JPMorgan: Memory Stock Pullback Is an Expectation Reset, Not a Cycle Reversal
Claire Weston
Asian memory chips have dropped roughly 30% from their June peak. JPMorgan argues the core driver is not fading demand but a market shift from the "AI expansion" narrative to "earnings delivery verification" — hyperscaler capex and HBM pricing are now the deciding variables.
Down 30% — what exactly is being repriced?
Memory stocks have fallen about 30% from the June high, while the Philadelphia Semiconductor Index dropped only about 11% — nearly a three-to-one gap.
This means → the market is not dumping semiconductors broadly; it is specifically repricing the premium that memory stocks built up fastest.
JPMorgan analyst Jay Kwon wrote on July 14 that the core narrative has shifted from "how big can AI infrastructure get" to "can the earnings already priced in actually be delivered."
Three pressures stacked up — what broke the mood?
Pressure one: expectations outran reality. Markets kept raising AI data-center investment forecasts and lifting the memory sector's total addressable market in lockstep. After share prices surged, investors began questioning whether sell-side demand estimates had raced ahead of hyperscalers' actual spending plans.
Pressure two: DRAM price gains slowed. After rounds of price hikes, year-on-year and quarter-on-quarter DRAM increases started to moderate after Q2 2026, cooling expectations for rapid earnings expansion.
Pressure three: Samsung earnings were cut early. Before Q2 results were released, the market had already trimmed Samsung's operating-profit forecasts. In plain terms = the sector leader hadn't even handed in the exam, but its score was already marked down — confidence took a hit.
HBM pricing — where is the biggest disagreement?
Most buy-side firms expect HBM — high-bandwidth memory, a high-speed DRAM designed specifically for AI chips — to roughly double in per-GB price by 2027, driving further earnings upgrades.
JPMorgan is notably more cautious: it estimates the current HBM industry average at about $1.80 per GB, even slightly below some high-end server DRAM products. This means → HBM's "premium" is not as wide as many assume.
The bank sees a 25%–30% year-on-year HBM price increase in 2027 as more realistic. Memory makers negotiate with cloud firms on a bundled DRAM-plus-NAND-plus-HBM profitability basis, so HBM pricing cannot rise without limit.
However, HBM contracts are repriced roughly every year, far shorter than the three-to-five-year deals typical for conventional DRAM. In plain terms = if AI demand keeps beating expectations, chipmakers still hold room for further price increases.
Long-term agreements — shackle or moat?
LTAs — long-term supply agreements between memory makers and large customers — were the most debated topic during the roadshow. Compared with a few months ago, investor sentiment has improved markedly; the discussion has moved from "do LTAs exist" to "how can makers use LTAs to lock in key AI clients."
JPMorgan expects more than half of future contract volumes will fall under the LTA framework. This means → the industry is shifting from quarterly price negotiations toward volume-for-stability deals.
The bank argues LTAs do not cap upside: new incremental orders can still be signed at higher prices, take-or-pay clauses provide order certainty, and products outside the LTA scope continue to rise with tightening supply-demand.
How long can the tight supply-demand picture last?
DRAM supply currently meets only about 50%–60% of order demand; NAND is at roughly 70%–80%. This reflects the fact that even after a sharp share-price drop, underlying supply-demand has not loosened.
JPMorgan expects the tight balance to persist through 2027–2028, even as DRAM wafer capacity keeps expanding.
Enterprise SSDs are emerging as a new growth engine. AI data centers are driving persistent upward revisions to enterprise SSD demand — especially from workloads like KV Cache Offload, a technique that lets AI models temporarily store intermediate computation data on SSDs. The supply chain forecasts enterprise SSD shipments approaching 500 EB in 2027, up nearly 50% year-on-year.
What should investors watch next?
After meeting more than 50 Hong Kong-based institutional investors, JPMorgan found roughly 70% of sentiment hinges on one variable: whether hyperscaler capex will continue to be revised sharply upward. Most investors expect global hyperscaler capex to climb further to $1 trillion–$1.5 trillion over the next three to six months.
This means → the sector's trajectory no longer depends solely on "is AI demand strong" but on two verifiable numbers: whether hyperscaler capex keeps delivering, and whether HBM profitability can reach the bullish assumptions already baked into share prices.
The verification window centers on the upcoming quarterly earnings cycle. In plain terms = this correction is about waiting for the report card, not about the industry turning cold.
Content is for reference only, not financial advice.