Crossgrain No. 1: They Will Be Just Fine

Crossgrain No. 1 examines how facts become stories, and how stories acquire authority.

The Story:

Artificial intelligence needs accelerator chips, which need High Bandwidth Memory, HBM, the specialized memory stacked beside AI processors. Supply is tight. 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 real.

We've heard every sentence above.

Here is my question.

If SK Hynix keeps getting better at this, does "three to one" survive? And if the shortage eases, will the premium still stand?

We don't know.

The contract is private. But the market talks as though shortage, three-to-one, and long-term contracts were the whole of the economics. They aren't.

The number is "3". The man who gave the market that number attached words to it. 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 what the story leaves out.

  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.

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 holds. 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

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.

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 dies on the way as if both used up capacity the same way.

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

Mehrotra's verb matters: get to. Three is the destination. So actual wafer consumption today must be higher 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

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 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 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 company declined and offered that the agreements contained "robust terms."6

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

Crossgrain their stories

Crossgrain is the cut made against the direction the wood wants to split.

Stories have a grain too. Words run along it, outward from where they were first said, and lose something at every step. To cut across it is to go back to the source.

A private contract leaves a hole in the story. The market does not leave it empty.

On Micron's June 2024 call, "three times" was read from the script.2 But the yield assumption came out only in the Q&A, when an analyst asked.3

Try a test. Take any confident sentence about these contracts and replace each word with what it actually specifies. Long-term: how long, with what exits? Robust: for whom, against what? Premium: over what, fixed on what date? Where the definition should stand, we find a blank.

The contracts stayed closed. The market moved anyway. Some of us are asked to pay for their stories.

A premium can also be a shackle

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

Yet 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: "Yes, it is correct that the margins for non-HBM today are higher than HBM margins."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, while HBM prices locked in annual and long-term contracts did not seem to move with them. The same rigidity that can let 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.

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 HBM demand has become "sticky."10 Samsung calls it an "unprecedented supercycle."11

These companies themselves are telling the structural story.

But the record shows three companies doing different things while the future was still unclear.

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 underuse of its Utah fab cost it $557 million in fiscal 2020.12

Samsung had HBM and retreated from it. In 2019 it dissolved its dedicated HBM team, judging the market not big enough to matter.13 When that judgment proved wrong, 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 was better. It stayed with HBM through years of thin demand. By 2019, it was already naming machine learning and artificial intelligence among the applications for HBM2E.15 But HBM had begun as a solution to bandwidth, power, and space constraints in graphics and high-performance computing, and first shipped inside a gaming card in 2015.16 And in October 2020, SK Hynix signed the largest deal in its history, $9 billion for Intel's NAND and storage business.17 NAND is commodity flash, the memory that stores files, the opposite of a strategic bet. When the bottleneck arrived, SK Hynix had what the future suddenly needed.

"This Time Is Different"

Their spending shows what they expected. In August 2020, IC Insights forecast Samsung's DRAM capital-expenditure budget falling 21 percent, SK Hynix's by 38 percent, and Micron's by 16 percent.18

The cuts made sense. The industry had just come through 2018–19, when prices were cut in half and margins collapsed with them. But a budget is not an opinion. It is money spent or withheld. A company that expects a structural break builds through the downturn. A company that expects another turn of the cycle cuts, keeps its cash, and waits. In 2020, they cut.

The factories that could have relieved the present shortage take years to build, and none was started in time. The cuts made in the last cycle helped create the shortage. The shortage is now presented to the market as evidence that the cycle is dead.

They are not weak. So?

The moat of the three suppliers 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 a conglomerate around its moat; Micron, $6.2 billion in U.S. government grants, more than the other two combined19; SK Hynix, the strongest position in the room.

They will be just fine.

But the 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 those gates.

As HBM grows more customized, supplier and customer lock together more tightly. Suppliers become harder to replace while growing more dependent on the buyer's schedule, architecture, and approval.

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

What the shortage does not tell us

The shortage can be real. But the story grows. It begins with HBM for AI accelerators. Then inference, custom chips, autonomous systems, the edge, robotics, and eventually AGI are folded into the same case. Any of them may become real, but each requires its own evidence. A bottleneck in one system does not prove the future of another.

The asymmetry

Samsung once dismantled its HBM team. It is still formidable.
HBM began as a graphics fix. SK Hynix still leads.
One of Micron's bets failed. Micron is still one of the three.

Their moat forgives their mistakes.

Once a moat becomes valuable, the old decisions begin to look more deliberate than they did at the time. The uncertainty disappears from the retelling.

But.

We have no moat.

Every boom has its number. This one's was "3".

Notes

  1. TrendForce press release, May 20, 2024, “HBM3e Production Surge Expected to Make Up 35% of Advanced Process Wafer Input by End of 2024”: HBM yields of “around 50–60%.” Second-hand: TrendForce News, May 24, 2024, “SK Hynix Revealed Progress for HBM3e, Achieving Nearly 80% Yield,” relaying the Financial Times (SK hynix’s head of yield: close to the 80% target for HBM3E) and Business Korea (40–60% for the through-silicon-via stage). The figures come from different sources and measure different things. ↩

  2. Micron fiscal Q3 2024 call, June 26, 2024. Sanjay Mehrotra, prepared remarks, supply section. The sentence opens “Industry-wide, HBM3E consumes approximately three times…”; the next sentence says the HBM4 ratio will be even higher. ↩ ↩

  3. Same call, Q&A. Mehrotra, answering Vivek Arya (Bank of America Securities), who asked what happens to the 3x ratio as stacks get taller and yields improve. The answer begins “Yes, I mean, I think for HBM3E, 3x here is the operative guidance”; all three quotations come from it. The opening capitals of the second and third quotations are the author’s. ↩ ↩ ↩

  4. Micron fiscal Q4 2024 call, September 25, 2024. Q&A, Mehrotra, answering Toshiya Hari (Goldman Sachs), who asked about upside to the 2025 HBM plan. The passage begins “And yields, we expect…”; the quotation in the text starts at “we,” capitalized. ↩

  5. Same call as note 3, same answer. Mehrotra: the 12-die stack “will have somewhat lower mature yields,” and the HBM4 trade ratio rises above 3x. ↩

  6. Micron fiscal Q2 2026 call, March 18, 2026. Prepared remarks, Mehrotra, opening paragraphs (“fundamentally recast memory,” “structural supply constraints”); the first five-year strategic customer agreement is announced there. Q&A: Krish Sankar (TD Cowen) asked whether price is committed for several years or renegotiated annually, and about cancellation terms. Mehrotra declined, calling the agreements confidential, and said they have “robust terms.” ↩ ↩

  7. TrendForce press release, June 2, 2026, “Tight DRAM Supply Gives Suppliers Greater Pricing Power in HBM, with HBM Contract Prices Expected to Surge Multiples Higher in 2027.” TrendForce’s own estimate of per-wafer revenue, built from die size, yield rates, and per-Gb pricing: HBM wafer revenue was overtaken by DDR5 64GB RDIMM in 1Q26, and HBM profitability fell below it from the same quarter. The release opens by saying the annual pricing mechanism the three suppliers use for HBM has kept contract prices from fully reflecting quarterly market increases. ↩

  8. Micron fiscal Q2 2026 call, March 18, 2026. Q&A, Mehrotra, answering Christopher Muse (Cantor Fitzgerald), who asked whether any industry players were moving bits from HBM toward DDR5 given the higher margins there. The answer opens with the sentence quoted. ↩

  9. Chey Tae-won, Welcoming Remarks, Trans-Pacific Dialogue 2026, Washington, D.C., February 20, 2026; printed in the Chey Institute’s Trans-Pacific Dialogue 2026 Conference Report, Part 1, pp. 10–12. He puts the margin on HBM at 60% (“over 60%” in one passage) and on conventional chips at 80%. ↩

  10. SK hynix Q2 2025 earnings call, July 24, 2025. Q&A, an unnamed executive speaking through an interpreter, answering Dana Chae (Korea Investment & Securities): “the demand for HBM has become very sticky.” ↩

  11. Reuters, March 18, 2026, report on Samsung Electronics’ annual shareholders’ meeting. Second-hand: the words “unprecedented supercycle” are as Reuters reported them. ↩

  12. Micron press release, March 16, 2021 (Form 8-K, Exhibit 99.1): Micron said it had determined there was “insufficient market validation” to justify the investment needed to commercialize 3D XPoint at scale. Underutilization costs at its Utah fab, as Micron reported them: $557 million in fiscal 2020 (Form 10-K for fiscal 2021, where the fab appears as MTU). ↩

  13. Second-hand: TrendForce News, April 2, 2024, “Samsung Reportedly Establishes New HBM Team, Looking to Improve AI Chip Yield,” relaying the Korea Economic Daily of March 29, 2024. Samsung dissolved its HBM team in 2019 because it wrongly believed the market would not grow significantly. ↩

  14. Korea JoongAng Daily, September 20, 2025, “Samsung Electronics passes Nvidia’s quality test for HBM3E chip supply.” Citing industry sources, Samsung’s 12-layer HBM3E passed Nvidia’s quality test after more than 19 months without approval; Samsung said it cannot confirm client-related matters. Conflicting report, second-hand: Aju News of the same day, as summarized by TrendForce News on September 22, 2025, said Samsung had not yet received Nvidia’s formal approval. ↩

  15. SK hynix press release, August 12, 2019, “SK hynix Develops World’s Fastest High Bandwidth Memory, HBM2E.” The subhead names “high-end machine learning” among expected applications, and the text calls HBM2E an optimal solution for systems including machine learning and artificial intelligence. ↩

  16. AMD press release, June 16, 2015: the Radeon R9 Fury X is “the world’s first HBM-powered graphics card,” with availability scheduled for June 24, 2015. AMD claims 60% more memory bandwidth than GDDR5, more than three times the performance per watt, and 94% less board area. Second-hand for the earlier history: Korea JoongAng Daily, September 2, 2025, “How HBM split the paths of Samsung, AMD and SK hynix” (AMD began its HBM project in 2007; Hynix agreed to partner in 2010; AMD and SK hynix helped JEDEC release the first HBM specification in 2013). ↩

  17. Intel and SK hynix joint announcement, October 20, 2020, “SK hynix to Acquire Intel NAND Memory Business”: SK hynix agreed to acquire Intel’s NAND memory and storage business for US$9 billion, paid as $7 billion at the first closing (expected late 2021) and $2 billion at a final closing (expected March 2025). Second-hand for “largest”: Korea Economic Daily, December 23, 2021, called it SK hynix’s largest-ever acquisition. ↩

  18. IC Insights, August 19, 2020, “DRAM Capex Spending Expected to Decline 20% in 2020”: forecast cuts to DRAM capital spending of 21% for Samsung (to $4.9 billion), 38% for SK hynix (to $4.0 billion), and 16% for Micron (to $3.6 billion). These are forecasts of budgets, not reported spending. ↩

  19. U.S. Department of Commerce, CHIPS Incentives Program announcements: Micron, up to $6.165 billion in direct funding (December 10, 2024); SK hynix, up to $458 million in direct funding plus up to $500 million in loans (December 19, 2024); Samsung, up to $4.745 billion in direct funding (December 20, 2024). The Department pays on completion of project milestones. ↩