AI’s ‘Continual Learning’ Redraws Memory Chip Map

As of September 19, 2026, the rise of 'continual learning' — AI that keeps absorbing new information even while it's serving users — is pushing memory chip demand beyond HBM into server DRAM and NAND. A global investment bank expects related demand to jump sharply next year, and Samsung Electronics and SK hynix are already rolling out competing new technologies to meet it.

1.6xNext year's HBM bit demand vs. this year (Citi forecast)

2.8x2028 HBM demand vs. this year (Citi forecast)

6.42%↑SK hynix share price gain on Sept. 18

42%↑Y/Y growth in AI-related goods trade, Q1 2026 (WTO)

Key points

'Continual learning,' where AI models keep training even as they answer queries, has emerged as a wildcard reshaping memory demand.
Citigroup expects HBM bit demand to run about 1.6x this year's level next year, and nearly 2.8x by 2028.
Samsung showcased a stacked memory-on-accelerator design, while SK hynix unveiled in-memory computing tech, at a US industry event.
Korean chip megacaps surged after Nvidia's CEO spoke, and new thematic ETFs tied to the trend keep launching.

The AI industry's newest buzzword is "continual learning" — an approach that blurs the once-clear line between training and inference. Traditional generative AI learns everything upfront, then just recalls it when answering. Continual-learning models keep absorbing new information even while actively serving users — industry insiders compare it to a student cramming new material mid-exam.

That shift is rewriting memory chip demand. Citigroup expects the trend, once it takes hold, to push high-bandwidth memory (HBM) demand well beyond today's levels: bit demand next year could run roughly 1.6 times this year's, nearly tripling to 2.8x by 2028. Citi argues that even amid memory-spec adjustments and lingering AI-safety concerns, this shift is the key reason supply shortages could persist for years.

The catch is that no single memory type can handle it alone. Models need to both ingest fresh data and repeatedly tap into vast stores of existing knowledge — and stashing everything in pricey HBM isn't realistic. That's expanding the role of server DRAM and NAND-based storage, and raising the stakes for tech that computes directly inside memory chips.

Samsung Electronics and SK hynix both debuted new technology to address this at a US AI infrastructure event. Samsung showed off "zHBM," a stacked design that places memory directly atop accelerator chips to shrink data-travel distance, targeting roughly 10x faster response times. SK hynix showcased PIM (processing-in-memory) tech, along with SALT-KV — which sorts data into different storage tiers by type — and a high-bandwidth flash architecture concept.

Markets are already pricing this in. Shares of both companies jumped after Nvidia's CEO said next year's chip sales could double this year's. Meanwhile, Korean markets have seen a wave of new ETFs tracking US memory makers and HBM-linked stocks. WTO data shows AI-related goods trade climbing 42% year-over-year in Q1 — far outpacing other categories — adding fuel to the bullish demand outlook.

In Korea right nowKorean investors are riveted by the chip-stock rally and the ETF gold rush it's spawned — several new funds tracking US memory makers or domestic HBM plays launched just this week. Brokerages argue AI infrastructure spending will keep climbing even with high interest rates, though some analysts warn the market's now flooded with near-identical thematic products that don't really differentiate themselves.

FAQ

What is 'continual learning'?

It's an approach where AI keeps learning from new data even while actively serving users, blending it with existing knowledge — effectively erasing the old line between training and inference.

Why does this matter for Samsung and SK hynix?

Because AI memory demand could broaden well beyond the HBM market the two companies already dominate, extending into server DRAM and NAND-based storage.

Could an 'AI slowdown' narrative derail memory demand?

Some have floated the idea of AI spending cooling off, but there's no evidence of actual order cutbacks yet, and analysts say it would be hard for the whole industry to slow down in lockstep.