From 8% to 30%! Memory expenditure skyrocketing

Obviously, this is driven by artificial intelligence demand.

SemiAnalysis projects that by 2026, memory spending will account for about 30% of total capital expenditures for hyperscale data centers, up from roughly 8% in 2023 and 2024. The company expects this share to rise further in 2027—meaning that within just four years, as DRAM prices surge to unimaginable levels while HBM supply remains severely short, memory spending’s share will increase by nearly four times.

SemiAnalysis projects that 2026 DRAM prices will more than double, and that average selling prices in 2027 will also grow by double digits. LPDDR5 contract pricing has risen more than threefold since the first quarter of 2025. The company estimates that public-market prices this quarter could exceed $10/GB.

Based on SemiAnalysis’s survey results, the vertically stacked memory HBM—core to AI accelerators—will still face a supply shortfall in 2027, while memory already makes up a large share of the estimated $250 billion in incremental spending by hyperscale data centers for this year.

This is already reflected in the pricing of AI servers. SemiAnalysis notes that, severely affected by rising memory costs, the price of B200 servers is expected to increase by as much as 20% by year-end. This aligns with the broader industry trend: manufacturers have all acknowledged large increases in component costs during recent earnings calls. Dell COO Jeff Clarke described the pace of cost increases as “unprecedented” during Dell’s FY2025 Q3 earnings call last November.

Counterpoint Research previously projected separately that by the end of 2026, the price of DDR5 64GB RDIMM modules could be twice that of early 2025. AI servers built on NVIDIA’s LPDDR platform see the most significant price increases because the amount of memory required per system is so large.

SemiAnalysis points to an interesting phenomenon: NVIDIA obtains so-called “VVP” (Very Very Preferred, extremely preferred) DRAM pricing from suppliers, “far below the prices paid by hyperscale data centers and the broader market.” SemiAnalysis believes this compresses NVIDIA’s own server costs and also lowers the overall market pricing benchmark, masking the severe supply shortages that other users are actually facing.

AMD’s situation is the opposite. Its AI accelerator SKUs typically have a higher unit price, so they can’t benefit from the same supplier-preferred pricing. AMD’s AI accelerator volume is also far lower than NVIDIA’s, which makes it “more susceptible to rising memory costs when its AI accelerator scale is far smaller than NVIDIA’s.” In other words, NVIDIA’s procurement scale in HBM and traditional DRAM gives it advantages that small-batch buyers can’t match.

SemiAnalysis’s conclusion is that although major cloud operators have partially reflected rising memory prices in their 2026 capital expenditure guidance, Wall Street’s expectations have not yet priced in the reassessment of 2027. Samsung, SK hynix, and Micron have all shifted capacity toward HBM and higher-margin enterprise DRAM, resulting in constrained supply of traditional DDR5 and LPDDR5. In addition, Micron’s $9.6 billion Hiroshima HBM plant and SK hynix’s expansion projects in Lichuan and Cheongju will not be able to reach meaningful production until at the earliest 2027 or 2028.

Source of this article: Semiconductor Industry Observer

Risk Warning and Disclaimer

        There are risks in the market; invest prudently. This article does not constitute personal investment advice, and it does not take into account any particular users’ special investment objectives, financial conditions, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article align with their specific circumstances. You bear responsibility for any investment made based on this.