This is clearly driven by the demand for artificial intelligence.

SemiAnalysis predicts that...By 2026, memory spending will account for approximately 30% of total capital expenditure for hyperscale data centers, up from approximately 8% in 2023 and 2024. The company predicts that this percentage will climb further in 2027, meaning that in just four years, with DRAM prices soaring to unimaginable levels and HBM supply remaining severely insufficient, memory spending as a percentage of total expenditure will nearly quadruple.

SemiAnalysis predicts that DRAM prices will more than double in 2026, with average selling prices expected to see double-digit growth in 2027. LPDDR5 contract prices have more than tripled since the first quarter of 2025, and the company estimates that open market prices could exceed $10/GB this quarter.

According to SemiAnalysis' findings, vertically stacked memory (HBM), which is at the core of AI accelerators, will remain in short supply in 2027, and memory currently accounts for a large portion of the estimated $250 billion in new spending on hyperscale data centers this year.

This is already reflected in the pricing of AI servers. SemiAnalysis points out that the price of B200 servers is expected to rise by as much as 20% by the end of the year, severely impacted by rising memory costs. This aligns with industry-wide trends, with manufacturers acknowledging significant increases in component costs in recent earnings calls. Dell COO Jeff Clark described the rate of cost increases as "unprecedented" during the Q3 FY2025 earnings call last November.

Counterpoint Research previously predicted that by the end of 2026, the price of DDR5 64GB RDIMM modules could be double that of early 2025. AI servers built on NVIDIA's LPDDR platform will see the most significant price increases due to the massive memory requirements of each system.

SemiAnalysis points to an interesting phenomenon: Nvidia obtained so-called "VVP" (Very Very Preferred) DRAM pricing from its suppliers, "significantly lower than what hyperscale data centers and the market as a whole pay." SemiAnalysis argues that this compresses Nvidia's own server costs and lowers the overall market pricing benchmark, masking the severe supply shortages actually faced by other users.

AMD's situation is quite the opposite. Its AI accelerator SKUs typically have higher unit prices, thus preventing it from enjoying the same supplier discounts. Currently, AMD's AI accelerator sales are far lower than Nvidia's, making AMD "more vulnerable to rising memory costs due to its significantly smaller AI accelerator scale compared to Nvidia." In other words, Nvidia's large-scale procurement of HBM and traditional DRAM gives it an advantage that smaller-volume buyers cannot match.

SemiAnalysis concludes that while major cloud operators have partially reflected rising memory prices in their 2026 capital expenditure guidance, Wall Street expectations have not yet reflected the repricing in 2027. Samsung, SK Hynix, and Micron have all shifted capacity to HBM and high-margin enterprise-grade DRAM, leading to supply constraints on traditional DDR5 and LPDDR5. Furthermore, Micron's $9.6 billion Hiroshima HBM plant and SK Hynix's expansion projects in Icheon and Cheongju are not expected to reach substantial production until 2027 or 2028 at the earliest.

This article is sourced from: Semiconductor Industry Observation