Morgan Stanley: Memory Pullback Is a Healthy Reset, AI Will Push Cycle Peak Out by at Least Several Quarters
Claire Weston
Memory stocks surged 70%–134% in two months, then pulled back ~15%. Morgan Stanley analyst Shawn Kim calls it a crowded-trade cooldown, not a cycle end — AI buildout is still bottlenecked on DRAM, earnings revisions keep rising, and the cycle peak may be pushed out several more quarters.
Down 15% — is the cycle over?
Memory stocks rose 70% to 134% over two months, then pulled back roughly 15%, sparking fears the cycle has peaked.
Morgan Stanley analyst Shawn Kim's June 10 note is explicit: this is a crowded trade forced to cool off, not a cycle ending.
This means → too much capital piled into memory at once; short-term profit-taking created selling pressure, but fundamentals haven't deteriorated.
Since generative AI kicked off, memory stocks have gone through three similar resets — drops of roughly 24%, 10%, and 22% — and none of them ended the bull run.
How did AI turn memory from a "consumer-electronics sideshow" into a compute necessity?
Morgan Stanley's core thesis: memory is no longer just riding the smartphone and PC cycle. AI inference, AI agents, and data-center infrastructure are turning DRAM, HBM, and NAND into part of "token capacity."
In plain terms = memory used to rise and fall with phone shipments; now it rises and falls with how many tokens AI models process — storage has become AI's fuel.
Each GPU generation is bottlenecked more on memory than on compute: A100 carried 80 GB per card; Rubin GPU/Superchip calls for 288 GB to 768 GB — a 4× to 7× jump in DRAM content. AI chip annual growth is running at roughly 60%.
This reflects a widening gap: compute is accelerating, but memory can't keep up — whoever clears the DRAM bottleneck first runs more AI workloads.
Prices are still climbing — who is hoarding, who is cutting orders?
Channel checks show AI customers are still scrambling for supply; consumer-electronics buyers are starting to buckle under price hikes, with order cuts emerging.
AI-related products' share of supply allocation is expected to rise from over 50% this year to nearly 70% next year — consumer electronics and module makers will keep getting squeezed.
DRAM prices have nearly doubled since February; 3Q26 opening quotes still carry a roughly 20% quarter-on-quarter hike. NAND contract prices for 3Q26 may rise another ~30%, driven by enterprise SSDs.
HBM — high-bandwidth memory, the high-speed storage paired with AI chips — is still being negotiated; 2027 prices could rise 50% to 100% year-on-year. 2026 HBM pricing is already locked via annual contracts.
Why are long-term supply agreements the key to a valuation re-rating?
Morgan Stanley argues that LTAs — long-term agreements locking in supply volumes and pricing 3 to 5 years out — are the real lever for re-rating the cycle, not price hikes alone.
This means → if LTAs cover over 70% of total supply, DRAM stocks' P/E could theoretically re-rate from the current ~5× to 8–10× — contract-locked profits look more like predictable long-term cash flow, which the market prices at a higher multiple.
So far, the market hasn't assigned much premium for LTA-backed earnings: earnings revisions have roughly matched the stock rally, and 2027 P/E still sits around 5.2×.
In plain terms = the stocks have surged, but earnings power has risen almost as fast — so on a valuation basis, they haven't obviously gotten "expensive."
The traditional cycle says we should be peaking now — why does Morgan Stanley say it extends?
A traditional DRAM upcycle lasts roughly 6 quarters; stocks typically front-run the peak by about 4 months, placing the high point near year-end.
But Morgan Stanley's call: after AI-agent demand began ramping in January 2026, the cycle peak may be pushed out by at least several more quarters.
The supply side helps too — new fab capacity won't start coming online until late 2027, and constraints on cleanrooms, EUV equipment, plus LTA-driven demand visibility mean producers may not expand as aggressively as in past cycles.
This reflects the biggest difference from prior cycles: AI stretches demand on one end, while physical bottlenecks and contract commitments restrain supply on the other — both sides are delaying the inflection point.
What risks could actually break this cycle?
Morgan Stanley lists four hard risks: ① An architectural efficiency breakthrough in AI that sharply reduces memory usage; ② AI inference and agent growth diverging from the current exponential trajectory.
③ New capacity flooding the market in late 2027 just as AI buildout slows — supply surge meeting demand deceleration would replay the historical glut.
④ Macro liquidity tightening that compresses valuations on high-beta tech assets.
In plain terms = the cycle isn't over, but it has entered a phase that demands higher evidence density — whether AI demand is durable enough, whether LTAs can lock in cycle profits, and whether supply won't turn a shortage into a glut too quickly are the three verification points that will decide what comes next.
Content is for reference only, not financial advice.