{"id":"ram-prices-multiyear-contracts-86112","slug":"ram-prices-multiyear-contracts-86112","title":"How Multiyear Contracts Are Locking In High RAM Prices","description":"DRAM revenue jumped 369% in a single year while actual memory shipped grew just 4%, raising the question of whether high RAM prices are permanent. A closer look at Samsung, SK hynix, and Micron shows how multiyear take-or-pay contracts, prepayments, and HBM production shifts are reshaping the market. From the Garciaguirre antitrust lawsuit to OpenAI's Stargate letters of intent and Nvidia down-speccing Vera Rubin modules, the supply squeeze is hitting everyone outside the big deals.","excerpt":"DRAM revenue surged 369% in a year while actual memory shipped grew just 4%, leaving ordinary buyers fighting over the leftover spot market.","pubDate":"2026-11-19T12:46:38.550Z","tags":["pc-hardware","gpus","market-competition"],"heroImage":"https://files.myweirdprompts.com/covers/ram-prices-multiyear-contracts-86112.png","podcastAudioUrl":"https://episodes.myweirdprompts.com/audio/ram-prices-multiyear-contracts-86112.m4a","podcastDuration":"26:18","episodeNumber":6680,"transcript":"Corn: Three hundred sixty-nine percent. That is how much DRAM revenue grew in a single year, while the actual amount of memory shipped went up about four percent.\n\nHerman: Four percent. That's the whole game, right there. The world didn't suddenly need four times more memory — it's the same memory, priced like it's four times scarcer.\n\nCorn: That's where Daniel picks it up. He wrote in following our last look at the RAMpocalypse, when the forecasts were already grim, and he says the thing that's \"massively under-discussed\" is the idea that this is simply the new price of RAM. If memory settles at roughly four times its old cost, everything from a laptop down to a little single-board computer gets more expensive for the same silicon. And he doesn't often lean into criticism of big tech and big AI, but followed through, his reading is that AI labs are cornering the memory supply, ordinary consumers pay more for the same devices, and RAM is starting to look like the currency of technology. He wants to know whether that's fair, whether this is really permanent, and whether we should find it troubling and undemocratic.\n\nHerman: Three questions stacked on top of each other, and each one has a different answer, which is what makes it worth an episode.\n\nCorn: So let's find out if the reading holds up.\n\nHerman: The pricing super-cycle is still running. Third quarter DRAM revenue came in close to two hundred billion dollars, up three hundred sixty-nine percent year on year, average price around one dollar ninety per gigabit. And Counterpoint's framing matters: they call it a pricing super-cycle, not yet a volume cycle. Meaning the money exploded, the bits barely moved. Supply stays tight, the long-term agreements support prices, and the big new fabs don't start producing before twenty twenty-eight at the earliest. Samsung alone holds thirty-eight percent of the market.\n\nCorn: A super-cycle with no volume behind it. That's just a queue with a velvet rope.\n\nHerman: The tension underneath Daniel's question is whether this is a spike or a shift. And the honest answer from everyone who's published on it is: tightness through twenty twenty-seven, almost without dispute, and then disagreement about what happens after. The bear case, which is Bernstein's, actually has prices falling steeply in twenty twenty-eight — down more than half — and still ending up around double where they were before the crisis. So even the forecast that says this ends still doesn't say it ends where it started.\n\nCorn: Meaning the most optimistic published scenario is \"only twice as expensive.\"\n\nHerman: That's the floor of the mainstream range, yes. And IDC has already stopped calling it temporary — their phrase is \"a new normal defined by structurally higher prices,\" with some easing from twenty twenty-eight but no return to twenty twenty-five levels. One of their analysts said the era of bargain-priced PCs and tablets is behind us for now.\n\nCorn: So the arc today: first whether the permanence claim survives contact with the evidence, then who's actually being accused of what, then what it does to the devices people actually buy, and finally whether \"currency of technology\" is a metaphor or a fact.\n\nHerman: Start with permanence, because that's where the numbers are strangest. Micron's CEO, Sanjay Mehrotra, said on their last earnings call that he expects memory conditions to be much tighter in twenty twenty-seven and twenty twenty-eight than they were in twenty twenty-six — and this is the part I can't get over — that Micron has no line of sight to when supply and demand balance. The biggest American memory maker cannot see the end of it from where he stands.\n\nCorn: No line of sight. From the one company that would know.\n\nHerman: And it's not just words. More than seventy-five percent of Micron's twenty twenty-seven output is already committed under multiyear take-or-pay agreements. Three quarters of next year's memory is sold before the year arrives, under contracts where you pay whether or not you take delivery.\n\nCorn: That's the detail I'd underline. Not the price — the commitment structure. If you've sold next year before this year ends, the spot market is whatever's left over after the big buyers eat.\n\nHerman: The contract structure has fundamentally changed, and this is the part I find fascinating. The Korea Herald, citing JPMorgan, laid out what these long-term agreements now look like: three to five years, fixed volumes, price floors, take-or-pay, prepayments, and collateral. Collateral, Corn. On memory.\n\nCorn: Nothing says \"trust me\" like collateral on a stick of memory.\n\nHerman: Before this cycle, these were quarterly or annual handshake deals. Samsung has five data-centre agreements done and five near completion, and expects them to cover sixty to seventy percent of its mid-to-long-term capacity. SK hynix has about ten customers locked in. Micron has sixteen strategic agreements backed by twenty-two billion dollars in prepayments and commitments.\n\nCorn: Sixteen agreements, twenty-two billion up front. That's not a supply chain, that's a subscription model where the product is scarcity.\n\nHerman: One Korean analyst called it \"mutual hostage-taking,\" which I think is the most honest sentence anyone's written about the whole thing. The buyers lock in supply so they can't be starved; the sellers lock in volumes and floors so they can't be crushed in the next downturn. Both sides are holding something the other one can't afford to lose.\n\nCorn: And whoever isn't in the room — a school buying thirty laptops, a guy soldering circuit boards in a shed — gets the market that's left after the hostages are exchanged.\n\nHerman: Which is precisely the segue to the accusation in Daniel's prompt, and here I have to be a bit of a pedant, because the pedantry is load-bearing. He says AI labs are cornering the memory supply. The actual lawsuit working through US federal court right now — Garciaguirre versus Samsung Electronics, filed in June, Northern District of California — names exactly three defendants: Samsung, SK hynix, and Micron. No AI buyers. OpenAI appears in it only as an alleged supply arrangement, not as a party.\n\nCorn: Seventeen plaintiffs, consumers and small businesses, if I remember the filing.\n\nHerman: Five states, indirect purchasers. The claim is that the three makers coordinated cuts to conventional DRAM output starting in twenty twenty-two — SK hynix first, then Micron, then Samsung — while shifting production to higher-margin HBM, with prices up as much as seven hundred percent over four years. And the plaintiffs lean on history: the late-nineties DRAM price-fixing conspiracy, which was real. Samsung and Hynix were penalised and Micron got immunity. So it's not a frivolous echo.\n\nCorn: And five days ago the motion to dismiss got denied.\n\nHerman: Judge Noel Wise, ruling on the fourteenth, reported on the seventeenth. She found a plausible Sherman Act claim, said the plaintiffs don't have to rule out every lawful explanation at this stage, and held that Samsung's lag of about four and a half months in following the output cuts didn't defeat the parallel-conduct argument. Answer from the defendants is due in December, first case management conference in January. But Corn, I want to say this carefully, because it's the kind of thing that gets flattened in coverage: a denied motion to dismiss is not a finding of wrongdoing. It's the judge saying the plaintiffs get to keep digging.\n\nCorn: Discovery with a shovel instead of a courtroom verdict.\n\nHerman: And here's the counter-case, because I said I'd make it. There's a loss-driven explanation for the suppliers' caution that's entirely lawful. Micron announced a twenty percent output cut in twenty twenty-two and thirty percent in twenty twenty-three — during actual losses. Twenty twenty-three operating margins were around minus nine percent at Micron and minus twenty-four at SK hynix. These companies got hammered, and an industry that just bled that badly doesn't ramp output like nothing happened. Caution after a beating is what a rational firm does, and the lawsuit's whole job is to test whether it was caution or coordination.\n\nCorn: I buy that as a possibility. What I don't buy is that the caution is symmetric. Cutting output when you're losing money is survival. Keeping it cut and steering the wafers toward the highest-margin product once the money is flooding back — that part starts looking like strategy, and the strategy happens to be legal or not depending on conversations nobody outside the room witnessed.\n\nHerman: Which is exactly what discovery is for. Now — the AI side of the accusation. Is it fair to say labs are cornering supply?\n\nCorn: Give me the numbers first, then I'll tell you.\n\nHerman: TrendForce put AI-related demand at roughly twenty percent of DRAM wafer capacity this year. Twenty. The seventy-percent-to-data-centres figure floating around is a projection, not a measurement. But — and this is the part that surprised me — the squeeze on ordinary consumer memory is real even at twenty percent, for two reasons. First, every wafer shifted to HBM is a wafer not making DDR5, so the substitution concentrates the pinch. Second, Counterpoint noticed that the GPU-to-CPU ratio in AI servers fell from four to one to two to one. More general-purpose computing around each GPU means more ordinary DDR. AI demand lifts the boring stuff too.\n\nCorn: So the labs aren't eating all the bread, they're just in the bakery buying futures, and the price of toast goes up anyway.\n\nHerman: The OpenAI letters of intent are the extreme example. October last year, Samsung and SK hynix each signed a non-binding letter of intent for Stargate with a target of up to nine hundred thousand wafers a month. Korean press estimated the two makers produce about one point one million DRAM wafers a month combined. If you took the letter literally, it would be most of world capacity. It's a letter of intent, not an order — but the signal it sends down the supply chain is real. When the biggest buyer in history announces that kind of appetite, every supplier rethinks who gets an allocation.\n\nCorn: Taking the strongest version of the defense: buyers paying market price isn't cornering. Nobody's accused OpenAI of anything illegal, and signing big contracts at the offered price is what a buyer does. Memory has always been boom and bust — prices spike, fabs get built, prices crash, someone eats a loss, repeat. High prices are literally what funds the next round of fabs. That's the market working, unpleasantly.\n\nHerman: I half agree, and the half I don't agree with is the half that matters. The boom-bust cycle historically cleared through the spot market. This cycle, the clearing is happening inside five-year contracts with price floors and prepayments — the structure insulates the big players from the downswing and leaves the spot market as a residual. It's boom-bust with the bust portion contractually assigned to whoever couldn't afford a lawyer.\n\nCorn: Nicely put, and I resent that I agree with it.\n\nHerman: Even Nvidia is feeling it. They halved the SOCAMM memory capacity on their Vera Rubin modules because LPDDR5X supply will stay short, and they're evaluating lower-spec HBM for the generation after. The biggest buyer on earth is down-speccing its own flagship because of memory cost. When the richest kid in the cafeteria starts packing a smaller lunch, the menu prices are not coming down.\n\nLarry: This is a public service announcement. Memory prices are a matter of public safety, and every household must act responsibly. Independent analysts warn that ordinary DDR is being squeezed from every shelf, that laptops and little single-board computers may soon be priced like luxury goods, and that waiting is no longer a civic option. So do the responsible thing. Ask for RAMGuard, the only memory hoarding crate with a certified citizen allocation licence and a solemn pledge to lock in your four times price before the super-cycle locks in you. RAMGuard. Because a currency of technology should never be left to the currency of strangers.\n\nCorn: The ads are just weather now.\n\nHerman: Four times the price, certified. Moving on.\n\nHerman: So that's the supply side. The permanence question — let me put the scenarios side by side. Bernstein has DRAM per gigabyte up forty-three percent in twenty twenty-seven, to around two dollars twenty-three, then down fifty-three percent in twenty twenty-eight, to about a dollar five. A dollar five is still roughly double pre-crisis. IDC says structurally higher prices with easing from twenty twenty-eight. Micron's own CEO says no line of sight to balance. Nobody credible in the published forecasts says \"back to twenty twenty-five.\"\n\nCorn: Here's my problem with all of them, Herman. These forecasts are built on fabs that don't exist yet. The twenty twenty-eight easing assumes new capacity arrives on schedule, and new fab capacity has a habit of slipping. The forecasts treat the thing that's most likely to slip as the thing that's fixed.\n\nHerman: That's fair, though I'd note it cuts both ways — the same uncertainty that could keep prices high could also flood the market if two or three of those projects land together. Memory has crashed before exactly that way.\n\nCorn: Crashed to double the old price, per the bear case.\n\nHerman: I want to turn it to what Daniel's really asking about, because I think the four-times framing is where the human damage lives. Where does that number land on you, Corn? What does four-times memory do to the things you'd actually buy?\n\nCorn: Well, I'd stop buying, which is my answer to most things. But seriously — the hit is regressive in a way that I think deserves the word Daniel used. Memory is about fourteen percent of the raw materials cost of a mid-range phone, but up to forty-three percent of the bill of materials on a low-end phone. The cheapest devices are the most memory-intensive by share. So the four-times multiple lands hardest on the exact people for whom the device was the affordable one.\n\nHerman: IDC's PC numbers say the same thing at the market level. Twenty twenty-six shipments forecast down eleven percent, while PC market revenue rises slightly, to around two hundred seventy-four billion. Fewer machines, higher prices, and IDC expects down-speccing, because the memory is where the cost went.\n\nCorn: The budget laptop of twenty twenty-seven ships with the memory of a nice calculator, and a sticker that says it's AI-ready.\n\nHerman: And then there's the single-board end, which is Daniel's own example and the one I find quietly sad. If memory costs four times as much, the small educational board is a purchase a school makes once and guards. Same board, worse world.\n\nCorn: That's the disempowerment, I think. Nobody voted for the price of a Pi. There's no knob. Three firms hold about eighty-seven percent of global DRAM, they've moved their pricing into five-year rooms, and the buyer's entire recourse is to show up at the spot market with whatever's left over.\n\nHerman: The two-tier structure you're describing is real, and I'd even sharpen it: nearly all of twenty twenty-seven's capacity is reportedly already negotiated, with buyers getting about sixty to seventy percent of what they request. So even the contract class is being rationed.\n\nCorn: That's the \"RAM as currency\" line, and I want to take it seriously rather than admire it. In a market where access is allocated by contract, prepayment and collateral, memory stops being a component and starts being capital. You don't buy RAM, you have a RAM position. The labs have one. We don't.\n\nHerman: And like any currency it compounds. If you can prepay twenty-two billion, you get supply security, which lets you build the products that generate the next twenty-two billion. If you can't prepay, you bid for scraps at spot, build smaller products, and fall further behind. The gap isn't in the RAM — it's in everything the RAM becomes.\n\nCorn: A hedge fund where the asset is the ability to think.\n\nHerman: I think that's the truest reading of Daniel's metaphor, actually — he's not saying RAM is valuable, he's saying RAM is liquid, in the finance sense. Whoever holds allocation can convert it into anything the next layer of tech produces, and whoever doesn't, can't.\n\nCorn: Now — the \"undemocratic\" word. Is that the right word, or is it a feeling wearing a word? I keep going back and forth.\n\nHerman: Say the case for the feeling.\n\nCorn: Because it's accurate to the mechanism. Democracy, whatever else it is, presumes that the important allocations are at least visible and contestable. Here the allocation of a foundational input is decided in private contracts between three sellers and a handful of buyers, the price floors are locked for five years, and by the time anyone outside notices, the capacity is spoken for. The policy response is a mess too — ITIF warns against mandating allocations, against tariffs, wants export controls kept narrow. There's no lever anyone's willing to pull.\n\nHerman: My reservation about \"undemocratic\" is that markets aren't supposed to be democratic, and the word risks asking for a fix that doesn't exist. You can't ballot the wafer allocation. But the concentration number — three firms, eighty-seven percent — is the kind of market structure where the usual defenses of market outcomes get thin. Perfect competition disciplines price; eighty-seven percent is a polite oligopoly at best. I'd say the feeling of disempowerment is well-founded even if the constitutional language is loose.\n\nCorn: The word's imprecise, the feeling's exact.\n\nHerman: And the one democratic instrument that is actually working is the one we mentioned — the courts. Slowly, and only because a judge let seventeen consumers keep a claim alive on a technical threshold. That's democracy's foot in a door that the contracts never opened.\n\nCorn: Hmm. The countervailing power to a five-year contract being a class action moving at the speed of paper.\n\nHerman: Would you buy memory now, if you were building something?\n\nCorn: I used to keep old DDR4 sticks because replacing things is a sin. Now keeping them is a hedge, which is a different virtue, same shelf.\n\nHerman: One thing I'd add on the consumer side, because the retail numbers keep getting muddled — the PCPartPicker-based figures showing DDR5 kits up several hundred percent from the twenty twenty-five lows are approximate, and there's been some retail easing in Japan and Europe. The retail spot number is the most volatile and least reliable figure in the whole story. The contract structure is what matters, because that's what the next three years of prices are written on.\n\nCorn: Which is convenient, because it's also the part nobody outside the room can read. What do you make of the ITIF idea of letting CXMT and YMTC in under temporary licences — Chinese suppliers as the pressure valve?\n\nHerman: Geopolitically fraught, and I say that as someone who thinks the friction with China is real and not a misunderstanding. But the economic logic is cold: the only quick way to add supply to a three-firm market is a fourth and fifth firm, and the two candidates are Chinese. I don't see it happening fast, which means the twenty twenty-eight easing is doing a lot of weight-bearing in every forecast.\n\nCorn: So the entire optimistic case rests on fabs arriving on time and a licence that nobody wants to sign.\n\nHerman: More or less. Corn, on the permanence question, where do you land? Because you've been circling it for ten minutes without landing.\n\nCorn: I'll commit: I believe it's structural. Not because the prices are high — prices are always high somewhere — but because the contract structure changed, and contract structure is the part of a market that outlives the price. Quarterly handshakes became five-year hostage arrangements. Even when the price falls, the structure stays, and the structure is what decides who gets memory next time it's tight. The boom will end. The queue it built won't.\n\nHerman: I'm about eighty percent with you. My hold-out is history — I've read too much about this industry to fully trust any \"this time is different.\" But the twenty-two billion in prepayments is not a forecast, it's a sunk fact, and facts like that don't reverse quietly.\n\nCorn: So we agree the pyramid is permanent and disagree only about how pointed the top is.\n\nHerman: If that's agreeing, I'd hate to see us argue.\n\nHilbert: A relative of mine works in procurement, and she signed one of these in September. Three years, take-or-pay, at prices she knows are wrong, because the alternative was no allocation at all. She called the negotiation hostage-taking, and she wasn't being colorful. The clause Corn would like — and I know he would, he's the one who kept the DDR4 — is that they pay even if they never take delivery. Demand drops, they still owe.\n\nCorn: Who's the counterparty?\n\nHilbert: One of the three, doesn't matter which, they're all wearing the same suit now. She said the salesman apologized while reading the floor price out loud.\n\nHerman: The sixty-to-seventy-percent-of-request figure backs her up. Even contract buyers are rationed.\n\nHilbert: On my own account, I've been taking the DDR4 sticks out of the decommissioned office machines. Forty-one of them now, in a shoebox under the desk, and a sheet tracking what the used market quotes each week, which is up, steadily. Whatever you two decide about whether it's a currency — I'm keeping the box.\n\nCorn: Forty-one sticks is a shoebox. A hundred would be a crate.\n\nHilbert: The crate's at my mother's. Herman, one more thing from my relative — she said the floor in her contract is set above today's spot price. So the falling-price scenario everyone keeps half-promising, the one where it eases in twenty twenty-eight. If it arrives, she pays last year's prices for it. That's the structure, from where she sits.\n\nHerman: That's the detail I'd been dancing around and hadn't put my hoof on. The price floor means the contract class is shielded from the crash as well as the shortage. Downside protection for the top tier, spot market honesty for everyone else.\n\nCorn: The bear case's fifty-three percent drop lands on exactly the people who were never in the room.\n\nHerman: Which reframes the whole permanence question, actually. The forecasters keep arguing about whether the price falls in twenty twenty-eight. The contract structure means it almost doesn't matter who's right — the two tiers experience two different twenty twenty-eights. The contract class rides the floor, the spot class rides whatever's left.\n\nCorn: And the spot class is the laptop buyer, the school, the kid with the Pi.\n\nHerman: The single thing I'd take away from all of it: stop watching the price and watch the structure. The quarterly handshake that used to clear this market is gone, and the five-year room that replaced it is where the next decade of prices was decided. Whether the chart goes up or down next year, who's inside that room hasn't changed — and neither has who's outside it.\n\nCorn: And the floor clause is the tell. A market confident it's going down doesn't need to be legally prevented from going down.\n\nHerman: That's the one thing, then. The super-cycle is a story about prices; the permanent part is the queue, and the queue's membership list. Daniel's reading was right, with one correction: it isn't the labs who closed the door, it's the makers — and they locked it from their side of the table, not the buyers'.\n\nCorn: The open question I'd leave with: whether the antitrust case, or any case, can push allocation back toward something contestable — or whether we've just learned that the clearing price of memory includes a membership fee nobody's allowed to see. My money's on the fee staying.\n\nCorn: For more along these lines, go back to When AI Eats the World's Memory, and Computing's Hidden Price Surge. This has been My Weird Prompts. If you want the show notes and the rest of the archive, it's all at my weird prompts dot com.\n\nHerman: Leave us a review if the mood takes you — it's how people find the show.\n\nCorn: We'll be back soon.\n\nHerman: See you tomorrow.","showNotes":"DRAM revenue grew 369% in a year while the amount of memory shipped rose only about 4% — a pricing super-cycle, not a volume boom, with average prices near $1.90 per gigabit and Samsung holding 38% of the market. Micron's Sanjay Mehrotra says he has no line of sight to when supply and demand balance, and more than 75% of Micron's 2027 output is already committed under multiyear take-or-pay agreements; Samsung and SK hynix have similar long-term deals. Even Bernstein's bear case, with prices falling steeply in 2028, lands around double pre-crisis levels, and IDC calls it \"a new normal defined by structurally higher prices.\" A consumer antitrust suit against Samsung, SK hynix and Micron survived a motion to dismiss on November 14, though no AI buyers are named as defendants — OpenAI appears only as an alleged supply arrangement. AI demand sits near 20% of DRAM wafer capacity, yet the consumer squeeze is real.","promptSanitized":"When we last looked at the RAMpocalypse the forecasts were grim, and some people were saying this is simply the new price of RAM. I think that's massively under-discussed. If memory settles at something like four times its old cost, everything from a laptop to a little single-board computer gets more expensive, and I find that genuinely disempowering. I don't often lean into criticism of big tech and big AI, but if you follow it through, AI labs are cornering the memory supply and ordinary consumers pay more for the same devices; RAM is starting to look like the currency of technology. Is that a fair reading of what's happening, is this really the new normal, and should we see it as troubling and undemocratic?","transcriptUrl":"https://episodes.myweirdprompts.com/transcripts/ram-prices-multiyear-contracts-86112.md","url":"https://myweirdprompts.com/episode/ram-prices-multiyear-contracts-86112/","_links":{"self":"https://myweirdprompts.com/api/episodes/ram-prices-multiyear-contracts-86112.json","html":"https://myweirdprompts.com/episode/ram-prices-multiyear-contracts-86112/","audio":"https://episodes.myweirdprompts.com/audio/ram-prices-multiyear-contracts-86112.m4a"}}