Managing the Server Memory Market Challenge of 2026

Introduction

This is the first in a short series of blogs meant to share insights from AMD experts to help customers address the challenges of the server memory market. The series will share insights on ways to reduce the need for and expense of memory, while retaining server performance. Subsequent blogs on this topic will drill down a bit further into memory optimization options and discuss how AMD and other enterprise management tools can help assess savings opportunities and modelling financial impact of IT procurement options.

Since late 2025 price increases and supply constraints have made planning and executing server acquisitions more challenging. While it is far from the only component with supply and price concerns, server memory has been an oft-cited example of a technology that is disrupting customer plans. The three leading memory vendors (Samsung, SK hynix and Micron) have acknowledged that rapidly growing demand for AI data centers has led to significant supply shortages, extended lead times and rising prices for standard DDR5 memory despite intense efforts to shift manufacturing capacity towards building the DDR5 and HBM used in AI servers.

In 2024 and early 2025, the practice of overprovisioning memory was relatively common in many businesses, the incremental costs were nominal, the memory was readily available and having robustly configured servers gave IT great flexibility in where they could deploy workloads. In today’s market, memory has transitioned from a secondary component cost to one of the primary drivers of overall server BOM cost. Pricing increases, constrained supply and extended lead times are directly impacting procurement decisions.

The combination of scarcity and cost may lead some IT managers to re-think the viability of new IT deployments. It is not unheard of for a system configuration that cost approximately $25,000 nine months ago to now be quoted in the $35-40,000 range, largely without corresponding performance gains. After all, increased cost while keeping performance and other aspects the same seems like a deal-breaker. Perhaps waiting until the supply/demand scenario improves is the smarter choice?

Worth Waiting?

What may seem like a prudent stand in isolation becomes less compelling in the dynamic business environment, and particularly so when the memory market constraints are expected to persist for several quarters (not weeks or even months). Companies invest in IT to provide robust, reliable infrastructure needed to support the innovation and growth. Companies have found tremendous economic value through IT—with modernization and consolidation projects that can pay for themselves and deliver compelling return on investment in months. While higher cost systems would make the economics seem less compelling, it is important to recognize that not investing also carries a cost. Deferred infrastructure investment can result in lost productivity, diminished competitiveness, reduced uptime and missed opportunities for consolidation and efficiency gains.

At the same time, organizations may want to avoid over-investing in a challenging market. As a result, defining deployment strategies that mitigate the current memory-related headwinds is essential. Given past tendencies toward over-provisioning and the wide variability in how memory affects workload performance, right-sizing configurations and selecting appropriate systems offer a practical path forward which still delivers business value. After all, system decisions extend beyond upfront cost: they can impact utilization, performance, and energy efficiency over the lifetime of the infrastructure.

One Size Does Not Fit All: How Do You Compare?

AMD analysis as well as some of the leading independent testing sites such as Phoronix and Serve the Home (STH) indicate that there is wide variability in how memory impacts workload performance. This variability reinforces that memory sizing should be workload-specific, not procurement-policy driven. And the high component costs and limited availability mean that overprovisioning of memory can be wasteful. To demonstrate how diverse memory needs are, consider Figure 1 showing some general recommendations from the AMD performance engineering experts that develop tuning and optimization guides with some of the world’s leading software vendors.

As you can see, application memory needs can vary from the very modest to the quite large. On the other hand, nearly all applications and workloads have some baseline need for memory capacity to deliver acceptable performance and stability. Certainly, these figures should be considered as general guidance points and not as final guidance on equipping your data center. This data is intended to provide a quick reference and estimate to help identify workload configurations where your organization may be over-investing in memory configurations. If application tuning experts often identify these as “sweet spot” configuration points, they are worthy of some consideration in many customer environments.

Moving IT Forward

This is just the first rock to turn over to look for opportunities to move IT forward despite current memory market headwinds. Our next blog will share additional methods and tools to evaluate memory provisioning decisions and “right size” IT infrastructure.