Memory has become an increasingly important factor in the cost of data center infrastructure, particularly as data center capex is projected to soar past $3 trillion by 2030. Strong demand for conventional DRAM, which is used across all server types, and high-bandwidth memory (HBM), which is used in AI accelerators such as GPUs, has driven memory prices sharply higher. Supply constraints have further intensified pricing pressures, increasing the cost of both general-purpose and AI-optimized servers.
DRAM Inflation Could Add Hundreds of Billions to Server Spending
DRAM pricing has risen significantly over the past year. Dell’Oro Group currently projects server DRAM average selling prices (ASPs) to reach approximately $10/GB in 2026, before gradually moderating toward the $5/GB range by 2030.
The impact on server spending is substantial. As a sensitivity analysis, if DRAM ASPs were instead held at approximately $3.50/GB from 2026 through 2030, roughly in line with pricing prior to the recent increase, higher DRAM pricing could add nearly 10% to cumulative server spending through 2030.
The impact will be most pronounced in 2026 and 2027, when we expect DRAM pricing to remain near its peak. Pricing should begin to moderate more meaningfully from 2028 onward as additional supply enters the market and the industry responds to the higher cost of memory.
Our outlook for the overall DRAM market has consequently increased significantly. The increase is not driven by pricing alone. We have also raised our forecast for DRAM bit demand alongside higher server unit shipments, supported by emerging agentic AI and storage-related workloads that are increasing demand for general-purpose servers in addition to AI-optimized systems.
However, we do not believe today’s elevated DRAM pricing can persist indefinitely without affecting the pace of infrastructure investment. Memory suppliers are expanding production capacity, including new capacity from China, while server and processor vendors are developing more efficient approaches to memory utilization. Together, these developments should help alleviate supply constraints and gradually normalize DRAM pricing over the forecast period.
HBM Faces a Different Cost Challenge
HBM represents an even more important consideration for AI infrastructure. Demand will continue to rise as high-end accelerators incorporate increasing amounts of memory to support larger models and more memory-intensive inference and reasoning workloads.
We project the average high-end accelerator to contain more than 500 GB of HBM by 2030. That estimate could prove conservative. AMD’s recently launched MI455X, for example, already incorporates 432 GB of HBM, illustrating how rapidly memory capacity per accelerator is increasing.
Unlike conventional DRAM, however, HBM may not benefit from the same long-term cost-per-bit declines. Increasing stack heights, density, bandwidth, and packaging complexity could keep the cost per bit elevated even as the technology advances. As a result, simply scaling today’s HBM architecture to ever-larger capacities could significantly increase the cost of future AI systems.
Continued advances in HBM manufacturing and advanced packaging will therefore be critical. Technologies such as hybrid bonding, higher-density HBM generations, improved yields, and eventually lower HBM ASPs could help offset some of the cost associated with rapidly increasing capacity per accelerator.
At the same time, the industry may need to reconsider how much HBM actually needs to reside directly alongside each accelerator. NVIDIA’s Rubin Ultra, for example, could ultimately incorporate less HBM than the approximately 1 TB originally envisioned. Improvements in model efficiency and memory management could reduce HBM requirements, while emerging technologies such as high-bandwidth flash (HBF) could complement HBM as a lower-cost, higher-capacity memory tier.
Memory Efficiency Will Become Increasingly Important
The AI infrastructure industry has so far focused heavily on increasing compute performance, but memory economics could become an equally important constraint on future system design.
For conventional servers, elevated DRAM pricing could add hundreds of billions of dollars to infrastructure spending over the next several years. For accelerated servers, rapidly increasing HBM capacity and complexity could place additional pressure on already-high system costs.
Memory suppliers also face an increasingly complex balancing act in allocating production capacity across conventional DRAM, HBM, and NAND flash. With demand rising across servers, AI accelerators, and storage, suppliers will need to prioritize capacity and investment across these markets. The recent increase in DRAM and NAND pricing could make capacity expansion in these markets more attractive, complicating suppliers’ decisions on how aggressively to prioritize HBM production.
The industry response will therefore need to come from both sides: more memory supply and lower manufacturing costs, but also more efficient use of memory within servers and AI systems. Improvements in memory architecture, packaging, software, and additional memory tiers could become increasingly important in determining how economically the industry can continue scaling AI infrastructure through the end of the decade.
