Nvidia has formally notified customers of AI server price increases exceeding 15%, driven by surging HBM memory costs and unmet demand.

Nvidia $NVDA plans to increase prices on servers equipped with its artificial intelligence chips by more than 15% in many cases, with the adjustments taking effect for systems shipped early next year, according to Bloomberg.

Systems built around Vera Rubin and Grace Blackwell chips are among those affected, though the exact percentage will vary depending on the chip generation and memory configuration. Nvidia did not immediately respond to requests for comment.

Contract server manufacturers supplying major cloud providers—Microsoft $MSFT Corp., Alphabet $GOOGL Inc.'s Google, and Oracle $ORCL Corp.—have recently notified customers of the impending price changes, Bloomberg reported, citing individuals familiar with the matter.

The primary catalyst for the hikes is the escalating cost of memory chips, a critical component of Nvidia’s GPUs and server architectures. These three companies collectively dominate global DRAM production; however, despite each ramping up manufacturing capacity, explosive growth in AI infrastructure spending has outpaced their supply capabilities, driving prices upward.

Nvidia currently operates with a 75% gross margin, and its chips retail for tens of thousands of dollars each—a premium driven by persistent supply shortfalls from contract manufacturer Taiwan Semiconductor Manufacturing Co. relative to buyer demand. That Nvidia—the market’s dominant force—is transferring these costs to customers rather than absorbing them underscores the strengthened negotiating leverage memory chip suppliers have gained amid surging AI capital expenditure.

Nvidia is also set to report fiscal second-quarter earnings next week.

The price adjustments arrive as Nvidia’s largest clients accelerate independent custom AI chip initiatives to lessen their reliance on Nvidia hardware. Google is already deploying its seventh-generation Ironwood chip, Amazon $AMZN's Trainium3 has been on the market since late 2025, and Microsoft’s Maia 200 inference processor began rolling out to U.S. data centers earlier this year. Nevertheless, all three remain heavily reliant on Nvidia hardware for the majority of their data center capacity.

The pace at which customers pivot to competing silicon in response to the price hikes will likely depend on their ability to secure sufficient memory inventory from the same manufacturers whose pricing pressures are currently squeezing Nvidia.