Nvidia warns customers: AI servers to cost over 15% more

Nvidia has reportedly told its biggest customers to expect price increases of more than 15 percent on AI servers shipping in early 2027. Affected are Grace Blackwell and Vera Rubin systems — exactly the hardware at the center of the current AI data center build-out.

The driver is not a GPU shortage but memory: high-bandwidth memory (HBM) has become scarce and expensive. SK Hynix has already sold out its entire 2026 memory production, and together with Samsung it raised HBM3E prices by close to 20 percent before the year even began. DRAM contract prices surged 90 to 95 percent in the first quarter of 2026, according to Tom's Hardware.

The scale follows from the architecture: a single Rubin GPU carries up to 288 gigabytes of HBM4, and an NVL72 rack holds more than 20 terabytes of HBM. Tom's Hardware notes that HBM production consumes "roughly four times the wafer area of equivalent conventional DRAM" — making memory the dominant cost factor. On systems priced in the millions, a 15 percent increase adds hundreds of thousands of dollars per rack.

The bill lands first with the hyperscalers: Microsoft, Google and Oracle are among the largest buyers. Their extra costs are likely to reach enterprise customers through cloud pricing over time. PC buyers are already feeling the memory squeeze through rising RAM prices.

For the industry, the warning marks a turning point: for the first time since the AI boom began, Nvidia is visibly passing on cost pressure instead of absorbing it. The bill for the AI build-out moves one step closer to the customers.