Nvidia is reportedly warning its largest customers of a fifteen percent price increase on artificial intelligence servers, citing soaring memory costs ahead of Grace Blackwell and Vera Rubin shipments.
Nvidia has informed several of its largest customers that prices for servers equipped with its artificial intelligence chips will increase by more than 15% in many cases, Bloomberg reported Saturday. The price adjustments will apply to Grace Blackwell and Vera Rubin systems scheduled to ship in early 2027, according to individuals familiar with the matter who spoke regarding communications not yet made public. The exact magnitude of each increase will vary based on the specific chip generation and memory configuration involved. Server manufacturers building under contract for major data center operators, including Microsoft, Google, and Oracle, have recently alerted their clients to the impending cost increases.
This development underscores the ongoing “RAMageddon” gripping the dynamic random-access memory (DRAM) market, where contract prices have surged at record rates throughout the year. Analysts project that conventional DRAM contract prices will climb between 58% and 63% quarter-over-quarter in the second quarter of 2026, following a first-quarter surge of 90% to 95%. The escalation stems from memory suppliers reallocating production capacity toward high-bandwidth memory (HBM) and server-grade products. In October of last year, SK hynix announced it had already sold out its entire 2026 memory production capacity. Additionally, Samsung and SK hynix increased 2026 HBM3E supply prices by nearly 20% prior to the start of the year.
Artificial intelligence workloads demand substantial memory configurations. Nvidia’s Rubin GPU ships with up to 288GB of HBM4 per package, while the NVL72 rack-scale system integrates 72 of these GPUs, delivering over 20TB of HBM within a single rack before accounting for the LPDDR memory attached to its Vera CPUs. Because HBM production consumes approximately four times the wafer area of equivalent conventional DRAM, memory has become one of the most significant line items in an AI server’s bill of materials, and its costs continue to escalate at a stratospheric pace.
Ironically, the supply constraints currently inflating Nvidia’s server costs are a direct consequence of the demand it helped generate. Over the past two years, the three leading memory manufacturers have shifted advanced fabrication nodes and new capacity toward HBM and high-capacity server DRAM, effectively starving commodity markets. As a result, consumer DDR5 pricing has more than doubled since late 2025. According to RAM price tracking data, a mainstream 32GB DDR5-6000 kit retailed for approximately $392 in August, compared to $110 to $140 a year earlier.
Nvidia has already begun transmitting these rising costs downstream, having increased prices on GeForce graphics cards earlier this month. The Bloomberg report indicates that this unrelenting pressure has now reached the apex of Nvidia’s product stack, where hyperscalers and PC builders will absorb the increases. A 15% price adjustment on rack-scale systems that retail for several million dollars each translates to hundreds of thousands of additional dollars per rack across deployments spanning thousands of units.
Nvidia operates with a non-GAAP gross margin of approximately 75%, among the highest in the semiconductor industry. The reported price hikes suggest the company intends to pass memory cost inflation directly to customers rather than absorb it—a financial cushion it can easily sustain. Meanwhile, continued supply constraints on its accelerators from TSMC remain unable to satisfy demand, further limiting buyers’ immediate negotiating leverage. Whether these increases accelerate hyperscaler adoption of AMD’s accelerators or drive investment into proprietary custom silicon will hinge on how rapidly those alternatives can absorb displaced demand. Regardless of the path chosen, all options rely on HBM sourced from the same three constrained manufacturers.