Citi says AI memory boom has further to run despite 20%+ sell-off in chipmakers

[Elliot_30]

Citi has put out a genuinely bullish note arguing the AI driven memory chip boom still has real room to run, even after Micron, Samsung and SK Hynix have all pulled back more than 20 percent from recent peaks as investors started questioning valuations and the durability of AI infrastructure spending

The banks core argument is that this cycle is structurally different from past memory upcycles, Citi wrote that the current expansion is likely to outperform the 01 to 07 upcycle given that AI demand is driving both DRAM and NAND demand simultaneously, rather than just one segment of the memory market getting pulled along, that dual demand driver is what Citi thinks gives this cycle genuinely more durable legs than previous boom and bust memory cycles investors have lived through before

A big part of the bullish case rests on how customers are actually buying now, unlike previous cycles where purchasing tended to be more spot market and reactive, customers are increasingly locking in supply through long term agreements running three to five years, Citi frames these contracts as a real signal of confidence that elevated demand will persist well beyond the near term, and they also give memory manufacturers themselves much better earnings visibility than theyve historically had, reducing the boom and bust volatility that used to define this sector

On the specific question of high bandwidth memory, the type of memory most directly tied to AI chip performance, Citi acknowledges an interesting wrinkle, ongoing HBM shortages are actually pushing AI companies to redesign their system architectures around more GPUs paired with less HBM per individual chip, but the bank still expects total HBM demand to keep accelerating regardless, forecasting HBM capacity per AI system will grow from 20.7 terabytes to 110.6 terabytes, a 434 percent increase, as GPU counts per system expand from 72 to 576, meaning even with less HBM per chip, the sheer scale of AI systems keeps pushing total demand sharply upward

Citi named SK Hynix as its preferred pick in the sector specifically, expecting the company to announce additional shareholder returns after management recently signaled its capital allocation options are under active review, the bank argued that as mid to long term earnings visibility becomes clearer, supported by the ongoing AI memory upcycle and the substantial advance payments secured through long term agreements, Hynix should be able to offer both a constructive market outlook and decent shareholder returns, Citi maintained its Buy rating, raised its 2026 and 2027 operating profit forecasts, and kept its target price at 3.1 million won

Citi Says AI Memory Boom Still Has Further to Run Despite Recent Sell-Offuk.finance.yahoo.com

[Anvil79]

The distinction between spot market purchasing and locked in three to five year agreements is genuinely the most important structural difference from past cycles, that kind of contracted demand visibility is exactly what tends to smooth out the boom and bust volatility this sector has always been notorious for

[RusticRidge]

A 20 percent plus pullback across Micron, Samsung and Hynix simultaneously being framed as an attractive entry point rather than a warning sign is a classic contrarian bank call, curious whether Citi is right that this is just a healthy correction or whether the market pullback is pricing in something the bank is underweighting

[Sega26]

434 percent projected growth in HBM capacity per AI system is a genuinely staggering number even accounting for how much hype surrounds AI infrastructure right now, if that materializes anywhere close to forecast it would represent a real structural shift in the memory market

[Hare51]

The architecture shift toward more GPUs with less HBM per chip is an interesting technical detail that could have gone either way for total demand, glad Citi actually addressed that directly rather than ignoring it, and their conclusion that total demand still rises despite less HBM per chip makes sense given how much GPU counts are scaling

[Steve59]

SK Hynix being singled out as the preferred pick with expected additional shareholder returns feels like a fairly conviction heavy call, would want to see the actual advance payment figures from those long term agreements to judge how solid that earnings visibility argument really is

[Christopher_27]

This cycle outperforming 01 to 07 is a bold claim given how much bigger and more mature the memory industry is now compared to two decades ago, though the dual DRAM and NAND demand driver from AI specifically is a new dynamic that older cycle didnt have

[Hyperdrive71]

If this note proves right and the sector genuinely does have further to run, it validates the broader thesis that AI infrastructure spending isnt just a speculative bubble but reflects real structural demand that memory manufacturers can build multi year business plans around

[Dolphin43]

Would be curious to see Citis downside scenario laid out just as clearly as this bullish case, every AI infrastructure bull case right now needs a credible answer for what happens if capex growth actually slows meaningfully next year, and this note doesnt seem to address that directly

[Bear24]

Its worth remembering banks issuing bullish notes right after a sector sell off is a pretty standard playbook move, doesnt make the analysis wrong but its worth weighing Citis own incentives and existing positioning alongside the actual argument here

[Restless Barrel]

The 20%+ sell-off does not automatically mean the underlying memory story is broken. If anything, it may be the market trying to separate long-term demand from the near-term expectations that had become pretty extreme. HBM is still a specialised product, and the interesting question is whether supply can catch up quickly enough to crush margins. The GPU architecture point is worth watching too. More GPUs with less HBM each could flatten demand per accelerator while still leaving total demand growing because the number of accelerators keeps rising. That is the classic case where unit economics and unit volumes pull in opposite directions.

[Frost Gary]

The long term agreement trend mirrors what were seeing across other parts of the AI supply chain too, Sandisk and other storage and memory suppliers have talked about similar multi year contracts, this genuinely does seem to be becoming the new normal purchasing pattern industry wide

[Natalie99]

Citi's argument makes sense to me, but the part I'd be careful with is treating AI memory demand as a straight line. Semiconductor cycles have a habit of turning a very real shortage into a very real oversupply once everyone builds capacity at the same time. A practical example would be a cloud provider doubling its accelerator fleet but becoming more efficient with memory per chip. That can still be great for total HBM consumption, just not as explosive as the earlier forecasts implied. The suppliers that manage yields, capacity and pricing best probably matter more than simply betting on the biggest headline growth number. So the sell-off could be either a buying opportunity or the market finally pricing in a normalisation. We probably need a few quarters of actual shipments and margins before knowing which story we're in

[Shane_8]

The architecture shift is exactly why I would avoid using the number of GPUs as a simple proxy for memory demand. There are several moving parts: HBM capacity per package, the memory bandwidth required by each workload, accelerator utilisation, and how quickly inference workloads scale. There is also a difference between demand for memory and profitable demand for memory. If every major supplier adds capacity because today's prices look fantastic, the eventual increase in supply can change the economics even while AI infrastructure keeps expanding. That happened repeatedly in older memory cycles. The bullish case still has plenty going for it, though. AI systems are becoming more memory-hungry in ways that are not captured by simply counting chips. Large models, long context windows and increasingly complex inference can all push bandwidth requirements higher. The interesting debate is less about whether AI will need memory and more about how much of that demand will translate into sustainable pricing power.