According to projections by SemiAnalysis, memory expenses are expected to constitute approximately 30% of total capital expenditures in hyperscale data centers by 2026, a significant increase from the 8% seen in 2023. This proportion is forecast to climb further in 2027. Over this brief four-year span, the share of spending allocated to memory is anticipated to nearly quadruple, driven by DRAM prices soaring to unprecedented levels and a persistent, severe shortage of HBM supply.
This surge is clearly fueled by artificial intelligence demand.
SemiAnalysis projects that memory spending will account for roughly 30% of hyperscale data center capex by 2026, up from about 8% in both 2023 and 2024. The firm predicts this ratio will continue its ascent in 2027. This implies that, within just four years, the portion of expenditure dedicated to memory will see an almost fourfold increase, coinciding with DRAM prices reaching unimaginable highs and HBM supply remaining critically tight.
SemiAnalysis anticipates that DRAM prices will more than double by 2026, with average selling prices expected to register double-digit growth again in 2027. Contract prices for LPDDR5 have already tripled since the first quarter of 2025, and the firm estimates that spot market prices could exceed $10 per GB this quarter.
Based on SemiAnalysis findings, HBM, the vertically stacked memory crucial for AI accelerators, is expected to remain in short supply through 2027. Memory already accounts for a substantial portion of the estimated $250 billion in new hyperscale data center spending projected for the current year.
This dynamic is already reflected in the pricing of AI servers. SemiAnalysis notes that the cost of B200 servers is projected to rise by up to 20% by year-end, significantly impacted by increasing memory costs. This aligns with industry-wide trends, as manufacturers have acknowledged substantial increases in component costs during recent earnings calls. Jeff Clarke, Chief Operating Officer at Dell, described the pace of these cost increases as "unprecedented" during the company's Q3 FY2025 earnings call in November.
Counterpoint Research has separately forecast that the price of DDR5 64GB RDIMM modules could double by the end of 2026 compared to early 2025. AI servers built on NVIDIA's LPDDR platform are seeing the most pronounced price increases due to the vast amount of memory required per system.
SemiAnalysis highlights an interesting observation: NVIDIA reportedly secures what is termed "VVP" pricing for DRAM from its suppliers, which is "far below the prices paid by hyperscalers and the broader market." The firm suggests this compresses NVIDIA's own server costs and depresses overall market pricing benchmarks, thereby masking the severe supply shortages actually faced by other buyers.
The situation for Advanced Micro Devices is markedly different. Its AI accelerator SKUs often carry a higher unit price, preventing it from securing similar preferential supplier pricing. With its current AI accelerator sales volume significantly lower than NVIDIA's, Advanced Micro Devices is "more vulnerable to memory cost increases while its AI accelerator scale remains much smaller than Nvidia's." In other words, NVIDIA's procurement scale for both HBM and conventional DRAM grants it advantages inaccessible to smaller-volume purchasers.
SemiAnalysis concludes that while major cloud operators have partially accounted for rising memory prices in their 2026 capital expenditure guidance, Wall Street expectations have not yet factored in the 2027 repricing. Samsung, SK Hynix, and Micron have all shifted capacity toward HBM and high-margin enterprise-grade DRAM, leading to constrained supplies of conventional DDR5 and LPDDR5. Furthermore, Micron's $9.6 billion Hiroshima HBM factory and SK Hynix's expansion projects in Icheon and Cheongju are not expected to contribute substantial production capacity until 2027 or 2028 at the earliest.
