Memory Is Becoming Infrastructure

In Brief

• My conclusion: this is not a normal memory cycle anymore. The market is still right to discount generic DRAM cyclicality, but it is probably underpricing the durability of scarce AI-linked memory capacity.
• The investable distinction is not memory versus no memory. It is high-end AI memory scarcity versus broad commodity memory beta.
• Multi-year supply agreements are the key tell. They suggest customers are reserving future capacity, not just chasing this quarter's spot price.
• I would not pay a full Nvidia-style multiple for Korean memory earnings, but I also would not value HBM-led earnings like a classic peak-cycle commodity profit stream.

The Observation

The obvious story in memory is that prices are rising. The sharper story is that the buyer behavior is changing.

In a normal memory upcycle, investors watch contract prices, inventories, and capex plans, then try not to overstay the moment when supply catches demand. The industry has trained investors to distrust peak margins. When DRAM looks best, the next downturn is often already being built.

The AI cycle complicates that reflex. Hyperscalers and AI infrastructure customers are not merely buying memory for the next device refresh. They are trying to secure future compute capacity, and memory bandwidth is one of the gates. If a data-center plan depends on guaranteed access to high-end memory over several years, the supplier is no longer selling only this quarter's components. It is selling future availability.

That is why recent reporting on three- to five-year memory supply agreements is more important than another bullish price forecast. Duration changes the economics. The question becomes less "where is the spot price?" and more "how much of future capacity has already been spoken for?"

My View

My view is that high-end AI memory deserves a partial re-rating, not a full escape from cyclicality.

The market is too blunt when it treats all memory profit as peak-cycle commodity earnings. HBM and other AI-constrained memory capacity now sit closer to the strategic bottleneck of the data-center buildout. When customers sign multi-year contracts to secure supply, some of the producer economics start to look like reserved infrastructure capacity rather than pure spot-market volume.

But the market would also be wrong to treat the whole memory complex like software or like Nvidia. Memory companies still manufacture into a cycle. Capacity can be added. Product leadership can shift. Lower-end pricing can be pressured by China. Customer urgency can fade if AI capex slows.

So the conclusion is narrow but important: the right multiple should go up for the scarce, contracted, AI-linked part of memory earnings, while the commodity part should still be discounted. The edge is in separating those two earnings buckets faster than the market does.

Why The Old Cycle Framework Is Incomplete

Memory has always had moments when people announce that the old cycle is dead. Usually it is not. High margins invite new capacity. Customers double-order. Inventories normalize. The same earnings that made the stock look cheap disappear.

This time, the old framework misses one new feature: AI data centers are long-lead-time systems. GPUs, HBM, advanced packaging, power delivery, networking, cooling, and grid access have to line up together. If memory is one of the scarce inputs, customers have an incentive to reserve it earlier and for longer.

That does not make memory software. It does make some memory capacity look more like infrastructure. Scarce future capacity can carry a different bargaining position than interchangeable commodity output.

The margin evidence is why this deserves attention. Recent Korean coverage framed SK Hynix and Samsung memory profitability as approaching, and in some estimates exceeding, the operating-margin profile associated with the best semiconductor platforms. That is unusual for a segment investors still reflexively discount as brutally cyclical.

The Valuation Paradox

The useful angle is not "memory stocks are up." Everyone can see that. The useful angle is that earnings have exploded while the market still appears unwilling to pay a platform multiple for them.

On the reported numbers, SK Hynix and Samsung have already rallied hard, yet their price-earnings multiples remain far below the multiples attached to TSMC and Nvidia. Part of that gap is rational. If current earnings are peak-cycle earnings, a low multiple is exactly what should happen.

But if even part of those earnings is protected by multi-year AI infrastructure commitments, the market may be using too punitive a denominator. A low multiple on peak commodity profit is not cheap. A low multiple on partially contracted capacity rent is different.

That is where I think the market may still be slow. It is not that Korean memory should trade like Nvidia. It is that HBM-led earnings should not be thrown into the same bucket as generic commodity memory earnings.

What I Would Watch

The first signal is contract breadth. If long-term agreements are limited to the tightest HBM products, this remains a narrower SK Hynix-style scarcity story. If they extend into broader high-value memory capacity tied to AI buildouts, the re-rating case gets stronger.

The second signal is capex discipline. The fastest way to kill this thesis is for producers to treat high margins as permission to overbuild. Scarcity only deserves a better multiple if supply response stays controlled.

The third signal is product mix. SK Hynix's HBM position is not the same as Samsung's broader memory recovery. Treating both companies as one generic "memory trade" misses the key distinction. I would pay more attention to who owns the scarce layer than to who simply benefits from rising memory prices.

The fourth signal is AI capex durability. If hyperscalers slow buildouts, push back on GPU orders, or find architectures that reduce memory intensity, the long-term visibility argument weakens quickly.

The Bottom Line

I come down on the side that the market is still underestimating the quality shift in AI-linked memory earnings, but only for the constrained part of the stack.

This is not a blanket argument to buy every memory stock after a huge rally. It is an argument that the old "low multiple because memory is always cyclical" shortcut is now too crude. Some earnings still deserve that discount. Some do not.

The best way to frame the opportunity is not "memory supercycle." It is "which part of memory has become reserved AI infrastructure capacity?" My answer: HBM-led, customer-committed capacity deserves more respect than the market's historical memory playbook gives it. Broad commodity memory still has to earn that respect cycle by cycle.

Source Notes

• MK Economy, "장기계약이 판 바꿨다…메모리 2028년까지 슈퍼사이클," fetched 2026-05-03T08:29:57Z.
• MK Economy, "증시 호황에 올해 10대그룹 시총 1500조," fetched 2026-05-03T08:20:57Z.
• BNK Investment Securities, "선진국 성장률 하향, ETF 중심의 수급 변화," fetched 2026-05-03T06:57:44Z.
• Hana Securities, "신고가 이후 순환매의 기준: 영업이익률," fetched 2026-05-03T06:57:44Z.
• Hana Securities, "삼성전자 -- 갈 길이 멀다," fetched 2026-05-02T18:55:58Z.