Why the AI Supercycle Keeps Stretching: Agent Demand, IPO Funding, and the Cheapest Asset Still Standing — Memory and Storage

Fri, 12 Jun 2026 15:30:00 +0900

This piece is the synthesis of a month of AI-infrastructure analysis. It connects Broadcom’s reaffirmed USD 100 billion target, Jensen Huang’s trillion-dollar Marvell comment, Goldman’s token demand vs. J.P. Morgan’s memory ASP peak-out, the USD 5.3 trillion AI data-center CapEx era, SpaceX’s listing and the Korean market, and Samsung-Hynix-Micron parity into a single map of the cycle.

TL;DR

The structure of the AI supercycle is shifting. (1) An explosion in token demand created by AI agents → (2) expanded funding by Big Tech → (3) public-market funding via mega-IPOs like SpaceX and Anthropic — the fuel hands off like a relay, and the lifespan of the cycle itself is lengthening.
Nomura has priced this shift in most aggressively. Per reporting, it set a Samsung Electronics target of KRW 590,000, an SK Hynix target of KRW 5,000,000, and a KOSPI target of 10,000-11,000, describing this year’s monthly memory revenue as “a vertical rise the likes of which we have never seen before.”
The longer the cycle runs, the more the real question becomes: “Which AI stock is still left?” The answer is surprisingly simple. Memory and storage — the final destination of demand, yet carrying the lowest multiples. Samsung Electronics and SK Hynix trade at a relative P/E of 0.80-0.82x versus Micron, and at roughly 6x on 2028 estimated earnings — among the lowest in the entire AI-chip basket.
The counter-argument is real too: J.P. Morgan’s slowing 2027 ASP growth, rate pressure, and earnings-bubble warnings. So the conclusion is not “buy at any price” but “bet on the unwinding of memory and storage’s undervaluation, while confirming the evidence of cycle extension step by step."

1. The Cycle’s Fuel Hands Off in Three Stages

Until now, the core of AI-cycle skepticism was “the money dries up first.” The logic: once the cash going into GPUs and data centers exceeds Big Tech’s cash flow, investment eventually rolls over and the cycle ends. Yet stitching together the news flow of the past month reveals a different picture — funding sources handing off one stage at a time, extending the cycle.

Stage 1 — Demand: Agents Detonate Token Consumption

The starting point is not humans but agents. Token demand in the era when a person typed a question into a chat window is in a different dimension from the era when agents search, run code, and read documents on their own to get work done. Goldman Sachs projected that token usage will rise 24-fold by 2030, and Nomura’s head of research, Chang-won Jung, went a step further, even using the phrase “AI-driven memory demand will grow 10,000x-20,000x in 5 years” (citing Sisa Journal reporting). What matters more than the precision of the numbers is the direction. It is now machine usage, not human usage, that determines the slope of the demand curve.

This demand is already confirmed in chip makers’ actual order books. Broadcom reaffirmed AI-semiconductor revenue exceeding USD 100 billion in 2027, and Jensen Huang publicly stated that Marvell could become a trillion-dollar company. As the customer base broadens to custom accelerators (XPUs), TPUs, and OpenAI’s own accelerators, demand has become multi-axis rather than resting on Nvidia alone.

Stage 2 — Funding: Big Tech Builds Even on Borrowed Money

The scale of money required to meet this demand has swelled to the unit of USD 5.3 trillion in data-center CapEx. Alphabet, Microsoft, Meta, and Amazon can no longer cover it with internal cash alone, so they are issuing more corporate bonds — and still not cutting investment. The reason is simple. Inference cost (cost per token) is falling 60-70% a year, creating a structure where margins improve the more you use. Whoever stops investing loses share.

Stage 3 — A New Funding Source: The Public Market Opens

This is the new chapter of this cycle. Until now, AI-infrastructure capital came from inside Big Tech (cash + corporate bonds). But with SpaceX filing its S-1 and, per reporting, pursuing a USD 75 billion offering, a third funding channel — the public market — is opening. Anthropic is likewise a mega-candidate whose listing is being talked about (we treat it as a scenario, as it is not officially confirmed). If these IPOs go through, retail and passive money will join the funding pipeline for AI-infrastructure investment.

In summary:

Stage	Funding source	Status	Meaning
Stage 1	Big Tech internal cash	Ongoing	Improving inference margins create reinvestment capacity
Stage 2	Corporate bonds / external borrowing	Expanding	Rate pressure exists, but no signal of investment cuts yet
Stage 3	IPO public-market funding	Emerging	SpaceX has an official S-1; Anthropic is at the speculative stage

The more diversified the funding sources, the lower the probability of the “money dries up and the cycle ends” scenario. This is the essence of the cycle’s extension.

2. What Nomura Saw: A Revenue Curve Rising Vertically

The house that translated this structural change into price targets most aggressively is Nomura. Organized per reporting (Sisa Journal, 2026-06-12 media briefing):

Item	Nomura’s call	Note
Samsung Electronics target	KRW 590,000	A large upgrade versus the current price
SK Hynix target	KRW 5,000,000	Top pick of the memory supercycle
KOSPI target	10,000-11,000	Cites a 60% probability of MSCI developed-index inclusion
Core rationale	“This year’s monthly memory revenue is a vertical rise, a pattern never seen before”	Chang-won Jung, Co-Head of Asia Research

The rationale matters more than the target itself. Nomura’s logic overlaps precisely with the picture we have tracked for a month. Past memory cycles rose and fell roughly once every 2-3 years, following PC and smartphone replacement cycles. This time, agents — a new demand source — have layered structural growth on top of the cycle, and as a result the monthly revenue curve is climbing at a slope unlike any prior cycle.

Of course, a foreign house’s aggressive target cannot in itself be a reason to buy. Still, the fact that this upgrade came right as foreigners began returning after 24 sessions of selling can be read as a signal that the direction of foreign flows and the sell-side view have started pointing the same way.

3. So What’s Left: The Most Powerful and the Cheapest

Even if you agree the cycle is lengthening, what to buy in an already much-higher market is a separate question. Here the valuation map gives the answer.

Let’s lay out again the figures we confirmed in the Samsung-Hynix-Micron parity analysis and the AI-chip P/E map follow-up.

Asset	Valuation	AI demand exposure	Verdict
US AI-chip large caps	High multiples maintained	Direct	Strong but already expensive
Micron	US-listed AI-memory premium	Direct	Premium is justified, but expensive versus Korea
Samsung Electronics	Relative P/E 0.82x vs. Micron, ~6x 2028E P/E	Direct (HBM4E catch-up + eSSD + foundry option)	Cheapest direct exposure
SK Hynix	Relative P/E 0.80x vs. Micron	Direct (HBM leader)	Discounted versus its tech edge
Storage (eSSD / NAND)	Early-stage multiple re-rating	Direct (inference / RAG / KV-cache storage)	A zone where reclassification is underway

The core logic compresses into three lines.

The final destination of demand is memory. Whether agents burn tokens or SpaceX builds satellite networks and AI compute, what is ultimately needed is HBM, server DRAM, and eSSD. Memory always sits next to the GPU, and as inference grows, storage demand follows. An AI server uses roughly 3x the SSD of a general server.
And yet it is priced as if it were late-cycle. Among assets exposed to the same AI demand, only Korean memory trades at roughly 6x on 2028 estimated earnings. It means the market has not yet fully shed the old-cycle frame of “memory is about to roll over.”
The longer the cycle runs, the harder this discount is to sustain. As the funding source hands off through three stages and the lifespan of demand extends, the “about to roll over” assumption breaks down, and the room for multiples to normalize grows. Nomura’s targets are numbers that pre-empt that normalization.

Storage sits one step earlier in the sequence. As we saw in the Fadu-SanDisk analysis, NAND and eSSD are in the early stage of being reclassified from “always the worse cycle asset than DRAM” to “the storage bottleneck of AI inference.” If the reclassification is correct, the upside is large — but so too are the items to verify (price, volume, new markets) remain many.

4. Red Team: When This Logic Is Wrong

The more comprehensive the piece, the more clearly the counter-argument must be written.

When J.P. Morgan’s peak-out is right. The year-over-year growth rate of DRAM and NAND ASPs may slow from 2027. As we saw in the P-Q-C separation analysis, this is a deceleration in the pace of increase, not a price crash — but share prices react first to momentum, not earnings. Momentum deceleration can arrive before the memory multiple normalizes.
When rates squeeze the cycle. As organized in the AI supercycle mid-game check, the more funding shifts toward corporate bonds and IPOs, the more sensitive the cycle becomes to rates. Infrastructure built on debt and public offerings wobbles first when rates rise. It is a yellow light right now, not a red one — but this signal must be watched continuously.
When IPOs absorb funds rather than supply them. If the SpaceX and Anthropic offerings are funded by selling existing AI positions, then in the short term liquid AI large caps like Samsung Electronics and SK Hynix could become the sell target (the funding source). The material for cycle extension is also the material for short-term volatility.
When rhetoric like “10,000x, 20,000x” is itself a sign of overheating. There is also the classic counter that the sell-side’s “vertical rise” rhetoric is itself a hallmark of late cycle. The warning from the earnings-bubble analysis — that when earnings are a bubble, share prices fall first, before estimates are cut — applies equally to the logic of this piece.

5. Checklist: How to Confirm Cycle Extension Step by Step

Things to confirm, instead of to predict.

Item to check	Strengthens the extension thesis	Weakens the extension thesis
Token / agent usage	High growth sustained	Growth rate slows
Big Tech CapEx guidance	Maintained / raised	Cut, or emphasis on “efficiency”
Corporate bond rates / spreads	Stable	Spreads widen
SpaceX IPO	Strong demand, stable trading	Offering shrinks, post-listing plunge
Anthropic listing status	Enters official process	Indefinitely postponed
HBM long-term contracts	Prices held / raised	Renegotiation, volume deferral
Korean memory relative P/E	Normalizes from 0.8x toward 1.0x	Re-declines below 0.8x
Foreign flows	Net buying sustained / broadening	Reverts to net selling again

Conclusion

The skepticism that what ends the AI supercycle is not demand but money was only half right. Demand is being detonated by agents, and money is relaying from Big Tech cash to corporate bonds and on to the public market, expanding its sources. A cycle with diversified funding runs longer than past ones.

And in a lengthened cycle, the most advantageous seat is the asset that is the final destination of demand yet still sits at late-cycle pricing. That is memory, and one step earlier, storage. Nomura’s KRW 590,000 for Samsung Electronics and KRW 5,000,000 for SK Hynix are not in themselves a buy signal but a milestone that this re-rating has begun spreading into the sell-side consensus.

The practical stance is one sentence. Confirm the evidence of cycle extension (token usage, CapEx, IPO reception, HBM contracts) step by step, weight toward the unwinding of Korean memory’s relative discount and the reclassification of storage, but ease off the pace if signals of rate pressure and ASP-momentum deceleration light up.

The Nomura targets and comments in this piece cite Sisa Journal reporting (2026-06-12 media briefing), and Anthropic’s listing is at a speculative stage prior to official confirmation. Mentions of specific names are not investment recommendations but examples illustrating an analytical flow; the actual investment decision and responsibility rest with the investor.

노무라 on Korea Invest Insights