Memory That Outearns TSMC: Why This Cycle Isn't Just a Cycle

This is the sector in my "Industry Research" series where I have the most conviction — because over the past year, I have dissected almost every company on this chain: SK Hynix, Samsung, Micron, and SanDisk. This piece puts them back into the industry landscape to answer a question that no single company's earnings report can answer: Is this the top of the cycle, or a qualitative change in the business?

01存储曾经是半导体里"最不体面"的生意

To understand how anomalous this cycle is, you need to remember what memory used to be like.

In the semiconductor food chain, memory (DRAM, NAND) has long been the least profitable, most looked-down-upon segment. Logic chips have Nvidia and Intel's design moats, foundry has TSMC's process monopoly, while memory — it's essentially a standardized, interchangeable commodity. A DDR5 die from Hynix and one from Samsung are indistinguishable on a motherboard.

The fate of a commodity is cycles. For the past three decades, the memory industry has repeated the same script: boom → everyone adds capacity → oversupply → price collapse → massive losses → production cuts → supply clear → boom again. I wrote in my piece on Howard Marks's The Cycle that commodity money is "hard money" — you don't earn on moats, you earn on betting the timing of supply and demand. Memory is the purest specimen of this script: even a slight misalignment in the capex decisions of the Big Three can flip the industry from fat profits to deep losses.

Historically, memory's profit peaks almost never lasted more than 2-3 quarters. That was the iron law of its cyclicality: you could buy at the bottom and sell at the top, but you could never count on it to "keep making money." Because if it kept making money, someone would expand capacity, and expansion always leads to the next crash.

So for thirty years, memory manufacturers sat at the bottom of semiconductor profit distribution. Logic and foundry ate meat, memory sipped soup — and when the cycle turned, there was no soup at all.

02这一轮的反常:利润率,超过了台积电

Keep that background in mind, and look at the numbers this time — only then do you realize how abnormal it is.

Hold on that row: SK Hynix 79%, Micron 75%, and Micron guiding ~81% next quarter — while TSMC, the industry's公认 "cash machine" of high margins, sits at only 54-56%.

This is an extremely rare sight in semiconductor history. The business that sat at the bottom for three decades, earning the hardest money, getting crushed by cycles, for the first time systematically — not just for an isolated quarter — has surpassed the apex predator of the food chain in profit margins. 79% gross margin is no longer hardware territory — it's closer to a software company.

If you still use the "cyclical stock" framework, your instinct would be: This is the top, time to run. Because under the old script, such fat profits can't last more than two or three quarters, then comes expansion, oversupply, and collapse.

But this time, several structural changes I've observed make me fundamentally doubt that instinctive conclusion.

03四个结构性证据,让"周期"这个词开始失效

I boil down the differences between this cycle and every past one into four points. Together, they point to one thing: What is driving this boom is no longer a self-reversing cyclical force, but a set of structural forces that could persist for years.

First, demand is locked in by "hyperscaler capex," decoupled from consumer electronics.

In the past, memory demand followed PC and smartphone shipments — classic consumer cycles that swing with the economy. But this cycle's core demand comes from AI infrastructure: the capex of hyperscalers like Microsoft, Google, Amazon, and Meta. The nature of this money is completely different — it's multi-year, strategic, top-down planned enterprise spending that won't pivot just because consumer demand softens in a given quarter. SK Hynix said it straight in its earnings: "AI is evolving from training to inference and agentic AI. As the data volume generated by AI agents grows, the demand base for various memory types is broadening." In other words, the demand foundation has shifted from fickle consumers to deep-pocketed cloud giants that plan in multi-year horizons.

Second, supply is physically constrained by HBM — you can't ramp fast even if you want to.

The root of past cycle crashes was "boom → frantic capacity expansion." But this time, expansion is blocked by a physical reality: manufacturing HBM consumes about 3x the wafer area of an equivalent capacity of standard DRAM. That means every time a producer shifts a wafer's worth of capacity to high-priced HBM, it pulls 3 wafers' worth from standard DRAM. The result is a cascading effect — HBM eats wafers, causing a structural shortage in standard DRAM, and prices surge accordingly (TrendForce data: commodity DRAM contract prices rose 90-95% QoQ in the latest round). More critically, HBM isn't just a matter of adding a line — it requires advanced packaging, yield ramps, and deep customer collaboration. This means supply increases are slow and capped — the part of the cycle script where "everyone frantically builds capacity and crashes prices" can't happen quickly this time.

Third, pricing has shifted from "spot" to "long-term contracts," smoothing price volatility.

In the past, memory prices were set in the spot market, which is the direct cause of violent cyclicality — spot prices can double in a month or halve in a month. But this time, HBM and some high-end DRAM have moved to multi-year long-term contracts. SK Hynix has already negotiated or is advancing long-term contracts with Nvidia, Google, and Apple; its full-year 2026 HBM capacity has been fully booked by customers, with some large customers even prepaying to lock supply. Micron similarly announced its HBM capacity sold out through end of 2026. When prices and capacity are locked by multi-year contracts, the old transmission mechanism — "spot price crash" — is largely severed.

Fourth, the competitive landscape has solidified into a three-player oligopoly, with positions locked by technology barriers.

This is the most easily overlooked point, yet possibly the most important. HBM's demands on yield, packaging, and customer collaboration create entry barriers far higher than standard DRAM. The result is a market highly concentrated in three hands, with a technology gap locking the pecking order. Look at the latest HBM share landscape:

A three-player oligopoly, technology-driven ranking, and extremely high entry barriers — this is a structure where it's very hard to start a price war. When only three players exist and anyone who falls behind gets kicked out of Nvidia's supply chain, the rational choice is to maintain discipline, tilt toward high-value products, and not slaughter each other with capacity-driven price cuts. This is already a different species from the old "seven or eight players brawling, whoever expands first crashes prices" memory market.

Put these four together — demand locked by multi-year cloud capex, supply constrained by HBM's physical nature, pricing smoothed by long-term contracts, and competitive structure locked by technology moats — and you'll find that every mechanism that once made memory a cyclical stock has been weakened or severed this time. That is why its fat profits have now lasted beyond the "two-to-three-quarter" iron law, and look set to continue.

04HBM是这场权力转移的中枢

If the above four points answer "why is it different," HBM is the physical nexus. Understand HBM, and you understand why memory has, for the first time, taken the main seat at the table.

HBM (High Bandwidth Memory) essentially stacks multiple layers of DRAM vertically and bonds them tightly to the GPU using advanced packaging, solving the biggest bottleneck in AI compute — bandwidth. No matter how powerful an Nvidia GPU is, if data can't be fed in fast enough, the compute cycles are wasted. HBM is the critical organ that "feeds the data."

This creates a shift unseen in thirty years: memory, for the first time, has become the bottleneck and value nexus of the entire AI compute chain. In the past, the bottleneck was in logic and foundry; memory was a supporting part. Now, whether Nvidia can ship its next-generation GPU on time and in volume depends heavily on whether HBM is sufficient and yields are good. SK Hynix's HBM capacity has essentially become the "admission ticket" to Nvidia's GPU supply chain — a sentence that would have been unimaginable three years ago.

How big is the value? Look at pricing: HBM3E costs about $300 for a 36 GB stack; HBM4 is estimated at ~$500 per stack — dozens of times the unit value of a standard DRAM die. And it's still climbing: HBM4's energy efficiency improves another 40% over HBM3E, deepening its irreplaceability in AI servers.

So the real story this time is not "memory prices are rising"; it's that a new species — HBM — has bifurcated from memory. It is no longer a commodity, but a high-value component with technology barriers, customer lock-in, and pricing power. The fat profits of Hynix and Micron are fundamentally because HBM has transformed them from "sellors of commodity dies" into "sellers of critical AI compute components." That is why their margins surpass TSMC's — the nature of what they sell has changed.

Worth noting: for US equity investors, this has a practical handle. Among the Big Three, the only directly investable US stock is Micron (MU) — Hynix and Samsung are Korean stocks. So for US equity investors, Micron is the most direct vehicle for this theme, and its HBM share overtaking Samsung and winning mass production for Nvidia's Vera Rubin platform is the most important positioning shift to watch this year.

05但我保留一个怀疑:周期,真的死了吗

If I were to stop here and just say "this time is different," I would betray what I've always believed.

I once wrote in my piece on Reinhart and Rogoff's This Time Is Different: The four most expensive words in finance are "this time is different." At the top of every bubble, there is always a seemingly unassailable logic arguing that "the old rules are broken, this time is structural, it won't crash again" — and then it crashes.

So I must remain wary of my own "four structural evidences" above. Memory is a business with a thirty-year cyclical history; cycles don't disappear just because of a new narrative. Here are a few real risks I see:

First, capital expenditure can reverse. The demand foundation this time is hyperscaler AI capex. But that money is not a law of physics; it's a business judgment. If AI application monetization keeps lagging (huge compute investment, but applications haven't made money yet), cloud giants could slow capex at any time — and at that moment, the "demand is locked" logic would reverse. This is the biggest sword hanging over the entire chain.

Second, capacity expansion is happening, just slowly. Samsung and Hynix have both announced large capacity expansions for 2026. HBM expansion is slow, but not zero. When all three players have ramped up capacity and demand growth even just moderates (not even declines), the supply-demand balance will shift. The core cyclical mechanism — "high profits will eventually attract enough supply" — has not been abolished, only delayed.

Third, long-term contracts can't protect quantity. They lock prices, but they don't lock total demand. If AI capex peaks, customers will simply sign fewer or no contracts in the next round. Long-term contracts postpone the impact, but the cost is that the impact is more concentrated when it comes.

Fourth, profit quality needs to be discounted. When I dissected SK Hynix, I pointed out: in its "77% net margin," about KRW 11.5 trillion (nearly 29%) came from non-recurring items like forex and asset revaluation. Core net margin was actually around 55%. The prettier the headline number, the more you need to beware how much is real, sustainable operating profit.

So my verdict is a qualified "different": the structural changes this cycle are real — demand nature, supply constraints, pricing mechanism, competitive landscape are indeed unlike the past, allowing fat profits to persist longer than any previous cycle. But "cycle weakened" is not "cycle killed." Memory has not become a permanent-moat business like TSMC's process monopoly; it has just gained a "window that doesn't look like a cycle" thanks to the new species of HBM. This window will be longer than in the past, but it will still end in some form — probably not from a price crash, but from a peak in AI capex.

For investors, the real thing to track is therefore not "how many more quarters can memory rise," but the single master switch: Are hyperscaler AI capital expenditures a multi-year secular trend, or a wave that will eventually recede? The answer to that question decides when this window closes.

06写在最后

Memory sat on the bench for thirty years. In this AI cycle, it has taken the main seat at the table for the first time.

Its story is worth studying repeatedly not because any single company's earnings are breathtaking, but because it presents a rare, ongoing industrial power shift: value is migrating from "compute" (logic, foundry) to "store and transmit" (HBM, bandwidth); a former commodity is growing the bones of infrastructure.

But I won't forget the iron law I believe in. Structures change; human nature does not. Every "this time is different" is half true (structure indeed changed) and half illusion (the assumption that the change is permanent). In this memory cycle, the real part is the qualitative change brought by HBM; the illusion is taking the persistence of AI capex for granted.

So if only one sentence remains —

The fat profits in this memory cycle are not the cycle's peak; they are a window. And when this window closes depends not on memory itself, but on the upstream cloud giants who are spending on compute — how long they are willing to keep this bet going.

Understanding this sector is essentially answering a bigger question: Is this AI capex feast the foundation of infrastructure, or just another "this time is different"? I have no answer. But I know which switch to watch.

——

Industry Research series: Memory Overview. Follow-ups will drill down the chain: HBM technology landscape, the divergent fates of DRAM/NAND/HBM, and a head-to-head among the Big Three.