Morgan Stanley: CPU Demand as a New Engine for the Storage Supercycle
0xBroomberg
J.P. Morgan raised its 2028 global memory TAM to $1.7 trillion, arguing the market has badly underestimated CPU-driven demand — not just GPUs — as the core variable.
Why is CPU demand the biggest expectation gap in this cycle?
The market has focused on GPUs pulling memory demand, overlooking that CPUs now play a rapidly growing role in AI systems. This means → the entire memory demand base was underestimated.
Nvidia guided $20 billion in standalone Vera CPU sales for 2026. AMD raised its server CPU TAM growth forecast from 18% to 35%+. Intel predicted the CPU-to-GPU ratio will eventually reach 1:1.
In plain terms = the old model was "GPU count drives memory needs." Now CPUs are consuming large volumes of memory too, adding a block of demand the models missed.
J.P. Morgan accordingly raised 2027–2028 server-grade DDR/LPDDR5 bit demand by 20%–22%, with AI CPUs' share of DRAM demand revised from 14% to 24%.
How extreme is a $1.7 trillion memory market?
Global memory revenue is projected to grow from roughly $214 billion in 2025 to $1.7 trillion in 2028 — roughly 8× in three years.
Operating margins for Samsung, SK Hynix, and Micron are expected to surge from 36% in 2025 to 75%–76%, far exceeding AI chips at 63% and cloud providers at 37%–39%.
This means → memory makers would become the highest-margin segment across the entire AI supply chain, entering historically uncharted territory.
At a 3.6× price-to-sales ratio, memory stocks still imply roughly 50% upside to 2027 TAM; on an EV/operating-profit basis, the implied upside reaches 89%.
Why is the HBM supply-demand gap still worsening?
Google TPU, Amazon Trainium, and other ASIC chips are adopting HBM3E heavily. ASICs' share of HBM bit demand is set to rise from 33% in 2026 to 39% in 2027, squeezing available supply further.
J.P. Morgan reversed its prior call for modest HBM3E price declines, now forecasting 2027 HBM3E ASP up 19% year-on-year and blended HBM ASP up 32%.
DRAM capacity allocated to HBM will rise from 24% in 2026 to 31% in 2028. This reflects HBM "crowding out" conventional DRAM capacity, providing additional price support for standard DRAM.
HBM bit demand carries a three-year CAGR of 85%, far outpacing the broader memory market.
Enterprise SSDs — why is this the overlooked NAND super-segment?
The enterprise SSD (eSSD) market — high-performance solid-state drives for data centers — will grow from over 500 EB in 2026 to more than 1,100 EB within two years, a 52% CAGR.
By value, eSSD TAM could exceed $300 billion over that period — larger than HBM. Put simply = everyone is talking about HBM, but eSSD has quietly grown into an even bigger market.
SLC-based (single-level cell) high-speed SSDs are gaining traction, with multiple solutions expected from 2H26. SLC yields fewer bits per wafer, which means → it tightens NAND supply-demand and eases downward price pressure.
How has memory's position within the AI ecosystem shifted?
Memory stocks' share of total AI-ecosystem market capitalization surged from 6% at end-4Q25 to 13% YTD in 2Q26 — the largest gain of any segment.
Memory's share of cloud providers' hardware capex by value rose from low-double-digit percentages early in the AI boom to an estimated 52% in 2026, potentially reaching 73% in 2027.
This reflects memory's shift from a supporting role to a strategic asset for cloud providers. After Nvidia disclosed a $3–4 trillion capex outlook through 2030, concerns about sustained order flow to memory makers are fading.
Can supply keep up — and what about Chinese competitors?
J.P. Morgan raised its three-year aggregate memory capex estimate from $300 billion to $450 billion. DRAM accounts for $364 billion, with 60% of incremental capacity allocated to HBM; EUV equipment procurement is the primary bottleneck.
Despite supply expansion, DRAM and NAND shortages are expected to worsen in 2027, on top of 220%–250% year-on-year price increases already projected for this year. Rising long-term agreement (LTA) contract ratios will anchor pricing and prevent wild swings.
Chinese memory makers (CXMT, YMTC) will see DRAM volume share rise from 6% in 2025 to 8%–11% in 2028, but product-mix constraints limit value-share gains to roughly 10%. In plain terms = volume is growing, but the revenue share stays small — the moat in high-end products holds.
Key variables to watch: whether Chinese makers shift meaningfully toward HBM capacity (which would tighten conventional DRAM further), and updates to export-control policies such as the MATCH Act.
Content is for reference only, not financial advice.