BofA Merrill Lynch: AI Demand Visibility Extends to 2028, Memory Won't Be Oversupplied
Alina Collins
Bank of America raised its 2030 global semiconductor market forecast from $2.3 trillion to $2.7 trillion, arguing AI-driven chip demand visibility now stretches to 2028 with no DRAM or NAND oversupply before then — a bullish call published the same day the Philadelphia Semiconductor Index fell 7.9%.
A $2.7 trillion forecast — where is the growth coming from?
BofA raised its 2030 global semiconductor market estimate from $2.3 trillion to $2.7 trillion, implying a 2025–2030 compound annual growth rate of 28%.
2026 is the breakout year: total semiconductor sales are forecast to grow 103% year-on-year, with memory chips surging 298% — DRAM up 309%, NAND up 295%.
This means → the chip industry took roughly 50 years to reach its first $1 trillion in annual sales. BofA believes AI can add another trillion within five years.
Why won't memory oversupply before 2028?
BofA projects DRAM and NAND supply-demand ratios will stay above 110% throughout the forecast period, with prices holding firm — no quarter-on-quarter decline expected before 2027.
In plain terms = money is being spent, but most capex right now is going into building cleanroom shells, not buying production equipment. Micron's FY2026 capex guidance tops $25 billion (up from $13.8 billion in FY2025), yet the bulk funds construction — equipment that actually produces chips won't arrive at scale until 2028.
Micron's new Idaho fab is expected to begin initial output in mid-2027 and ramp through 2028; its Singapore HBM advanced-packaging plant should start contributing in 2027 and reach full production in 2028. Geopolitics, packaging capacity, and power constraints are also throttling supply expansion.
Long-term agreements (LTAs) provide 2–3 years of supply-and-pricing visibility; the Micron–Anthropic deal is a recent example. This reflects downstream customers locking in future capacity with long contracts, further tightening available supply elasticity.
HBM from $35 billion to $246 billion — what's driving this?
HBM — high-bandwidth memory, a high-speed memory standard purpose-built for AI accelerators — is the core driver of this memory supercycle. BofA forecasts HBM revenue will grow from roughly $35 billion in 2025 to about $246 billion by 2030, a 34% CAGR.
HBM capacity per AI accelerator is expected to rise from 187 GB in 2025 to 464 GB by 2030. Nvidia's upcoming Vera Rubin system (expected H2 2026) packs 288 GB of HBM4 per accelerator, confirming the trajectory.
This means → pricing is rising too: HBM is projected at roughly $17.50/GB in 2027–2028, up from $14.30/GB in 2026 — volume and price moving up together gives memory makers wider profit leverage than headline revenue alone suggests.
Why will equipment spending spike in 2028?
BofA raised its 2028 wafer-fab equipment (WFE) spending forecast by 23% to $250 billion (up 32% year-on-year), and lifted 2027 to $190 billion (up 31% YoY).
Three forces converge: ① cleanroom capacity built over 2025–2027 reaches completion around 2028, triggering concentrated equipment procurement; ② 2nm GAA — gate-all-around, a transistor architecture that wraps the gate on all four sides instead of three — accelerates into volume production in 2026–2028, requiring more equipment per wafer, with High-NA EUV lithography (the next-generation, higher-precision scanner) adding further spend; ③ HBM upgrades from HBM3 to HBM4/5 and NAND migration from 300 to 400+ layers both demand additional tooling.
In plain terms = WFE as a share of semiconductor revenue will drop from the historical mid-teens to roughly 11% — but that is not shrinking demand. It is memory-chip price inflation inflating the denominator. The more meaningful metric is equipment spend per wafer, which keeps rising.
Big target-price hikes — but who gets left out?
BofA raised Micron's price target from $950 to $1,500 (maintaining Buy) and Applied Materials' target from $540 to $720, shifting the valuation base year from 2027 to 2028.
This means → BofA is betting that 2028 will be the dual inflection point for equipment spending and memory capacity — if the timeline holds, current valuations still have room.
But consumer electronics is explicitly excluded from the AI dividend: smartphone chips are forecast to decline 13% YoY in 2026, PC chips to fall 9% — a stark contrast to the AI-datacenter boom.
The report dropped on a day the SOX fell 7.9% — how to read the gap?
This bullish report — arguing AI demand visibility extends to 2028 — was published the same day the Philadelphia Semiconductor Index fell 7.9%.
This reflects a sharp divergence between short-term market pricing and medium-term fundamental research — short-term capital was de-risking while sell-side analysts were raising forward estimates on the same day.
In plain terms = whether 2028 actually becomes the dual inflection point for equipment spending and memory capacity is the key validation checkpoint for this optimistic thesis. The price the market is paying today and the target price the analysts are setting are, for now, telling very different stories.
Content is for reference only, not financial advice.