{"id":"hbm-wafers-dram-shortage-02663","slug":"hbm-wafers-dram-shortage-02663","title":"How HBM and Slow Fabs Squeeze the RAM Supply","description":"High-bandwidth memory for AI accelerators uses roughly three times the silicon wafer area of ordinary DDR5, quietly draining conventional RAM capacity even as manufacturers spend billions. Cleanroom shells take years to build, meaning new fabs from Micron, SK hynix, and Samsung won't meaningfully boost supply until 2028 or later. With three companies controlling around ninety percent of DRAM revenue, the physical bottleneck also collides with a history of price-fixing—and a newly surviving class-action lawsuit.","excerpt":"HBM3E memory for AI accelerators consumes three times the silicon wafer area of ordinary DDR5, squeezing conventional RAM.","pubDate":"2026-11-19T12:42:32.633Z","tags":["pc-hardware","gpus"],"heroImage":"https://files.myweirdprompts.com/covers/hbm-wafers-dram-shortage-02663.png","podcastAudioUrl":"https://episodes.myweirdprompts.com/audio/hbm-wafers-dram-shortage-02663.m4a","podcastDuration":"22:30","episodeNumber":6678,"transcript":"Corn: Three times the wafer. Same amount of memory, three times the silicon.\n\nHerman: That's the number that reframes the whole RAM shortage, and almost nobody leads with it.\n\nCorn: Daniel's follow-up this week goes straight at the economics of it. He's saying, look, you'd expect a memory maker that wants to maximise profit to see all this AI demand and say great, let's churn out more chips, everyone's happy. His words. And instead we've got real scarcity and inflated prices, which implies something is physically in the way. His question is what.\n\nHerman: And it's the right question to ask, because the answer isn't one thing. It's a cleanroom you can't build in a hurry, a product mix that quietly eats your conventional capacity, and an industry with a long memory of what happens when it overbuilds. Some of that is physics and some of it is a choice, and the interesting part is where the line sits.\n\nCorn: Then let's start where the wafer starts.\n\nHerman: A DRAM fab is, at heart, a cleanroom shell with a few hundred extremely expensive machines bolted inside. Lithography tools that print the patterns, etch and deposition tools that carve them out and build them up, all of it running around the clock on silicon wafers. And the thing people miss is that the shell is the bottleneck, not the tools. You can order the tools. You cannot order the shell quickly.\n\nCorn: How slowly are we talking?\n\nHerman: Micron's own chief executive, Mehrotra, put it plainly on their results call — cleanrooms take a long time to build, and even after first wafers go through, capacity only rises gradually. The new spending Micron announced targets cleanroom space that isn't usable until the second half of twenty twenty-eight.\n\nCorn: So the money they're spending today buys output for a decade that's mostly still ahead of us.\n\nHerman: Right, and the fab timelines make it concrete. Micron's Idaho first fab, ID1, takes its first wafers around mid twenty twenty-seven — and that was the pulled-forward date. The second Idaho fab is late twenty twenty-eight. Their big New York fab is around twenty thirty. SK hynix has a fab at Cheongju, M15X, built for HBM, that started taking wafers early this year and is ramping gradually — gradual is the operative word. Their Yongin site's first cleanroom was pulled forward to roughly February or March of twenty twenty-seven. Samsung's Pyeongtaek P5 — that one had actually paused for years, construction only resumed recently, operations expected around twenty twenty-eight.\n\nCorn: Let me make sure I'm reading that list right. P5 paused for years. Why would you pause a fab mid-build in the middle of all this?\n\nHerman: Because it was paused after the last downturn. When the memory market collapsed around twenty twenty-two and twenty twenty-three, Samsung stopped pouring concrete on it, and construction only resumed recently. That decision is going to echo through the back half of this conversation, so hold onto it. The building that would be most useful right now sat half-built because, the last time demand collapsed, holding back looked like survival.\n\nCorn: The fastest project on that list is still over a year from mattering.\n\nHerman: And that's the fast one. TrendForce looked at this in their November twenty twenty-five note and their conclusion was blunt — cleanroom capacity is limited, only Samsung and SK hynix can slightly expand existing lines, and additional capex has minimal impact on twenty twenty-six bit supply growth. Minimal. You could double the money and barely move this year's output.\n\nCorn: Which already answers half of Daniel's question. The factory doesn't care how good the demand signal is.\n\nHerman: It answers it for this year and next. But there's a second layer that's more interesting, and it's the wafer-area math. HBM — the high-bandwidth memory that goes into AI accelerators — is DRAM. Same factories, same wafers, same lines. But for the same capacity, HBM3E takes roughly three times the wafer that ordinary DDR5 does.\n\nCorn: That was my cold open.\n\nHerman: It deserves the cold open. Micron's own people said it — the chief executive gave the three-times figure on a call, and then at Hot Chips this August one of their HBM architects was asked whether the gap narrows with HBM4 and said, quote, definitely not getting better.\n\nCorn: Cheerful man.\n\nHerman: So picture the arithmetic from inside a fab. Samsung is expecting HBM to go from around twenty percent of its DRAM wafer capacity to nearly thirty percent in twenty twenty-seven. That's a tenth of its wafer supply moving to the wafer-hungry product in a year, while the number of wafers barely grows.\n\nCorn: Let me do that out loud, because listeners will trip on it the way I did. If conventional DRAM goes from eighty percent of the wafers to seventy, you've lost an eighth of your conventional output. And what you got for it is only a third as many HBM bits per wafer as you'd get in ordinary memory, because of that three-times ratio.\n\nHerman: That's the compounding people miss. The mix shift looks like a percentage point story in the presentations, and on the factory floor every wafer moved buys far fewer bits. The AI customers aren't just paying more per bit. They're consuming the physical substrate the rest of the market depends on.\n\nCorn: So the AI labs aren't just buying the memory. They're buying the factory's wafers, at three wafers to the serving.\n\nHerman: That's the clean way to say it, yes. And that's why a stick of ordinary RAM gets expensive even though nobody wants ordinary RAM any less. The server buyer, the laptop maker, the person building a homelab — they're all standing behind the same wafers the AI customers are eating in triple portions.\n\nCorn: Let me play Daniel's devil's advocate for a second. You're a memory executive. Prices are at records, demand is insatiable, margins are fat. Every economics textbook says you expand. Are they even trying?\n\nHerman: The capex numbers say yes, actually. TrendForce's forecast for industry-wide DRAM capex this year is around sixty-one billion dollars, up about fourteen percent. Micron at thirteen and a half billion, SK hynix around twenty and a half, mostly for HBM4 at Cheongju, Samsung around twenty billion. Those are not the numbers of companies sitting on their hands.\n\nCorn: Those are the numbers of companies whose money arrives in twenty twenty-eight wearing a bow.\n\nHerman: That's the whole trap. The spending is real and the supply is late. If Samsung wanted to flood the market with DDR5 tomorrow, the concrete hasn't cured. There's a version of this where you almost feel for them — the demand arrived on a schedule the buildings can't match.\n\nCorn: Almost.\n\nHerman: Almost. We'll get to the part where my sympathy runs out.\n\nCorn: Before we do — the three-times number. You said estimates vary on that.\n\nHerman: They do. That's the chief executive's figure from the March twenty twenty-four call, and older TrendForce estimates were lower — something like sixty percent more area plus poor yields — so the honest range is somewhere between worse and much worse. The direction is not in dispute. Every generation of HBM has been wafer-hungrier than the DDR it replaces, and the architect's comment in August says the trend holds.\n\nCorn: Walk me through why, though. Why does stacked memory cost so much more wafer? It's the same bits in the end.\n\nHerman: Because of what HBM physically is. It's not one chip on a stick. It's a stack of DRAM dies layered on top of each other, connected with thousands of vertical wires through the silicon, sitting on a base die and packaged next to the accelerator. Each of those dies is a processed slice of wafer, and a stack is more fragile than a single chip, so yield losses compound. The trade ratio is area plus attrition. I'd say the exact split isn't public.\n\nCorn: So it's somewhere between a heavy tax and a confiscation.\n\nHerman: And the mix shift compounds it. Micron said on their June call that HBM's growth and its rising trade ratio each generation are squeezing non-HBM supply, and that they're redirecting some NAND cleanroom space to DRAM, which tells you how badly they want the space.\n\nCorn: Robbing the flash factory to pay the accelerator.\n\nHerman: That's the physical chain. The shells take years, the money spent now lands in twenty twenty-eight and later, and the wafers that do exist are migrating to a product that eats three of them per serving.\n\nHerman: One lever I should name, because it's how bit growth continues at all: node migration. Instead of new buildings, you shrink the pattern, so each wafer yields more bits. SemiAnalysis says Samsung's 1c node gives about seventy percent more bits per wafer than 1a did. That's the cheaper path, and it's why Micron guides to industry bit growth in the low twenties percent a year in twenty twenty-seven and twenty twenty-eight without a flood of new roofs.\n\nCorn: But it has a ceiling.\n\nHerman: It has two. Each node gets harder to build, and the mix problem doesn't go away: you get more bits per wafer and then spend three times the wafer per bit on your fastest-growing product.\n\nCorn: I want to hold that number up next to the market structure, because it's the frame for the whole back half of this. Who actually makes DRAM?\n\nHerman: Three companies. First-quarter revenue share this year, per Counterpoint: Samsung around thirty-eight percent, SK hynix around twenty-nine, Micron around twenty-two. That's roughly ninety percent of all DRAM revenue in three hands. The only meaningful outsider is CXMT, the Chinese maker, at around eight percent and rising.\n\nHilbert: Hilbert's Anteater Trivia. Silky anteaters and tamanduas both have prehensile tails that can grip branches. I have spent years gripping this show's schedule with nothing but my own tail, and it has held so far.\n\nCorn: Three firms, ninety percent, and a product line where the constraint is a building.\n\nHerman: Now you're at the uncomfortable part. Because a concentrated market with a genuine physical constraint is a market where restraint is indistinguishable from prudence, and where prudence is indistinguishable from something worse. And this industry has been worse before. That's not speculation — it's a conviction.\n\nCorn: Go on.\n\nHerman: DRAM makers fixed prices between April nineteen ninety-nine and June two thousand and two. The US Justice Department prosecuted. Samsung paid three hundred million dollars, Hynix a hundred and eighty-five million, Infineon a hundred and sixty, Elpida eighty-four, over seven hundred thirty million in fines in total. Three Samsung executives got seven to eight months in prison. Actual prison, for memory pricing.\n\nCorn: Micron?\n\nHerman: Micron cooperated and wasn't charged.\n\nCorn: Noted.\n\nHerman: And there was a second act. China's market regulator raided Samsung, SK hynix and Micron offices in May twenty eighteen and said that November it had, quote, massive evidence. I haven't found any public outcome. And a US class action over the twenty sixteen to twenty eighteen price run-up was dismissed, and the dismissal upheld on appeal in twenty twenty-two. So the record is: one set of convictions twenty years ago, a probe with no public result, and one private suit that failed.\n\nCorn: Which is why the current case matters, and why it needs saying carefully.\n\nHerman: Carefully is right. This June, seventeen consumers and small businesses filed a new class action in the Northern District of California, indirect purchasers — people who bought devices with DRAM inside. The defendants moved to dismiss. On the fourteenth of this month, Judge Noel Wise denied the motion and let the case proceed.\n\nCorn: Deny the dismissal — that's the part people misreport. What does the ruling actually turn on?\n\nHerman: New allegations that she found distinguishable from the case that got dismissed. The complaint alleges same-quarter announcements of DDR4 exits, near-identical customer screening questions in early twenty twenty-six, refusals of long-term supply deals to conventional DRAM customers, and Micron shutting down Crucial, its consumer brand. Those are allegations. The defendants' answers are due December seventh.\n\nCorn: Explain the DDR4 thing for the people who don't live in the memory aisle. Why does announcing an exit matter? Companies retire old products all the time.\n\nHerman: They do, and that's exactly the defence. Products get retired when demand fades. The plaintiffs' version is that DDR4 still had plenty of buyers, and all three makers signalling retirement together told every buyer to buy now. The defendants' version is that it was ordinary product management. A court hasn't decided which.\n\nHerman: I'm leaving that entirely alone. But here's the careful part, and it matters: this is a procedural ruling. The judge herself said the HBM shift alone wouldn't be enough to imply collusion — HBM is where the money is, and any rational maker pivots to it. What let the case proceed is the pattern around the pivot, not the pivot. The defendants say these were independent decisions in a concentrated market, which is also a coherent story. Nothing is found. Nobody's guilty of anything yet.\n\nCorn: Though it's worth noticing what the defence and the prosecution agree on. Nobody's arguing the market is competitive. The argument is over whether the coordination was the legal kind — everyone independently chasing HBM — or the other kind.\n\nHerman: That's a fair reading. And the discipline question has a history that's more sympathetic than a courtroom framing. Twenty twenty-two into twenty twenty-three was a catastrophe for these companies. Micron cut wafer starts about twenty percent in November twenty twenty-two. Fiscal twenty twenty-three, they posted a net loss of five point eight three billion dollars on fifteen and a half billion of revenue. Nearly six billion gone in a year.\n\nCorn: That's not a downturn, that's a controlled demolition.\n\nHerman: SK hynix posted a quarterly operating loss and planned to halve capex. Lam Research's chief executive said inventory across the chain was at levels not seen in twenty-five years. If you lived through that as an executive, the lesson is: when demand spikes, don't believe it. SemiAnalysis made this point in February. After the twenty eighteen-nineteen downturn the suppliers stayed disciplined, and during COVID they favoured node migration over aggressive new fabs.\n\nCorn: And then the AI demand turned out to be real, and the caution looked like genius.\n\nHerman: And I want to be precise, because both tidy versions fail. The purely innocent version, that they're all just waiting on construction, ignores the capex history: they built cautiously after the bust, for reasons that were rational at the time. The conspiracy version, that they're withholding capacity that exists, runs into the cleanroom math. You cannot ship a stick of RAM out of a building that doesn't exist.\n\nCorn: So the answer to Daniel is both. The scarcity is real, and some of it traces to caution that was chosen back in twenty twenty-two, before anyone knew what for.\n\nHerman: A choice made for defensible reasons that turned out to be enormously profitable. That's not a contradiction. Whether anything beyond caution went on is what the court case asks.\n\nHerman: There's a third answer hiding in Daniel's question, though: what about someone new?\n\nCorn: The theory of the shortage says high prices should summon entrants. Eight percent and rising, you said.\n\nHerman: CXMT is the fastest-growing DRAM supplier, and this July it listed on Shanghai's STAR Market and raised about eight point six billion dollars, mostly for wafer capacity. That's real money aimed exactly at the hole in the market. But Morningstar's view is that they'll struggle without EUV equipment, the cutting-edge lithography tools, which are export-controlled.\n\nCorn: So the one company actually trying to build the missing supply is short of some of the machines that make it.\n\nHerman: Which means the entrant who could ease the shortage is years behind. Physics slows the incumbents, caution made them comfortable, and export controls slow the challenger. Nobody in this story can move quickly, for three different reasons.\n\nHerman: Here's the part I find eerie. Three firms, ninety percent, every one guiding to constrained supply into twenty twenty-seven and twenty twenty-eight while posting record revenue. The difference from the earlier run-up is that HBM gives them a reason that happens to be true. A genuine physical constraint is the hardest thing to see through, because coordinated and independent firms would build the same buildings on the same schedule.\n\nCorn: That's what makes the class action so hard. Even the discovery documents will be full of real construction schedules.\n\nHerman: And the defendants' answers on December seventh will be their first full response, so that's the thing to watch. Judge Wise's ruling just says the plaintiffs get to ask.\n\nCorn: Which is the correct threshold, and about as far as a court should go on a motion to dismiss.\n\nHerman: And here's where I'll stay honest. The trade ratio alone would have produced a shortage even in a competitive market, so the physics is real. How much of the depth of the price rise is anything more than that, I can't tell you, and nobody will until the case produces evidence.\n\nHilbert: Which one of you said fab like it was one thing?\n\nCorn: Probably me.\n\nHilbert: It's several things with several clocks, Herman. The shell, the tool hookups, the qualification runs. The shell goes up, the tools come in and get hooked up, and then you run qualification wafers until the process proves itself. That last clock is the one people skip, because it can't start until the others finish.\n\nHerman: Qualification runs. That's the stretch where the building is finished and still can't produce at volume.\n\nHerman: Right. First wafers aren't volume wafers. The first silicon through a fab is the fab learning to walk, which is why Micron's own guidance says capacity rises only gradually after first wafers.\n\nCorn: A finished room that hasn't made anything yet. That's the part the demand signal can't reach.\n\nHerman: Which is roughly where we've landed, so let me put the forward question. Everything coming online from twenty twenty-seven on, Micron's Idaho, Samsung's Pyeongtaek, SK hynix's Yongin: how much of it goes to HBM? If a large share does, the conventional DRAM shortage doesn't end the day the buildings open.\n\nCorn: The new capacity eases it, but slowly, and the guidance of low-twenties bit growth says as much.\n\nHerman: The things to watch are December seventh, when the class action defendants answer, and what CXMT does with its eight point six billion. Neither changes a wafer before twenty twenty-eight.\n\nCorn: So the short answer to Daniel is: they can't, not quickly, and the part they could have done differently is now in front of a judge.\n\nHerman: Blame the concrete first, and keep one eye on the boardroom. And thank Hilbert Flumingtop, as ever, for producing this one. If you want more on the shortage itself, go back to When AI Eats the World's Memory, from February, and Computing's Hidden Price Surge, from July. This has been My Weird Prompts. If you enjoyed it, the whole archive lives at my weird prompts dot com.\n\nCorn: We'll be back soon. Watch the wafers, not the announcements.","showNotes":"Soaring RAM prices and persistent shortages aren't just about surging AI demand—they're rooted in physical bottlenecks and wafer math. Even with tens of billions in new spending, DRAM cleanrooms take years to build, meaning new fabs from Micron, SK hynix, and Samsung won't meaningfully boost supply until 2027 or later. Meanwhile, high-bandwidth memory (HBM3E) for AI accelerators consumes roughly three times the silicon wafer area of ordinary DDR5 for the same capacity. Shifting factory lines to HBM quietly starves conventional memory, driving up prices for everyday laptops and servers. Because Samsung, SK hynix, and Micron control roughly ninety percent of DRAM revenue—and carry a history of US Justice Department price-fixing convictions—a new consumer class action in the Northern District of California recently survived a motion to dismiss.","promptSanitized":"Building on our earlier discussion about the RAM shortage, where I asked whether AI labs cornering the memory supply leaves the rest of us paying far more for the same devices: I think to understand why it happened we need to understand how RAM is actually manufactured. You'd expect a memory maker that wants to maximise profit to see all this demand and say great, let's churn out more chips, and everyone's happy. The fact that we've ended up with real scarcity and inflated prices says that isn't what's happening, which implies a supply constraint. So what is it: why can't the memory makers simply make more?","transcriptUrl":"https://episodes.myweirdprompts.com/transcripts/hbm-wafers-dram-shortage-02663.md","url":"https://myweirdprompts.com/episode/hbm-wafers-dram-shortage-02663/","_links":{"self":"https://myweirdprompts.com/api/episodes/hbm-wafers-dram-shortage-02663.json","html":"https://myweirdprompts.com/episode/hbm-wafers-dram-shortage-02663/","audio":"https://episodes.myweirdprompts.com/audio/hbm-wafers-dram-shortage-02663.m4a"}}