AINEWS
JPMorgan's Jay Kwon calls a two year memory chip shortage, with demand broadening from GPUs to CPUs. The 25% Q3 sector drop is an EPS reset, not a demand break.
Memory chips, the components that store data on every AI server, are now the binding constraint on AI buildout, and the shortage has years to run, according to JPMorgan strategist Jay Kwon's Monday research note. The call sits inside an unresolved contradiction: the supply squeeze is tightening even as memory stocks have just corrected 25% in the third quarter.
Demand is broadening from GPUs, the processors that handle AI's matrix math, to the CPUs that orchestrate servers and feed GPUs data. That shift is what investors have underestimated, Kwon writes, and it underpins both his raised memory market size estimate and his bullish call on Korean names plus Micron, Kioxia, and Winbond. Micron, the only major US-listed pure-play memory company, told attendees at JPMorgan's TMT conference that the crunch will outlast 2026 because AI demand is outpacing what high-bandwidth memory (HBM, the stacked DRAM chips that sit next to AI GPUs), conventional DRAM, and NAND flash can supply.
Memory stocks dropped roughly 25% in the third quarter, and Micron, which is up 202% year-to-date against 13% for the S&P 500, has shed 10% over the past month on overvaluation fears, per Yahoo Finance AlphaSpace. Kwon's own note names the drivers: EPS misses against high expectations, slower-than-expected cloud-service-provider AI capex, and a phenomenon he calls "memory content optimization," meaning cloud customers tuning their servers to use less memory per CPU, or less memory per GPU, to save cost.
The Sandisk corroboration sharpens the picture. On the storage company's prior earnings call, CEO David Goeckeler said Sandisk has "over four years of demand visibility" from its largest customers, a sign that hyperscaler demand is committed well past the typical 12-to-18-month contract window, per the Yahoo Finance report. That does not contradict Kwon; it lengthens the planning horizon. Sandisk sells NAND flash, the cheap, persistent storage used in solid-state drives, which sits on a different node and a different supply curve from HBM, but the visibility framing is the same: customers are signing multi-year supply commitments because they cannot afford to be short.
Deloitte's independent analysis goes further than Kwon does. Its team argues the memory crunch is greater than expected and may not ease until 2029, two years past Kwon's call. The two-year-versus-five-year gap is the duration tension to watch. If Kwon's call is right, the 25% Q3 correction is a routine valuation reset and 2027 is the year of tightness. If Deloitte is right, the correction is a pause inside a longer squeeze and the 2027 loosening is itself a soft patch.
What would break Kwon's two-year call is concrete. First, memory content optimization spreading from a few customers to most hyperscalers: the goal would be to compress content per server from, say, 1.5TB toward 0.8TB without performance loss, and that would offset unit volume. Second, cloud capex slowing structurally: the major CSPs have guided to record 2026 capex, but if those guides get cut by 15% or more in late 2026 or early 2027, the demand picture softens. Third, the supply response itself: HBM and DRAM capacity additions at Micron, SK Hynix, and Samsung, plus Chinese entrants, are landing in 2026 and 2027, and faster-than-expected yields would close the gap earlier than Kwon models.
The risk view from Everstream Analytics is more pessimistic, treating the shortage as a worsening macro condition rather than a two-year price story. That framing matters for downstream device and server buyers: if memory inflation continues through 2027, AI server unit costs stay elevated, consumer PC and phone pricing absorbs the same dynamic in smaller form, and any product launch whose bill of materials is memory-heavy is exposed.
For now, the visible signal is that supply is locked and visibility is longer than the tape suggests. Kwon's third-quarter reset is an earnings and valuation story, not a demand break. The next milestones are the cloud capex guides in late October and Micron's November quarter, when memory content per server will be a number investors can finally compare across cycles.