The Memory Wall: Morgan Stanley Confirms AI's Next Bottleneck and the $276 Billion Opportunity

The Core Thesis: From Compute Wall to Memory Wall

Morgan Stanley's analysts, led by Shawn Kim, argue that the industry has reached an inflection point. Their framing is direct: "GPUs determine how fast AI runs, while memory determines how far AI can go" . The structural root of this shift is a staggering divergence in growth rates:

• DDR5 memory bandwidth increased by only 14% over the past two years (2024–2026) .
• AI inference token demand surged more than 320-fold over the same period .

This imbalance means that as AI scales toward inference-heavy, agentic, and long-context workloads, memory — not compute — is now the binding constraint .

'Chipflation': The New Economics of Memory

Morgan Stanley coined the term "chipflation" to describe a market where memory chip prices rise sharply and stay elevated because demand persistently exceeds supply .

• Memory chip prices have risen six-fold over the past year . A summary of Morgan Stanley's June 2, 2026 report confirms "DRAM Price Growth: 6 fold (Past year)" .
• The bank notes that DRAM prices are back to their highest levels in nearly 30 years .
• Media coverage of the reports cites DRAM prices up 125% in 2026 .
• Memory now accounts for 73% of a CPU server's bill of materials (BOM) .

A two-tier market has emerged: hyperscalers secure supply via long-term agreements (LTAs), while non-AI buyers face allocation risk and even higher prices .

Reshaping AI Infrastructure Investment Through 2030

The bank's projections for AI infrastructure spending are enormous and point to a dramatic reallocation of capital toward memory:

• Cloud storage spending is projected to reach $418 billion by 2030 .
• Memory's share of cloud providers' capital expenditures will rise from 12% to 40% by 2027 .
• The top five hyperscale cloud providers could spend up to $1.4 trillion in total capex by 2028 .
• Morgan Stanley estimates nearly $2.9 trillion in global data center construction cost alone through 2028 .
• The bank identifies a $276 billion addressable market for novel memory technologies — innovations needed to overcome capacity, bandwidth, and cost constraints .

Six Paths to Break the Memory Wall

Morgan Stanley outlines six areas critical to overcoming the memory wall :

1. Advanced process nodes — shrinking memory cell geometry
2. Memory architecture redesign — fundamental changes to memory cell and array design
3. Advanced packaging — 3D stacking and heterogeneous integration (e.g., HBM)
4. CXL (Compute Express Link) — enabling memory pooling and disaggregation
5. Processing-in-memory (PIM) — moving compute into the memory array
6. New materials — novel dielectrics, conductors, and switching materials

Spillover into Consumer Electronics

AI's memory demand is crowding out supply for consumer hardware, creating tangible shortages :

• PCs: A 15% memory shortfall in 2027, equivalent to 58 million PCs affected .
• Smartphones: A 12% memory shortfall in 2027, equivalent to 134 million devices affected .

Sony and Lenovo have already raised prices, and Microsoft attributes roughly $25 billion of its $190 billion 2026 budget to elevated chip costs .

Market Size and Revenue Projections

Morgan Stanley sees the memory market nearly quadrupling in size:

• Memory market growing from about $220 billion in 2025 to about $890 billion by 2030 .
• Memory industry profits are projected to grow 35%–40% by 2027 .

Top Investment Picks

Morgan Stanley's recommended plays on the memory bottleneck include :

• Micron — memory manufacturer with DRAM and NAND exposure
• Samsung Electronics — dominant memory producer with HBM leadership
• SK Hynix — leading HBM supplier, a key beneficiary of the AI memory super-cycle
• SanDisk (Western Digital's spin-off) — NAND flash exposure
• ASML — EUV lithography equipment, identified as the next bottleneck after memory

Analyst Shawn Kim summarized the thesis: "memory is the new AI bottleneck." As agentic AI workloads scale, the investment opportunity is expanding beyond GPUs into CPUs and memory .