Memory Costs Surge, Nvidia AI Servers to Rise 15%!
According to Bloomberg, citing people familiar with the matter, Nvidia has notified some of its major customers that prices for AI servers equipped with its AI chips are expected to increase by more than 15% in most cases due to a sharp rise in memory chip costs.
The price increases are expected to apply to Nvidia AI systems scheduled for shipment in early 2027, covering product lines powered by its next-generation Vera Rubin and Grace Blackwell platforms. The actual increase will depend on the generation of Nvidia’s AI chips and the memory configuration used in each system. Some customers have reportedly described the adjustment internally as a revision to Nvidia’s pricing structure.
In addition to selling AI accelerators, Nvidia also provides complete AI server rack systems directly to customers and supplies key components to server manufacturers. As Nvidia raises the prices of its AI infrastructure products, other server makers that rely on Nvidia GPUs and related components are also expected to adjust their pricing.
Bloomberg reported that companies building AI servers for major data center operators, including Microsoft, Google, and Oracle, have already informed customers about upcoming price increases for Nvidia-based AI systems.
Morgan Stanley previously estimated that Nvidia’s Vera Rubin VR200 NVL72 rack system could cost around $7.8 million for hyperscale customers. Memory components alone account for approximately $2 million of the total cost, representing a 435% increase compared with the memory cost of Nvidia’s previous-generation GB300 rack system. The surge is largely driven by the rapid increase in memory chip prices since last year.
Based on a 15% price increase, each Nvidia AI server rack could see additional costs of more than $1 million.
Meanwhile, a recent report from semiconductor supply chain research firm Edgewater Research indicated that Nvidia has signed new multi-year HBM and DRAM supply agreements with SK hynix and Micron to secure long-term memory availability for its AI products. The agreements were reached even after Nvidia reportedly reduced the DRAM requirements for its next-generation Rubin AI chips, with actual memory usage potentially lower than initially expected.
As one of the most influential companies in the semiconductor industry, Nvidia’s need to adjust pricing amid rising memory costs highlights the growing pricing power of memory manufacturers such as Samsung Electronics, SK hynix, and Micron Technology as AI infrastructure demand accelerates.
Nvidia’s AI accelerators serve as the foundation for AI training and inference workloads, and their performance relies heavily on the capacity and bandwidth of accompanying DRAM and HBM memory. Samsung, SK hynix, and Micron currently control the majority of global DRAM production capacity. Although these companies continue expanding output, supply growth has struggled to keep pace with explosive AI-driven demand, pushing memory prices significantly higher.
The imbalance between supply and demand has transformed DRAM from a traditional commodity into a strategic resource for AI infrastructure. Many technology companies, including Apple, have faced increasing component costs and have been forced to raise prices for some products.
Despite being one of the semiconductor industry’s most profitable companies, Nvidia is now facing pressure from rising memory expenses. Nvidia’s AI chips can sell for tens of thousands of dollars per unit, while its AI accelerator business has maintained exceptionally high margins. However, continued demand growth and limited alternatives have driven Nvidia’s pricing power, which is now being challenged by escalating memory costs.
Earlier this month, Tom’s Hardware also reported that Nvidia had increased prices for some gaming-focused PC graphics cards.
Whether Nvidia’s higher AI server prices create opportunities for competitors will depend on whether rival chipmakers and companies developing their own AI accelerators can secure sufficient DRAM and HBM supplies.
Currently, technology giants including Amazon, Microsoft, Google, Meta, and OpenAI have developed their own AI chips, but they still rely heavily on Nvidia GPUs for large-scale AI data center deployments. Even if these companies expand the use of internally developed AI processors to reduce dependence on Nvidia’s systems, their ability to do so will ultimately depend on securing stable memory supplies from Samsung, SK hynix, and Micron.
