Institutions: Memory Cycle Peak Call May Be a Year Too Early
Nashnova编辑部
Memory stocks sold off on a triple-negative last Friday yet fully rebounded within a week. JPMorgan and Jefferies both see a meaningful pricing inflection only in 2028 — the market may be front-running the downturn by at least a year.
How much is memory worth inside an AI rack?
Morgan Stanley calculates that memory content value jumps from roughly $374,000 per GB300 rack to over $2 million per VR200 rack — a 435% increase.
Total rack cost rises less than 100% over the same step; memory's share of rack cost climbs from about 9% to 26%. This means → an ever-larger slice of AI hardware spending is going straight to memory.
JPMorgan corroborates from the demand side: memory's share of cloud-capex rises from under 10% pre-AI to 31% in 2026 and 49% in 2027.
In plain terms = AI capex expansion is becoming intensely memory-dense. Recent "less memory" reports reflect architectural efficiency gains, not a collapse in total demand.
Peak margins — does that mean peak earnings?
The market's core worry is that margins have topped. JPMorgan disagrees: it expects memory operating margins to hold above 70% through 2028, not crater.
Margin × volume = earnings. If margins plateau at historic highs while shipments keep rising, earnings can still compound.
This means → "peak margin" and "peak earnings" are two different calls — watching margin direction alone and calling the top ignores the volume leg.
How long can prices keep rising? Can supply keep up?
Jefferies expects DRAM ASP to rise 40%-50% QoQ in Q3 and 30%-40% in Q4. JPMorgan forecasts 2027 DRAM ASP up 29% YoY, then another 7% in 2028.
Supply cannot respond quickly: JPMorgan estimates the market needs roughly 300,000 additional DRAM wafer starts to reach balance by 2028.
Fab construction takes two to two-and-a-half years, and converting lines to HBM (high-bandwidth memory — stacking DRAM dies for AI chips) adds extra wafer-yield penalties. This means → even if expansion starts today, meaningful new capacity only arrives around 2028.
Memory companies are paying big dividends now?
JPMorgan expects Samsung Electronics and SK Hynix to deliver a combined 16%-20% shareholder return over the next two years via special dividends and buybacks.
This reflects a historically rare feature of this cycle: the memory industry — traditionally known for boom-bust swings — is returning substantial cash to shareholders.
In plain terms = investors waiting for the cycle to play out now collect real cash returns along the way — a cushion previous memory cycles never offered.
If fab valuations are priced in, who catches the next ball?
If memory shipments keep climbing, the beneficiary chain extends beyond fabs — production, test, and packaging equipment, plus interconnect infrastructure such as CXL (a new interface standard for high-speed links between CPUs and memory), enterprise SSDs, and high-bandwidth flash all absorb rising demand.
This means → if fab valuations already reflect expectations, the next leg of upside may come from these "picks-and-shovels" suppliers.
What is the market really front-running?
The key open question: when will memory stocks begin to price in a 2028 downturn? Historically, share prices lead fundamental inflection points.
But if JPMorgan and Jefferies are right — a meaningful pricing inflection is a 2028 story — the market's current downturn pricing may be at least a year too early.
Content is for reference only, not financial advice.