. Server DRAM roughly doubled in the first quarter of 2026, and Counterpoint reported 80-90% quarter-on-quarter rises across DRAM, NAND and HBM. Memory now makes up about 25% of the cost of a high-end rack. Deloitte expects AI-server DRAM prices to quadruple over 2026.

Why it matters: Memory, not just the AI chips themselves, is now driving up the cost of building AI infrastructure. That cost ultimately feeds through to what AI labs and cloud providers charge for compute.

The rivals: Samsung, SK Hynix and Micron, which make most of the world's memory, stand to benefit from the price surge.

The catch: Nvidia has not confirmed the reported increases, and Reuters said it could not independently verify them.

What's next: The report landed days before Nvidia's earnings on 26 August, putting fresh scrutiny on how rising memory costs will show up in the company's guidance.