Crossgrain No.1 - They Will Be Just Fine

We know the AI memory story. Do we ask what it is made of?

The Story:

Artificial intelligence needs accelerator chips. Accelerator chips need High Bandwidth Memory — HBM, the specialized memory stacked beside AI processors. Supply is tight. New factories take years to build. Only three companies on Earth — SK Hynix, Samsung, Micron — can make the good stuff at scale. And HBM consumes roughly three times the manufacturing capacity of ordinary server memory. Therefore the shortage is structural, the suppliers have pricing power, and the premium is deserved.

We've heard every sentence above.

Here is my question.

When SK Hynix gets better at making HBM, who keeps the savings? Does Nvidia's price fall as SK Hynix's cost falls? Does SK Hynix pocket the difference? Was the improvement already assumed in the original price?

We don't know.

The contract is private, so our uncertainty is not the problem. The confidence built on top of it is. The market talks as though shortage, three-to-one, and long-term contract complete the economics. They don't. They are where the questions begin. Yet the questions tend to stop exactly where the numbers start sounding authoritative.

The number is "3". The man who gave the market that number actually attached some words to it, though. The number passed into analyst reports, shortage forecasts, and valuation models. The qualifiers did not travel with it.

The wafer before the story

Every memory chip begins as a wafer: a dinner-plate disc of silicon, the raw acreage of the industry. A wafer committed to HBM is a wafer that cannot become ordinary memory. That much, every headline gets right.

What happens next is where the story starts hiding the fundamentals. Allow me three sentences to explain. 1. The wafer is cut into chips, and the good ones are shaved thin and stacked — eight, twelve, or sixteen high — into a tower. 2. Microscopic copper channels are drilled vertically through the silicon so signals can travel between floors, and a foundation chip underneath manages traffic to the processor. 3. However, every floor added is another place for value to disappear.

Picture eight matchboxes stacked into a tower, then shipped across an ocean. Even with a 95 percent survival rate per layer, the eight-stack arrives intact about 66 percent of the time. A sixteen-stack, about 44 percent. Real production doesn't follow that clean arithmetic — chips are tested before stacking, and failures cluster rather than strike independently — but the direction is merciless: every added floor is another chance for the tower to die.

Industry estimates in 2024 put HBM success rates anywhere between roughly 40 and 80 percent, depending on the supplier and on which stage of the process was being measured.[1]

The width of that range is itself information. Outsiders repeat the word yield (the success rate) without knowing which yield they mean.

And a dead tower does not give its wafer back. It has already consumed the silicon, the cleanroom hours, the machines, the energy. The cost has already been incurred.

Headlines compress all of this into one line: memory capacity is being redirected to HBM. Correct but incomplete. That line treats a tower that works and a tower that died on the way as the same kind of capacity consumption. Economically, they are not. One is the cost of making the product. The other is the cost of not yet being good enough at making it.

Both tighten supply today. Only one necessarily remains tomorrow.

The number, missing its words

The three-to-one figure that anchors shortage stories has a specific author. Micron's CEO, Sanjay Mehrotra, told investors:

"HBM3E consumes approximately three times the wafer supply as D5 to produce a given number of bits in the same technology node."[2]

D5 refers to DDR5 DRAM, ordinary server memory. Strip the jargon and the sentence says: producing the same amount of memory requires roughly three times as much wafer capacity when the product is HBM rather than DDR5.

Three to one. 

Then an analyst asked what happens to the ratio as production improves. Mehrotra answered: the three-to-one holds "with mature yields." And he qualified the qualifier: "Our assumption of mature yields here is world-class mature yields."[3]

With mature yields. World-class mature yields.

"Three to one" is not a reading of what factories consume today. It is a statement of where a first-rate factory should arrive after the learning is done. Mehrotra put the tense on record himself: "We expect to get to mature yields on our HBM in fiscal year '25. Yields are always an upside opportunity."[4]

Get to. Future tense. Three is not the odometer. Three is the destination. Then, the actual wafer consumption must therefore be worse than three. 

Then a new generation arrives, the towers get taller, and the climb resets: Micron itself says twelve-layer stacks will settle at "somewhat lower mature yields" than eight-layer ones, and that the next generation's ratio will exceed three.[5]

So split the famous number open. Part of it is the product's shape: taller towers, drilled channels, a foundation floor — silicon surrendered on purpose, forever. No engineer fixes that away. The other part is the climb: scrap, misalignment, towers that died in assembly. That part shrinks as the factory learns.

Some of the three belongs to the product. Some belongs to the producer's current ability to make it. The market prices them as one thing.

Who owns the learning curve?

Micron never hid which direction the benefit flows. "As the yield will be ramping up for us," Mehrotra said, "of course, we will be able to get the benefit on the lower cost as we go forward."[3]

The factory improves. The supplier's cost falls. What happens next is not a semiconductor question. It is a contract question.

If the selling price stays fixed while yield improves, the supplier keeps the gain. If the price resets against an agreed cost curve, the buyer takes a share. If the deal was priced from day one on the assumption of mature yield, the buyer may have already collected — in advance.

A fixed price can favor either side: when costs fall it favors the seller; when market prices spike it favors the buyer. The phrase long-term contract tells us almost nothing until we know which variables are fixed, which are indexed, and when they reset.

That information is not public. When an analyst asked Micron whether its first five-year supply agreement contained multi-year price commitments, annual renegotiation, cancellation rights — the mechanics, in other words — the company declined, citing confidentiality, and offered that the agreements contained "robust terms."[6]

That is what the outside world received: an adjective and a duration.

Maybe Nvidia negotiated a mechanism that claws the yield gains back. Maybe SK Hynix keeps them. Maybe maturity was priced in on day one. Anyone outside the negotiating room who gives us a definite answer is guessing.

The point is that people who never ask how the learning curve is divided still speak confidently about what these contracts are worth.

A premium can also be a shackle

The story calls HBM a premium product whose margins are protected by long-term contracts.

Let's read what actually happened. In the first quarter of 2026, the research firm TrendForce estimated that the revenue a wafer earns as HBM had fallen below what the same wafer would earn as ordinary server memory — and that HBM's profitability had fallen below it too.[7]

Micron's CEO confirmed the direction in one sentence: "Non-HBM margins are currently higher than HBM."[8]

The chairman of SK Group put rough numbers on it: around 60 percent margins on HBM against around 80 percent on standard memory.[9]

Ordinary memory prices had surged, but HBM prices — locked in annual and long-term contracts — didn't seem to move with them; the same contract rigidity that lets a supplier keep its yield gains also stopped it from catching a price surge. A contract transfers risk in more than one direction, and we cannot know who won merely by hearing that it was long-term.

We will have to see whether HBM's "premium" prices hold. But the 2026 episode already shows that this premium is not a fact that persists. It is a number that has to be argued back into existence, every single generation. 

The product comes with a narrator

The current story is that AI has changed the business structurally. Micron says AI has "fundamentally recast memory as a defining strategic asset" and points to "structural supply constraints."[6] SK Hynix says memory is no longer a pure commodity market and that demand has become "sticky."[10] Samsung calls it an "unprecedented supercycle."[11]

The structural story is not something imposed on these companies. They are telling it themselves.

And it may even be true. But two different claims are hiding inside it. One: the industry has structurally changed. Two: the incumbents saw the change coming and positioned for it. The first can be true while the second is false.

So check the second against the record. It is public, and it reads the same at all three companies.

Micron spent years and billions on a new class of memory called 3D XPoint, positioned for a market that never arrived, and shut it down in 2021 citing "insufficient market validation" — after losing roughly $400 million a year.[12]

Samsung had HBM and retreated from it: in 2019 it dissolved its dedicated HBM team, judging the market too small to matter.[13] When that judgment detonated, Samsung spent years rebuilding what it had dismantled, and was reported to have passed Nvidia's qualification testing for advanced HBM only in late 2025.[14]

SK Hynix kept the right product. But HBM began in 2008 as a fix for power and space problems in graphics cards, shipped first inside a gaming card in 2015,[15] and in October 2020, with GPT-3 already public, SK Hynix wrote its biggest check — $9 billion — for Intel's storage-chip business: a bet on storage, not on the bottleneck that the AI boom would expose.[16]

Micron built the wrong product. Samsung walked away from the right one. SK Hynix kept the right one — aimed at a different market — and spent its largest bet on the wrong one.

SK Hynix deserves real credit for keeping it. Staying with an unfashionable technology through years of thin demand is a form of strategy, and it is why SK Hynix leads today. But endurance is not prophecy. That is not a record of three companies reading the same signal and moving early. It is one company enduring, one retreating, and one misfiring — and all three arriving at the same structural narrative after demand made the answer visible.

They did not spend as though they knew

Capital tells the same story more quietly. GPT-3 arrived in 2020. That year, Samsung cut its memory-factory investment by 21 percent, SK Hynix by 38 percent, Micron by 16 percent.[17]

There is a respectable defense: the industry had just survived the 2018–19 memory crash, when prices halved and margins collapsed. Cutting back was discipline, not blindness. Fine — but the defense has a price. If they cut because they didn't yet trust the demand signal, they were not foreseeing today's shortage. If they foresaw it and cut anyway, they chose not to act on their foresight. Neither version supports the story now attached to their market position: that this time is different.

The factories that could relieve the shortage take years to build, and they were not started in time — which is why the shortage exists. The scar from the last cycle helped create this shortage. Now the shortage is offered as proof that the cycle is dead.

Perhaps this time is different. Then again, historically, market participants have paid extraordinary sums for believing those words near the top of every memory cycle.

The moat is real. So is the gate.

None of this means the three suppliers are weak. Their moat is among the deepest in the industry: decades of process knowledge, tens of billions in equipment, customer approval cycles that take years, failure rates that would kill any newcomer.

Samsung has the moat plus a conglomerate around it. Micron has the moat plus $6.2 billion in U.S. government grants — more than the other two combined.[18] SK Hynix has the moat plus the strongest position in the room.

They will be just fine.

But the moat does not belong entirely to them. Market share in HBM is decided at the customers' approval gates, including Nvidia's. Samsung's ordeal proved that factories, capital, and thirty years of memory expertise do not let you walk around it.

As HBM grows more customized, supplier and customer lock together more tightly — harder to replace, yes, and also more dependent on one buyer's schedule, architecture, and approval. That is not sovereignty. It is mutual dependence with unequal visibility.

The companies know the contract. We only know the press release.

What the shortage does not tell us

The shortage can be real. The three-to-one can be real. The moat can be real. But none of it tells us: how much of today's wafer burden is the product's shape and how much is the climb; what yield assumption was priced into the contracts; who captures the cost reduction as factories improve; or whether long-term agreements have abolished the memory cycle — or merely redistributed it.

One thing I can tell you: the story grows. It begins with one bottleneck: HBM for AI accelerators. Then the collateral widens — inference, custom chips, autonomous systems, the edge, robotics, eventually AGI. Any may become real. But real demand does not need to annex territories to sound true. Each is a separate claim. A bottleneck in one system should not be allowed to collateralize an entire technological future.

The asymmetry

Samsung is no less formidable for having once dismantled its HBM team. SK Hynix loses no ground because HBM began as a graphics fix. Micron's factories do not vanish because one bet failed. 

A factory takes five years and tens of billions to replace. An approval relationship takes years to reproduce. Process knowledge cannot be ordered from a catalog. 

Once a moat becomes valuable, a story of foresight gets attached afterward — because a moat without a story feels unfinished to whoever is selling it, and unsatisfying to whoever is buying it. The foresight is the part that was added in post-production.

But what about us? We have less room for error. So I suggest that we crossgrain the narratives before we're asked to pay for their stories.

Crossgrain is the cut made against the direction the wood wants to split. With the grain, the story opens easily: AI demand, structural shortage, three suppliers, three-to-one, long-term contracts, premium product. Across the grain, the questions are less comfortable.

You knew the premium. You knew the moat. You knew the "3".

Did you ever ask what any of them contained?

Notes

[1] The 40–80 percent range combines estimates that refer to different stages of HBM production rather than one standardized finished-stack metric. TrendForce estimated HBM3E TSV-related yields of approximately 40–60 percent in early 2024. SK Hynix's HBM3E yield was reported near 80 percent, while Samsung estimates were generally lower. The figures should be read as evidence of wide supplier and measurement variation, not as directly comparable audited yields.

[2] Sanjay Mehrotra, Micron Q2 FY2024 earnings call, March 20, 2024.

[3] Sanjay Mehrotra, Micron Q3 FY2024 earnings call, June 26, 2024, response to Vivek Arya. Mehrotra said the three-to-one guidance included HBM3E's larger die, stack and logic-die requirements, and mature-yield expectations ("Our assumption of mature yields here is world-class mature yields"), adding: "As the yield will be ramping up for us, of course, we will be able to get the benefit on the lower cost as we go forward."

[4] Sanjay Mehrotra, Micron Q4 FY2024 earnings call, September 25, 2024.

[5] Micron Q3 FY2024 earnings call. Mehrotra stated that twelve-die HBM3E stacks would have "somewhat lower mature yields" than eight-die stacks and that the HBM4 trade ratio would exceed three.

[6] Micron fiscal Q2 2026 earnings call, March 18, 2026. Micron described AI as having "fundamentally recast memory as a defining strategic asset," cited "structural supply constraints," and announced its first five-year Strategic Customer Agreement. Asked whether the agreement contained multiyear price commitments, annual price renegotiation, and cancellation provisions, management declined to disclose terms, citing confidentiality, describing them as "robust."

[7] TrendForce, "Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM," June 2, 2026. TrendForce estimated that HBM per-wafer revenue and profitability fell below DDR5 64GB RDIMM in 1Q26, attributed part of the change to annual HBM pricing mechanisms lagging rapid increases in conventional DRAM prices, and reported that 2027 HBM4 negotiations were targeting prices several times higher.

[8] Sanjay Mehrotra, Micron fiscal Q2 2026 earnings call, March 18, 2026, as reported by The Elec and TrendForce.

[9] SK Group Chairman Chey Tae-won, reported by News1, The Elec, and TrendForce, March 2026.

[10] SK Hynix Q2 2025 earnings call. The company said HBM had shifted memory away from a purely commodity market, made demand more "sticky," and increased the bargaining power of leading suppliers.

[11] Samsung Electronics annual shareholder meeting, March 18, 2026, reported by Reuters. Co-CEO Jun Young-hyun described an "unprecedented supercycle" and said Samsung was pursuing three-to-five-year supply contracts.

[12] Micron SEC Form 8-K, March 16, 2021, citing "insufficient market validation." Annual operating losses of approximately $400 million were reported by industry publications. Texas Instruments acquired Micron's Lehi fabrication facility later in 2021.

[13] Korea Economic Daily, March 29, 2024, subsequently summarized by TrendForce. Samsung reportedly dissolved its HBM team in 2019 after judging that the market was unlikely to grow significantly.

[14] Korea Economic Daily, September 19, 2025. Samsung reportedly passed Nvidia's qualification for twelve-layer HBM3E about eighteen months after completing development, qualifying after SK Hynix and Micron. An industry executive quoted in the coverage described the expected volumes as "less about revenue and more about pride."

[15] AMD and SK Hynix began joint HBM development around 2008. JEDEC standardized HBM in 2013. AMD's Radeon R9 Fury X, launched June 2015, was the first major consumer product to ship with HBM.

[16] SK Hynix announced its approximately $9 billion acquisition of Intel's NAND and SSD business on October 20, 2020. The business now operates as Solidigm.

[17] IC Insights 2020 estimates: Samsung DRAM capital expenditure −21 percent, SK Hynix −38 percent, Micron −16 percent.

[18] U.S. Department of Commerce and NIST CHIPS award announcements, December 2024. Micron: approximately $6.2 billion in direct grants finalized for its Idaho and New York projects (a separate, smaller award for its Virginia facility is excluded here); Samsung: approximately $4.745 billion; SK Hynix: approximately $458 million in grants plus access to up to $500 million in loans.

Crossgrain, by Jung. Written between Korea and the United States.

Essays: conventional numbers and the distinctions they overlook.

Coaching: U.S. PhD SOPs & Global Mid-Career Strategy.