BofA Lifts Memory Forecasts to $573bn in 2026 DRAM Sales as Legacy DDR4 Spot Trades 69% Above DDR5
BofA Global Research rebuilt its industry memory model this week on the resumption of SK Hynix coverage, and the headline output is a global DRAM industry sales path of $134bn in 2025 to $573bn in 2026, then $847bn in 2027 and $917bn in 2028. The 2026 figure is 328% year-over-year growth in an industry that grew 52% in 2025 and 86% in 2024. Numbers that size normally signal a modelling error. Here they signal that the entire revenue base repriced inside twelve months.
The decomposition is where the note earns its keep. DRAM bit shipments grow 25% in 2026 and 19% in 2027 — respectable, but nothing that explains a quadrupling. DRAM ASP grows 242%. NAND runs the same shape: bits up 20%, ASP up 267%. Almost the whole revenue increase is price, and price at this magnitude is a function of what did not get built rather than what got ordered. The capex line makes that explicit. DRAM capex rises 68% in 2026 and NAND capex 57%, while DRAM wafer capacity rises 12% and NAND capacity rises 5%. Nearly a 6:1 ratio between capital deployed and capacity added in DRAM, and better than 11:1 in NAND. That gap is HBM conversion, equipment inflation, and cleanroom lead times absorbing the money before it becomes a wafer. Investors reading the capex headlines as a supply response are reading the wrong line item; capa growth of 12% against bit demand growth of 25% is not a supply response, it is a shortfall being financed.
The spot table published alongside the forecasts contains the more interesting anomaly. As of early August, 16Gb DDR5 traded at $51.3 while 16Gb DDR4 — the older, denser-per-dollar, supposedly end-of-life node — traded at $86.7. Legacy sits at a 69% premium to the current node, and it got there via 911% year-over-year appreciation against DDR5’s 733%. The 8Gb DDR4 part is at $42.1, up 744%. Anyone still describing this as an AI-driven DRAM cycle has to account for the fact that the parts AI accelerators do not use are the parts appreciating fastest. What is actually happening is that the majors converted mature capacity toward HBM and DDR5, and the industrial, automotive, networking, and set-top buyers who need DDR4 are now bidding against each other for a pool nobody is refilling.
NAND repeated the pattern in miniature this week. The 256Gb wafer, the oldest node on the sheet, rose 16% week-over-week; 1Tb rose 6% and 512Gb rose 5%. Look at the quarterly column and 512Gb is still down 2% QoQ while 256Gb is up 20%. BofA attributes the move to rush orders from Tier 2 and white-box OEMs building for September and fourth-quarter model launches, which is a demand explanation for a supply-shaped price curve. The corroborating tape-level datapoint is Nanya Technology, a pure legacy DRAM house with no HBM exposure, reporting July sales of NT$44bn, up 49% month-over-month and 720% year-over-year. Korean semiconductor exports of $42bn were down 9% MoM but up 179% YoY, and BofA’s proprietary memory indicator sits at 183 against mid-cycle and up-cycle reference levels of 100 and 130 respectively.
The forward path is the part that will get argued over. BofA models a final round of fourth-quarter contract hikes — DRAM up 8% QoQ, NAND up 3% — followed by stable pricing or mild corrections across every quarter of 2027, then declines of 8% in DRAM and 14% in NAND in 2028. That is a plateau, not a rollover, and the arithmetic holds because 2028 DRAM sales of $917bn still exceed 2027’s $847bn as bit growth of roughly 19% absorbs the ASP decline. The house view is a high-margin normalization rather than the cliff that has ended every prior memory cycle. Kioxia’s guidance for a 35% QoQ revenue increase in the third quarter, implying an ASP hike near 20% against an industry average in the low-to-mid teens, is the supporting evidence on the NAND side. Hynix’s own second quarter was the outlier in the other direction, with DRAM ASP up only 30% QoQ against 45-70% at peers, before a modelled 25% third-quarter recovery and more meaningful HBM4 shipments.
Capital allocation tells you what the producers themselves believe. Samsung’s CFO committed on the 30 July call to returning 50% of 2024-26 free cash flow net of the W10tn annual dividend, and BofA reads that as a W30tn-plus special dividend in the third or fourth quarter, W40tn-plus of buybacks in the first half of 2027, and W30tn of year-end dividends payable April 2027, with a further W30tn-plus of treasury shares needed for employee bonuses. SK Hynix is expected to return 50% of FCF weighted toward repurchases, roughly W40tn-plus against W20tn-plus in cash dividends. Two companies sitting on the tightest pricing environment in the industry’s history are choosing to hand the cash back rather than build the fabs that would end it. That decision is the supply thesis, stated by the only parties in a position to falsify it.
The equipment read-through is not uniformly positive. BofA cut its price objective on Wonik IPS to W85,000 from W95,000 on a 10% reduction to 2027 EPS after weak second-quarter CVD sales against US and Japanese competitors, reiterating Underperform, while keeping Duksan Neolux at W70,000 with a Buy on OLED materials and turbomachinery. A memory upcycle of this magnitude does not lift every Korean supplier; share loss inside a growing tool market still shows up as an earnings miss.
The number to watch is the DDR4-to-DDR5 spread. If legacy stays at a premium into the fourth-quarter contract round, this is a supply-exit cycle with a long tail and the 2028 correction estimate is too early. If it compresses while DDR5 holds, the AI demand story is doing the work after all.
Source: BofA Global Research, “Global Memory Tech — Weekly theme: Memory industry forecasts, NAND spot rally, Samsung’s dividends,” 7 August 2026 (Simon Woo, Dai Shen, Vivek Arya, Mikio Hirakawa, Matt Shin). Spot price data per DRAMeXchange.