Morgan Stanley Raises Q4 2026 Memory Price Outlook as AI Supply Constraints Persist
Редакция ЦОД:Главное
Morgan Stanley, as reported by TechNews, has raised its fourth-quarter 2026 forecasts for contract-price increases across DRAM, NAND and enterprise SSDs, citing AI demand and constrained supply. The bank also expects a material expansion in new DRAM bit supply only in the second half of 2027.
- Morgan Stanley forecasts Q4 2026 contract-price increases of 10–15% for DRAM, 15–20% for NAND and 23–28% for enterprise SSDs.
- The bank expects US$98 billion of DRAM CAPEX in 2026 and US$150–160 billion in 2027, but sees meaningful new bit supply only in H2 2027.
- In an eight-rack interconnected AI configuration, Morgan Stanley estimates networking’s share of value rises from 10% to 20%.
- Projected 2026 shipments of NVIDIA VR200 NVL72 are below 10,000 racks, versus an initial target of about 15,000.
- Morgan Stanley’s supply-chain preference remains compute first, networking second and memory third.
Morgan Stanley has increased its forecast for fourth-quarter 2026 contract-price growth to 10–15% for DRAM, 15–20% for NAND and 23–28% for enterprise SSDs. Its earlier estimates were 3–8% for DRAM and 0–5% for NAND.
For AI infrastructure operators and server buyers, the revised outlook indicates continued procurement exposure across memory and storage as clusters scale. The source also reports that US hyperscalers have begun placing additional RDIMM orders for 2027, although the material does not provide volumes or terms for those orders.
Morgan Stanley expects total DRAM CAPEX to reach US$98 billion in 2026, up 82% year on year, and US$150–160 billion in 2027, or approximately 60% year-on-year growth. However, it expects a meaningful increase in physical DRAM bit supply only in the second half of 2027. This timing could leave the market tight even as memory producers increase investment.
The bank’s projections extend beyond conventional DRAM. It forecasts a 121% year-on-year increase in average HBM selling prices in 2027. For NVIDIA’s VR200 NVL72, the source assigns memory a 29% share of rack cost for a 54 TB configuration, compared with 19% for Rubin Ultra NVL72. Separately, projected VR200 NVL72 shipments for 2026 are below 10,000 racks, versus an initial target of about 15,000 racks. The source does not disclose the basis for that original target or the reasons for the projected shortfall.
Networking is increasingly material in Morgan Stanley’s AI supply-chain framework. In a configuration of eight interconnected racks, the bank estimates that networking’s share of system value rises from 10% to 20%, while memory represents 17%. Morgan Stanley ranks AI investment opportunities as compute first, networking second and memory third, with a selective approach to memory.
For data-center planners, these are forecasts rather than confirmed procurement outcomes. Still, simultaneous price pressure in server DRAM, HBM and enterprise SSDs could raise the IT-equipment cost of AI deployments and increase the value of longer-term supply arrangements. If multi-rack AI clusters develop along the configuration described by Morgan Stanley, the inter-rack network and its supply chain could account for a growing share of infrastructure budgets.
The forecasts point to potential cost and availability pressure for server memory and enterprise storage during AI data-center build-outs. Morgan Stanley’s assessment that networking rises from 10% to 20% of system value in an eight-rack interconnected configuration also suggests that inter-rack fabric is becoming a larger budget item alongside accelerators.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.