The Deepest Memory Shortage on Record: Why Goldman Sachs Says the Crisis Runs Through 2028

While not part of the "Big Three" designation, the report also named Micron Technology (MU) and SanDisk (SNDK) as key beneficiaries of the same AI-driven memory squeeze .

The Numbers: A Widening Supply-Demand Gap

Goldman’s models show undersupply intensifying, not resolving:

• DRAM: Shortage of 5.0% in 2026, deepening to 5.9% in 2027 .
• NAND: Shortage of 4.4% in 2026, deepening to 4.6% in 2027 .
• HBM: Even tighter than conventional DRAM and NAND in both years, though exact percentages weren't broken out in the same format .

These figures represent a deepening and an extension of a prior forecast from February 2026 that had already flagged the largest DRAM supply gap in 15 years .

Why "This Time Is Different": Three Structural Breaks

Past memory booms, driven by PC and smartphone upgrade cycles, typically ended when supply caught up and consumer demand softened. Goldman argues this cycle is built on three structural shifts :

1. AI Demand Is Non-Deferrable: AI server buildout now consumes over half of all DRAM supply. This is not discretionary consumer spending that can be delayed; hyperscalers are locked into multi-year infrastructure buildouts.
2. Capital Expenditure Discipline: Historically, high prices triggered aggressive new fab construction. This time, major manufacturers are showing restraint, deliberately constraining supply growth.
3. Permanently Rising Memory Intensity: High-bandwidth memory and high-density NAND are no longer just components for gadgets; they are integrated into the fabric of AI infrastructure, enterprise storage, and data centers.

A Historic Shift in Valuation

Perhaps the most actionable insight for investors is Goldman’s call for a framework shift from price-to-book (P/B) to price-to-earnings (P/E) .

The bank argues that the profitability and earnings visibility of memory makers have structurally improved. Despite this, most of these stocks trade at mid-single-digit P/E ratios, a valuation that fails to capture a multi-year earnings cycle. Goldman formally anchored its new targets to P/E multiples, using 9x as a baseline .

The Known Risks: A Boom's Classic End

Goldman Sachs does not dismiss the historical lessons of prior memory cycles. The report acknowledges that:

• A slowdown in AI investment or a cut in hyperscaler capex could rapidly destroy demand .
• If manufacturers abandon their current capex discipline en masse, the shortage could flip into a classic oversupply glut .
• High component costs are already slowing growth in PCs and smartphones, a form of demand destruction that could accelerate .

While the bank sees the AI architecture as a strong buffer against these risks, it warns that a sudden surge in capital expenditure would bring the old cycle logic back with a vengeance .

The broader analyst community echoes the sentiment, with firms like IDC describing a "potentially permanent" strategic reallocation of silicon wafer capacity, though its timeline was more cautious, initially flagging risks into 2027 .