NextFin News - South Korean chipmakers Samsung Electronics and SK Hynix advanced in Seoul on Tuesday, September 22, 2026, as Meta Platforms' newly launched Muse AI agent reignited investor enthusiasm for the memory chips that power artificial-intelligence infrastructure. The move extended a rebound that had already pushed the benchmark KOSPI back above 7,000 the prior session on record semiconductor exports, and it underscores a market betting the AI memory shortage has years to run - not months.
The timing is deliberate. Meta opens its Connect developer conference on September 23 at its Menlo Park campus, with the keynote scheduled for 4 p.m. Pacific time, where the company is expected to deepen its push into AI agents and smart glasses. Muse, released on September 8 as Meta's first task-executing personal agent, is the software hook; the memory chips made in South Korea are the physical bottleneck that determines how many of those agents can run at once. Investors are treating the two as a single trade.
The Situation: A Rally Built on Agents and Atoms
Muse is not another chatbot. According to Meta's launch materials, it is a personal AI agent that can access a user's email, calendar, payment methods, and other applications to carry out multi-step tasks with minimal supervision - booking travel, making purchases, managing communications. The company framed the ambition in sweeping terms, with CEO Mark Zuckerberg saying: "Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about. Your agent will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more."
The product is available only in the United States, restricted to users 18 and older, and reachable through a dedicated iOS and Android app, a web destination at muse.ai, or directly inside WhatsApp. A free tier allows up to 100 million tokens per week per user, while paid tiers - Power at $20 a month and Maximum at $100 a month - open higher usage ceilings. Meta has said glasses support is coming soon but is not yet available. The launch moved Meta's own shares more than 6% higher on the day of release.
Why does a software agent lift chip stocks in Seoul? Because every agentic session consumes memory - high-bandwidth memory (HBM) stacked next to AI processors, plus server DRAM and NAND flash for context windows that are expanding from hundreds of thousands to millions of tokens. The market's read is straightforward: more agents, running longer, for more users, means more memory per query and more queries per user. The memory makers sit at the narrowest point of that funnel.
The Seoul market action needs context. On Monday, September 21, the KOSPI closed at 7,007.72, up 1.65%, reclaiming 7,000 for the first time in seven sessions, according to Korea Exchange data. Samsung Electronics led the rally, closing at 274,000 won, up 4.98%, after rising as much as 5.36% intraday. SK Hynix finished at 1.87 million won, up 0.59%. The catalyst that day was trade data: semiconductor exports reached $34.12 billion from September 1 through 20, up 259.4% from a year earlier and setting a new monthly record, according to the Korea Customs Service. Tuesday's advance extended that momentum with a fresh software catalyst layered on top.
The combination matters. Export data confirms the demand is already showing up in the customs ledger; the Muse launch is what convinces investors the demand has not yet peaked. One is backward-looking proof, the other forward-looking optionality. Together they explain why the memory trade, which many declared exhausted after July's volatility, is back on the front foot.
Analysis: The Mechanism Behind the Move
From Agents to Memory: The Transmission Channel
The first-order link is simple arithmetic. An Nvidia H100-class accelerator carries 80 GB of HBM; the H200 carries 141 GB and the B200 carries 192 GB, according to configuration data tracked by Silicon Analysts. An independent researcher writing under the pseudonym Zeitgeist estimates that a single session with a 128,000-token context window requires roughly 20 GB of memory, that a knowledge worker running ten AI agents in parallel would need around 152 GB, and that the next wave of accelerators could carry as much as 288 GB. At one hundred agentic sessions per person per day, the estimate implies the world would need roughly 60 times more memory than will be produced in 2026.
That arithmetic is why memory has become the swing factor in the AI build-out. GPUs get the headlines, but memory is where the physics bites first: you cannot compensate for a memory shortfall by adding more logic. The transmission channel from "Meta ships an agent" to "Samsung and SK Hynix rise" runs through this bottleneck - each incremental agentic workload raises both the memory capacity required per session and the number of sessions per user, compounding demand on both axes.
The pricing power follows. Bank of America estimates the 2026 HBM market will reach $54.6 billion, a 58% increase from the prior year. Goldman Sachs forecasts HBM demand tied to custom, ASIC-based AI chips will surge 82%, accounting for one-third of the market. SK Hynix, which UBS channel checks place at roughly 50% of total HBM bits, is the most direct beneficiary; Samsung, the clear No. 2 in HBM supply, is the leveraged second. TrendForce data puts second-quarter 2026 DRAM industry revenue at $154.73 billion, up 59.5% from the previous quarter - the shortage is already priced into supplier revenues, not just expectations.
"The demand for AI is robust, and it doesn't matter if the Fed hikes or not," said Jung In Yun, chief executive officer at Fibonacci Asset Management Global in Singapore. "The earnings prospect is overwhelming the rate hike fear."
That quote captures the market's hierarchy of drivers. Rate sensitivity matters for valuation multiples, but earnings revisions matter more when supply cannot catch up with demand. When the supply curve is inelastic, the marginal buyer of memory is not price-sensitive - and neither is the stock.
Cyclical or Structural: The Call That Determines the Trade
This is where the analysis must take a stand, because getting it wrong flips the conclusion. Memory is historically the most cyclical industry in semiconductors - boom, bust, inventory correction, repeat. The July 2026 rout, in which the KOSPI shed roughly 18% over two sessions and Samsung and SK Hynix each fell double digits, was a textbook cyclical scare: investors priced a slowdown in AI infrastructure spending and an inventory hangover.
But the current leg is different in one critical respect: the shortage is not being driven by a temporary inventory swing. It is being driven by a step-change in workload architecture - from single-turn chatbots to persistent, parallel agents with expanding context windows. That is a structural shift in how compute is consumed, and it changes the shape of memory demand in a way that does not mean-revert on its own.
The evidence for the structural read: SK Hynix Chief Executive Kwak Noh-jung has said the global memory industry faces its worst-ever supply shortage in 2027, with demand continuing to exceed the company's production capacity well beyond 2030 despite aggressive expansion. KB Securities analysts project the memory shortage will persist through 2028. Daishin Securities reports that customers have requested more than 40 billion Gb of HBM for 2028 demand, with a single customer's request exceeding this year's estimated global HBM shipment of 36.6 billion Gb. When forward orders exceed an entire year's global output, the constraint is structural, not cyclical.
The cyclical counterweight is real and must be separated out. Roughly 55% to 70% of 2026 revenue growth at Samsung memory, SK Hynix, and Micron comes from price rather than volume, according to TS Lombard analysts - and price growth is the cyclical component that will mean-revert when capacity catches up. The structural component is the rise in memory content per AI workload, which persists regardless of the price cycle. The correct read is therefore a barbell: a cyclical price wave riding on top of a structural volume floor. Investors who treat the whole move as cyclical will exit too early; investors who treat all of it as structural will be caught by the price correction when new fabs come online.
Short-knife close: the agents are structural; the pricing is cyclical. The trade works as long as you know which leg you are holding.
Second-Order Thinking: What the Market Has Not Priced
The first-order trade - more agents, more HBM - is already conventional wisdom. The second-order effect is less discussed: agentic AI shifts memory demand from the data center outward, toward the edge. Meta has signaled that Muse will extend to its smart glasses line, and Connect 2026 is widely expected to showcase deeper integration between Meta's AI software and consumer devices. If agents migrate from cloud-only execution to on-device or near-device inference, the memory mix shifts again - toward higher-density mobile DRAM and NAND for local context caching. That would broaden the beneficiary set beyond the pure HBM leaders.
The third-order effect is an expectation gap between the hyperscalers and the memory suppliers. For much of 2026, software and cloud names underperformed the chipmakers underpinning the AI boom - a reversal of the 2023-2024 pattern where applications led. If agents monetize faster than the market expects, the multiple compression on software names could reverse, and the relative trade that has favored memory over platforms would unwind. The Muse launch is an early data point that the platform layer may be about to recapture the narrative.
There is also a geographic second-order effect. South Korea's benchmark has become a leveraged proxy on global AI capital expenditure: memory companies make up roughly half the KOSPI by weight and have been the dominant contributor to the index's 2026 gains. A rally in Seoul on a Meta product announcement is not parochial - it is the global AI trade expressing itself through the market with the purest memory exposure. That concentration cuts both ways, as July's nearly 11% single-day collapse demonstrated.
The Adversarial Case: Why This Rally Could Be a Trap
The strongest argument against chasing this move is that the memory cycle has fooled investors before, and the valuation cushion is thinner than it looks. Both Samsung and SK Hynix have traded at roughly three times forward 2027 earnings, according to Nomura Securities, which in September maintained Buy ratings with target prices of 670,000 won for Samsung and 4.7 million won for SK Hynix. Cheap on earnings, yes - but memory earnings are the most leveraged to the price cycle in the industry. If HBM pricing peaks in 2027 as new capacity from all three major suppliers comes online, those forward earnings estimates compress faster than the multiples expand.
There is also a concentration risk that is impossible to ignore. Samsung and SK Hynix together account for more than half of the KOSPI by market capitalization - the highest two-stock concentration on record for the Seoul main board. Goldman Sachs has flagged that MSCI Korea single-stock weights for both names now sit above the regulatory caps that govern many diversified U.S. funds, forcing mechanical trimming by global Korea allocators. A market where half the index is two memory names does not diversify; it amplifies. The July rout, which triggered a combined buy- and sell-side sidecar count that surpassed the previous annual record of 26 set during the 2008 global financial crisis, is the live demonstration of that fragility.
The counter-thesis, stated fairly: this rally is a cyclical reflex on a software announcement that has not yet produced a single dollar of incremental memory revenue, in a market that is structurally over-concentrated and mechanically prone to violent reversals. The earnings are real, but the multiple being assigned to the next wave of demand assumes the shortage persists unbroken for years.
Answering it requires separating the two legs again. The cyclical price component is indeed vulnerable - and that is exactly why the structural volume floor matters. Even if HBM average selling prices moderate in 2027, the bit-demand growth from agentic workloads and custom silicon - the 82% surge Goldman projects - sustains utilization. The concentration risk is real but is a risk of the index, not necessarily of the individual companies; an investor exposed to the memory thesis through the KOSPI is taking concentration risk on top of cycle risk, while direct exposure to the chipmakers isolates the cycle. The mechanical-trimming point is valid for index funds but cuts the other way for dedicated Korea or semiconductor mandates, which have been net buyers on weakness.
The falsifying signal is specific: if DRAM contract prices fail to rise quarter-over-quarter for two consecutive quarters while AI server shipments continue to grow, the structural-shortage thesis is wrong - it would mean capacity has caught up with demand faster than the order book suggests, and the entire premise of persistent scarcity collapses. Watch the quarterly contract-price prints from industry trackers and AI accelerator shipment data; a divergence between rising shipments and flat prices is the tell.
Conclusion: What to Watch, and What Could Break
The mechanism cashes out into a clear asymmetry. The direct beneficiaries are the memory suppliers with the deepest HBM qualification: SK Hynix first, Samsung second, Micron third. The exposed are the index investors who think they own "Korea" but own a leveraged memory proxy, and the GPU-centric investors who assume memory will scale elastically to meet their roadmap. The second-order beneficiaries, if agents migrate to the edge, are the mobile-memory and NAND suppliers - a broader set than the HBM duopoly narrative implies.
By time horizon: in the short term (weeks), sentiment and liquidity dominate - the Connect keynote on September 23 and any glasses or agent-capability announcements can extend the rally or trigger profit-taking after a sharp run. In the medium term (quarters), fundamentals dominate - the third- and fourth-quarter 2026 DRAM contract-price prints and the HBM supply-demand balance reported by industry research firms will determine whether earnings revisions keep climbing. In the long term (years), the structural question dominates - whether agentic AI becomes a persistent, parallel workload that permanently raises memory content per user, or whether it proves to be a feature that consolidates into fewer, larger models running on existing hardware.
Three scenarios frame the path:
- Base case: HBM supply remains tight through 2027, pricing stays firm, and SK Hynix and Samsung grind higher as earnings revisions continue. Trigger: quarterly contract prices up sequentially and HBM lead times extending beyond six months, as industry trackers report.
- Upside case: Meta's Connect announcements accelerate agent adoption, edge deployment expands the memory mix, and the shortage deepens into 2028 as Daishin's 40-billion-Gb order book implies. Trigger: Muse usage metrics or glasses shipments that imply a step-change in memory content per user.
- Downside case: New HBM capacity from all three suppliers comes online faster than expected, contract prices flatten for two consecutive quarters while AI shipments grow, and the KOSPI's concentration forces mechanical de-risking. Trigger: two flat quarterly DRAM contract prints amid rising shipments.
The signal that would prove the bullish structural read wrong is the one named above: flat DRAM contract prices for two straight quarters alongside growing AI server shipments. That combination would show the shortage was cyclical after all.
Here is the judgment condensed: Meta's Muse did not create the memory shortage, but it gave investors a credible reason to believe the shortage will outlast the cycle - and in a market where half the index is two chip stocks, that belief is the only thing standing between the KOSPI and its next double-digit drop.
Explore more exclusive insights at nextfin.ai.
