Pricing of memory chips is notoriously volatile. The production lines that make them cost billions of dollars and need to run at full capacity. This means that spot market pricing is extremely sensitive to any supply-demand imbalances. Prices crater when there is excess supply and soar when supply tightens. The IT market is currently witnessing the latter in the extreme.

The DRAM chip price increases in the second half of 2025 were anticipated by market researchers such as TrendForce, as well as by major producers including Micron, Samsung Electronics and SK Hynix. However, the brunt of the pricing shock was expected to be borne by consumer electronics, smartphones and PCs, rather than server products, with the overall market finding a new balance by early 2026. This did not happen.

The culprit is the added capacity strain caused by demand for AI compute. Because they can reap higher margins, the world’s largest memory manufactures have shifted more of their production lines from generic DRAM to high-bandwidth memory (HBM) used in data center GPUs made by Nvidia and AMD.

This shift, together with growing demand for high-performance DDR5 memory modules — which are also required by AI servers — rapidly depleted inventories, first at memory makers and then at IT vendors. By January 2026, DRAM spot prices had more than doubled year-on-year. By the end of the quarter, server memory prices in the supply chain are expected to have quadrupled compared with the first quarter of 2025. Depending on the amount of memory, server prices are likely to follow suit in the coming weeks and months.

In a related supply chain event, prices for NAND chips used to make large-capacity solid-state drives (SSDs) have also doubled, while the prices for conventional hard drives continue to climb.

As a result, organizations planning to expand their IT footprint in the year ahead will pay more than they initially budgeted for. Unprecedented price increases will inevitably delay a significant proportion of the planned IT projects as contracts are renegotiated; this will likely lead to some scaling back or even cancellations — slowing capacity uptake in data centers.

Major contributing factors to persistently high memory and SSD prices and their effects:

• HBM is gobbling up chipmaking capacity: A single high-end data center GPU used in AI workloads packs as much memory as a typical server. More crucially, HBM chips require roughly three times more wafer capacity than DRAM to reach the same memory capacity, due to their lower areal density and poorer yields. Despite these extra costs, profit margins on HBM products are still more attractive to memory makers than other products in their portfolio, which reduces available capacity to meet demand elsewhere.
• Demand is expected to endure: The demand for memory and storage silicon shows little sign of easing. Analysts at Morgan Stanley estimate that text-based AI inference alone could account for 35% of global DRAM output in 2026 and 92% of NAND. Memory makers report that their 2026 HBM production is already sold out for GPUs and warn of tight overall supply across all product lines.
• Extra manufacturing capacity will take time to materialize: Memory makers are increasing capital expenditure to expand manufacturing capacity. However, these new production lines are unlikely to come online until late 2026 or 2027. Almost all DRAM worldwide is made by just three companies: Micron in the US, and Samsung and SK Hynix in South Korea. High capital costs, deep technological knowledge and fierce market competition present major barriers to new entrants.
• Server pricing increases have only just begun: Server prices from major OEMs such as Dell, Lenovo and HPE are going up by at least 15%. Further price revisions are likely during the first and second quarters of the year. Inventory stockpiling by OEMs will likely exacerbate shortages.
• Doing more with less: As a result of the memory shortage, some organizations will postpone non-essential IT upgrades and refreshes, and it is reasonable to expect longer average server lifecycles in 2026. Not all workloads are affected by system memory capacity to the same degree, and modules can be easily moved between servers. In some cases, organizations will be able to sacrifice DRAM capacity or performance without a noticeable impact on the workload. 
  Some IT departments may limit hardware refreshes to CPUs, rather than entire servers — opting to keep existing DRAM modules in place. Others may deploy modern servers with minimal amounts of DRAM and upgrade them later as workloads grow. To be successful in either of these approaches, organizations will need to monitor workload performance and have accurate data on the physical location and configuration of their IT assets. An asset inventory alone will not be sufficient.

The following expert contributed to this report:
Rich Kenny, Managing Director, Interact