Global memory technology: BofA Global Research sees a long-term memory super-cycle, with global DRAM and NAND sales reaching $2.0tn by 2030.

Overview

This global memory-industry update argues that AI infrastructure is reshaping DRAM and NAND demand, supporting an extended pricing and earnings cycle rather than a conventional short-lived upturn. BofA Global Research raises its long-term market forecast and remains constructive on HBM and memory fundamentals despite expected periods of price normalization.

Core views

BofA Global Research raises its 2030 global DRAM-plus-NAND sales forecast to $2.0tn from $1.8tn and describes the outlook as a long-term memory super-cycle. The report notes that the industry previously peaked at $0.17tn in 2018 and generated $0.21tn in the 2025 upturn, making the new forecast a major step-up. Its case rests on a 3Q26 annualized sales run rate above $1.0tn: quarterly sales of $278bn imply a four-quarter equivalent of $1.1tn. It forecasts DRAM plus NAND sales of $0.9tn in 2026 and $1.4tn in 2027, followed by a 21% four-year CAGR from 2027 to 2030. The report attributes this to 2026 contract-price and ASP gains of three to four times year on year, high-teen annual bit growth through 2026-30, persistent supply tightness, and a richer mix of HBM, SOCAMM and enterprise SSD products. On its illustrative valuation logic, $2.0tn of sales, a 50% operating margin, a 25% tax rate and 13x P/E imply a $10tn fair market capitalization for the memory industry. AI systems are the central demand driver. Servers, particularly systems using HBM, are expected to account for more than half of DRAM shipments and sales in 2026 because memory content per server is rising sharply. The report does not expect lower memory specifications in new GPUs or ASICs: it assumes Rubin uses 288GB to 384GB of HBM and Rubin Ultra reaches 1,000GB. HBM sales are forecast to rise to $77.4bn in 2026, up 124% year on year, and $152.8bn in 2027; the report expects a roughly 50% industry operating margin in 2026 and no 2027 downturn because volume growth and cost reductions should offset price cuts. It expects HBM revenue to reach $276.9bn by 2030, while HBM capacity rises to 1TB in Rubin Ultra in 2027 and more than 1.5TB in Feynman Ultra in 2030. NVIDIA’s GPU roadmap, AMD’s MI-series, cloud ASICs and hyperscaler AI infrastructure are presented as the mechanism increasing HBM demand. The report also links memory demand to continued hyperscaler spending. It forecasts combined capital expenditure by Amazon, Microsoft, Alphabet, Meta and Oracle to grow about 100% year on year to $730bn in 2026 after 65% growth in 2025, with annual capex exceeding $1tn in 2027-28. Cloud revenue is projected to grow 35-45% year on year in 2026-28, while AWS operating margin remains above 35%, Azure above 40%, and Google’s margin is expected to exceed 35%. These economics underpin continued AI investment, although the report expects several hyperscalers to experience temporarily negative free cash flow in 2026-27 as infrastructure spending outpaces near-term cash-flow growth. Near-term pricing evidence remains strong. Channel checks indicate most DRAM products rose 20-30% quarter on quarter in 3Q despite increasing long-term agreements, while NAND rose more than 15%; the report expects at least a single-digit ASP increase in 4Q. New hyperscaler agreements reportedly set higher DRAM ASPs for 1H27 than in 4Q26. Accordingly, BofA makes no 2H estimate changes but raises 2027-28 DRAM ASP assumptions by 8-12% and NAND ASP assumptions by 2-3%, lifting corresponding revenue forecasts. It expects DRAM ASP of about $17, $17 and $16 per 8Gb equivalent in 2026, 2027 and 2028, while NAND ASP is about $9-10 per 256Gb equivalent. The report nevertheless expects a 2028 soft landing: DRAM ASP is forecast to fall about 5% and NAND ASP about 13%, before renewed growth in 2029-30 as more advanced AI chips use high-end memory. Supply conditions reinforce the pricing view. DRAM is expected to remain tight and mostly fully utilized, while NAND inventories may rise through 2027 but remain normal or only slightly elevated. The report distinguishes HBM-related investment from conventional supply expansion: DRAM equipment spending rises strongly because of HBM, but conventional DRAM and NAND spending remains broadly disciplined. DRAM capex is forecast at $97.7bn in 2026 and $128.6bn in 2027; NAND capex is forecast at $33.0bn and $39.2bn, respectively, partly reflecting a low 2025 base. The report argues that supply reallocation toward HBM and server DRAM has supported unprecedented conventional DRAM prices, including 16Gb DDR5 at about $55 and DDR4 at about $90, far above the prior roughly $10 peak in October 2017. The BofA Memory Indicator remains the report’s high-frequency confirmation of the super-cycle. It was 180 in July after reaching 190 in March and April, compared with a mid-cycle level of 100 and a prior 2017-18 peak of 120. The indicator combines year-on-year momentum in ASPs, revenue, spot prices and exports; the report raised its ceiling to 240 because of the unprecedented recent surge. Supporting data include Korean semiconductor exports of $47bn in August, up 14% month on month and about double January’s $21bn; year-on-year Korean semiconductor-export growth of 209%; Nanya August sales of NT$44.7bn, about triple January’s NT$15.3bn; and China IC imports of $65.8bn in August, up 84% year on year. The report notes that DRAM spot prices remain exceptionally strong, while NAND spot prices softened in September after an August recovery. For equities, the report says memory stocks corrected in July-August amid global uncertainty and AI-peak concerns but recovered in early September. It states that DRAM names remained two to three times their early-2026 levels and NAND names three to four times those levels. It argues that fundamentals remain strong because of HBM demand, NVIDIA commentary and hyperscaler capex, and notes that most DRAM stocks trade at four to seven times 2026-27 EPS.

Analysis framework

The report combines a top-down forecast of DRAM and NAND supply, demand, bit growth and ASPs with bottom-up comparisons of major memory producers’ quarterly results and long-term capacity plans. It cross-checks the outlook against channel checks, spot and contract pricing, export data, company sales, AI accelerator memory specifications, hyperscaler capex and the BofA Memory Indicator.

Methodology notes

• Industry AnalysisSupply-demand framework

  Top-down DRAM and NAND supply-demand, capacity, bit-growth and ASP forecasting.

  The report estimates revenue by combining projected bit demand and average selling prices, then tests whether capacity, inventories and utilization point to sustained tightness or normalization.

• Industry AnalysisVolume-price decomposition

  Revenue decomposition into shipment/bit growth and ASP changes.

  The report attributes the expected expansion in memory sales to both high-teen bit growth and unusually strong pricing, with ASP revisions driving the 2027-28 forecast increases.

• Other

  BofA Memory Indicator.

  The indicator aggregates year-on-year momentum in ASPs, revenue, spot prices and exports to assess the strength of the memory cycle; the report uses it as corroborating, indicative evidence rather than a benchmark.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• SK Hynix

  Major HBM producer expected to retain a stronger HBM earnings outlook than Samsung.

  Strengths

  Successful 12-Hi HBM ramp and a high base of HBM execution.

  Weaknesses

  2026 year-on-year growth and margin improvement may be less pronounced because of its high 2024-25 base.

  Comparison

  The report expects Samsung’s 2026 HBM growth and margin improvement to be stronger year on year, while SK Hynix retains the stronger overall HBM earnings outlook.

  Risks

  HBM pricing and volume depend on continued AI-accelerator demand.

• Samsung

  Major DRAM, NAND and HBM producer with potential HBM upside.

  Strengths

  Potential normalization in HBM pricing and partnerships with NVIDIA and hyperscalers.

  Weaknesses

  Its HBM position lagged SK Hynix’s successful 12-Hi ramp in 2025.

  Comparison

  The report expects stronger 2026 year-on-year HBM growth and margin improvement than SK Hynix because Samsung starts from a lower base.

  Risks

  AI-peak concerns and HBM pricing dynamics could affect the outlook.

• NVIDIA

  Key downstream AI-accelerator driver of HBM demand.

  Strengths

  Its B300, Rubin and Rubin Ultra roadmap materially increases HBM content per system.

  Comparison

  The report also identifies AMD GPUs and cloud ASICs as additional HBM-demand drivers.

  Risks

  Memory-demand assumptions depend on the pace of GPU and AI-system deployment.

• AMD

  Additional AI-accelerator source of HBM demand.

  Strengths

  The report expects 10-30% HBM content growth in AMD GPUs, led by MI350X and MI400.

  Comparison

  AMD’s roadmap is considered alongside NVIDIA and hyperscaler ASIC programs.

  Risks

  Demand depends on adoption of its AI accelerator platforms.

Key data

• 2030 global DRAM plus NAND sales forecastUS$2.0tnRaised from US$1.8tn.
• 3Q26 annualized memory sales run rateUS$1.1tnBased on 3Q26 sales of US$278bn.
• 2027-28 DRAM revenue forecast revision+8-12%Driven primarily by higher ASP assumptions.
• 2027-28 NAND revenue forecast revision+2-3%Driven by stronger ASP expectations.
• 2026 HBM sales forecastUS$77.4bn+124% year on year.
• 2027 HBM sales forecastUS$152.8bnThe report expects continued growth rather than a downturn.
• BofA Memory Indicator180 in July 2026Record-high level versus 100 mid-cycle and a prior 120 peak.
• Top hyperscaler 2026 capexUS$730bnAbout 100% year-on-year growth forecast for Amazon, Microsoft, Alphabet, Meta and Oracle combined.

Impact & implications

The report concludes that AI-led growth in memory content, particularly HBM and enterprise SSD, is changing both the demand mix and the pricing power of the industry. It expects tight DRAM conditions and disciplined conventional supply to support earnings, while identifying a softer 2028 pricing phase before renewed growth in 2029-30.

Risks

• The report expects DRAM ASP to fall about 5% and NAND ASP about 13% in 2028 as the market enters a soft landing.
• NAND spot prices were expected to soften in September and 4Q26 after the August recovery.
• Smartphone and PC production is assumed to decline in 2026-27 because of memory shortages and expensive bills of materials.
• Record AI infrastructure spending could temporarily drive several hyperscalers to negative free cash flow in 2026-27.

What to watch

• DRAM and NAND spot and contract pricing, particularly whether DRAM remains tight and whether NAND softness persists.
• 3Q and 4Q memory ASP developments and higher DRAM pricing under hyperscaler agreements for 1H27.
• HBM demand and product ramps for NVIDIA B300, Rubin and Rubin Ultra, AMD MI-series, and hyperscaler ASICs.
• Hyperscaler AI capital expenditure, cloud-revenue growth and cloud operating margins.
• The BofA Memory Indicator, Korean semiconductor exports, Nanya sales and China IC-import data.