Nvidia has told its largest customers that AI servers will cost more than 15% more in many cases, with some configurations seeing increases of approximately 17%, for shipments starting early next year. The increases affect systems built around the Grace Blackwell and Vera Rubin chip generations, with the magnitude depending on chip generation and memory configuration. Contract manufacturers have already notified customers including Microsoft, Google, and Oracle.
Memory costs are the primary driver. The Vera Rubin NVL72 rack requires 20.7TB of HBM4 and 54TB of LPDDR5X per unit. Samsung, SK Hynix, and Micron control most of the world's DRAM production, and output has not kept pace with AI infrastructure demand. TrendForce expects DRAM supply to remain tight through 2027 as AI server demand pulls production toward HBM. The same memory cost pressure is affecting Apple, Qualcomm, AMD, and AWS; AWS has already raised GPU prices by 20%. A 17% price increase adds approximately $5 billion to the cost of a 1GW data center build.
Nvidia reported record quarterly revenue of $96.2 billion and confirmed that demand exceeds supply, with chip prices rising by 15% or more. Despite the record results, shares initially fell about 4% before recovering to finish approximately 4% higher. Analysts noted that Nvidia's chip design accounting structure, where chip designs are not counted as exports while the resulting chips count as imports, causes U.S. GDP growth to be understated by roughly 0.3%.
Nvidia has also invested in land and power infrastructure, including through entities such as Cloverleaf and SB Energy, partly to reduce deployment delays that leave expensive hardware idle.