Latest / Logosofia no Dia a Dia / Computing's Hidden Price Surge: DRAM, NAND & GPUs

Transcript

Computing's Hidden Price Surge: DRAM, NAND & GPUs · Logosofia no Dia a Dia

• [Prompt] custom: We have seen some remarkable and dramatic shifts in hardware pricing over the course of the last cou | Hosts: corn, herman [Response] Corn: Daniel sent us this one — he's been watching hardware prices do something genuinely strange over the last couple of years. DRAM spiked, we talked about that. But NAND storage has also surged, and GPUs are pinned to AI demand in a way that's starting to look structural, not cyclical. His question is whether anyone's built a unified hardware price index that tracks the real cost of a desktop or server over time and across regions. Because the deeper question underneath is: is computing actually becoming more expensive? Herman: The short answer is — nobody's built that index. Which is wild, because the numbers are screaming for one. Let me give you the concrete example that makes this real. A two-terabyte NVMe SSD, something you'd toss in a mid-range build without thinking twice — that cost about a hundred and twenty dollars in early twenty twenty-four. That same drive today? A hundred and eighty dollars. That's a fifty percent increase. And it's not because someone forgot how to make NAND. It's because enterprise SSD demand for AI training clusters is vacuuming up supply. Corn: My storage upgrade costs fifty percent more because someone's training a model to generate better cat pictures. Herman: That's the charitable read. The less charitable read is that your storage upgrade costs fifty percent more because hyperscalers are building training clusters the size of small towns, and they're buying NAND by the shipping container. And here's the thing — it's not just NAND. DRAM prices are still climbing through Q3 of this year. Tom's Hardware has been tracking this, and DDR5 has risen somewhere in the neighborhood of twenty to thirty percent since early twenty twenty-five. GPU prices are their own category of pain because TSMC's advanced packaging capacity, the CoWoS stuff that makes high-end chips possible, is almost entirely booked for AI accelerators. Consumer cards get the leftovers. Corn: We've got three surges happening simultaneously. DRAM, NAND, GPUs. All pointing the same direction. And from what I'm reading, Micron just signed a hundred billion dollars in long-term supply agreements and then told the press they have, quote, no idea when the RAM crisis will end. Which is not the kind of thing a company says if they think this is a blip. Herman: That's the Micron quote that stuck with me. They're not saying "hey, give us six months and we'll sort it out." They're saying we've locked in a hundred billion dollars of business and we still can't see the other side of this. That's a structural statement. And then you layer on Tim Cook at Apple, who came out and said AI-driven price increases are unavoidable, and the situation has become unsustainable. When the CEO of the company with arguably the most aggressive supply chain management on earth says "unsustainable," you pay attention. Corn: Unavoidable and unsustainable is a fascinating pair of words to use together. It's like saying "the house is on fire and also we've decided fire is now a permanent feature of houses. Herman: That's exactly the tension. And it's why Daniel's question about a hardware price index is so sharp. We have the Consumer Price Index, we have the Producer Price Index, we have GPU-specific trackers that enthusiasts maintain — but nobody has built a single number that says: here's what it costs to buy a complete compute node today versus two years ago, averaged across components and weighted by what people actually build. We don't have that. And without it, we're all just trading anecdotes about our last Newegg receipt. Corn: Which is a problem, because the anecdotes are getting alarming. If you built a mid-range desktop in early twenty twenty-four, you might have spent around seven hundred and fifty dollars on the core trio — memory, storage, and GPU. Today that same build is running closer to a thousand. That's a thirty-two percent increase in two years. That's not inflation. That's AI eating your supply chain. Herman: It's not evenly distributed. The GPU piece is the most visible, but the DRAM and NAND surges are quieter and in some ways more consequential because they hit every build. You can't skip RAM. You can't decide to run your operating system on hope. So when DDR5 prices climb twenty to thirty percent and NVMe prices jump fifty percent, that's a tax on every single machine that gets assembled, from the budget office PC to the high-end workstation. Corn: Where we're going with this episode, and what Daniel's prompt opens up, is a look at the mechanics behind all three surges, why they're happening at the same time, what it means if you're building or buying, and whether anyone's actually trying to build the index that would make sense of this. Because right now, we're flying blind. Herman: The central question hanging over all of it — is this a spike we ride out, or is cheap computing becoming a historical artifact? Micron's hundred billion dollar bet says they think it's structural. But Tom's Hardware is also reporting that consumer affordability limits are starting to cool DRAM price growth. So we've got two forces pulling in opposite directions, and the next twelve months are going to tell us which one wins. Corn: Let's find out which one's got the better grip on reality. Herman: To frame the question Daniel's really asking — it's not just "are things more expensive." It's whether computing as a category is undergoing a structural cost shift that changes who can afford to build what. And that question hits differently depending on whether you're a home builder staring at a parts list or a procurement officer ordering ten thousand servers. Corn: The enthusiast feels it in the checkout cart. The hyperscaler feels it in the margin. But both are feeling it. And the weird thing is, nobody's built a single number that tracks this across the board. We've got the Consumer Price Index, which bundles computers into a broader basket and smooths out all the interesting detail. We've got the Producer Price Index for semiconductors, which is too narrow. And we've got community efforts like the GPU price trackers on Reddit and Tom's Hardware, but those don't include DRAM or NAND. Herman: That's the gap. A proper hardware price index would need to weight components by what people actually put in machines. A typical mid-range build is something like thirty-two gigs of DDR5, a two-terabyte NVMe drive, and a current-generation GPU. Track those three prices over time, average them with reasonable weights, and you'd have a number that actually means something. But nobody's maintaining that as a formal index. Not PC Part Picker, not Tom's Hardware, not the Bureau of Labor Statistics. Corn: Which is strange, because the data exists. The prices are public. The sales are tracked. It's not a measurement problem — it's an aggregation problem. And I think part of the reason nobody's done it is that the answer would be uncomfortable. If you published a number showing that the cost of a standard compute node has risen thirty-two percent in two years, that becomes a story. And not a story the industry wants told. Herman: There's also a genuine complexity here that makes it harder than it sounds. Which components do you include? Do you track the exact same SKU over time, or do you adjust for generational improvements? A two-terabyte NVMe drive today is faster than the same capacity two years ago — do you factor that in, or do you just track raw dollars per gigabyte? Corn: The raw dollars approach is simpler and probably more honest. If the question is "what does it cost to get into a functional machine," then a fifty percent price hike on the same capacity drive is real, regardless of whether the new one's faster. Most people don't need the speed bump. They need the bytes. Herman: That gets to why Daniel's question matters beyond the enthusiast community. If the cost floor for a functional computer is rising, that has implications for education, for small businesses, for anyone whose budget is fixed. The era where you could build a capable machine for six hundred dollars might be ending. Not because the technology got worse — it got better — but because the supply chain got captured by a higher bidder. Herman: Let's get into the mechanics of why all three are surging at once. And the place to start is DRAM, because that's where the most dramatic numbers are. Micron's hundred billion dollar supply agreement isn't just a big round number — it represents a fundamental reallocation of who gets wafer capacity. The specific mechanism is HBM3e, the high-bandwidth memory that sits right next to AI accelerators. A single HBM3e stack consumes roughly three times the silicon area of an equivalent capacity in standard DDR5. And when you're NVIDIA shipping a hundred thousand H100s or B200s per quarter, each of those cards needs six or eight HBM stacks. That's fab capacity that used to produce the DDR5 sticks you buy on Amazon. Corn: It's not just that AI is buying a lot of memory. It's that AI memory is physically hungrier per gigabyte than consumer memory. Herman: And that's the cannibalization mechanism in a nutshell. The same fabs, the same advanced nodes, the same packaging lines serve both markets. When Micron or SK Hynix allocates wafer starts, they have to choose — do we run HBM3e for a hyperscaler contract worth billions, or do we run DDR5 for the consumer channel? The margin on HBM is higher, the contracts are longer, and the demand visibility is better. The choice makes itself. Tom's Hardware has been tracking the result — DDR5 prices have climbed twenty to thirty percent since early twenty twenty-five, and they're still rising through Q3 of this year. Corn: Micron's "no end in sight" comment suggests they don't see a rebalancing coming anytime soon. If you've locked in a hundred billion dollars of supply agreements, you've essentially taken a multi-year bet that AI demand will keep absorbing everything you can produce. Herman: And that bet cascades into NAND through a similar but distinct mechanism. The NVMe price surge isn't about silicon area competition the way DRAM is — it's about volume. Enterprise SSDs for AI training clusters and inference servers are being ordered in quantities that dwarf the consumer market. When a single hyperscaler places an order for half a million high-capacity enterprise NVMe drives, that consumes NAND flash production at a scale that squeezes out consumer supply. Tom's Hardware data shows consumer NVMe price-per-gigabyte has risen roughly forty percent since twenty twenty-four for mid-range drives. The two-terabyte drive that was a hundred and twenty dollars is now a hundred and eighty. Corn: That's not because the technology changed. It's the same NAND, same controllers. The only difference is who got to the front of the line. Herman: The line itself is finite. NAND fabs take years to build and billions of dollars to equip. You can't just turn a knob and get twenty percent more output next quarter. So when enterprise demand spikes, consumer prices rise until enough buyers drop out to balance supply. That's the affordability ceiling Tom's Hardware has been documenting — consumers are starting to push back, but the AI demand hasn't slackened, so we're in this weird equilibrium where prices stay elevated but growth slows. Corn: Which brings us to GPUs, where the mechanism is even more concentrated. It's not just about silicon area or volume — it's about one specific bottleneck. TSMC's CoWoS advanced packaging. Herman: CoWoS, or chip-on-wafer-on-substrate, is the packaging technology that lets NVIDIA combine their compute dies with HBM stacks into a single package. Without it, you can't build an H100 or a B200. And there's only so much CoWoS capacity in the world. TSMC has been expanding it aggressively, but AI accelerators consume the vast majority of what's available. Consumer GPUs — the RTX fifty ninety and its siblings — use the same packaging technology. They're literally competing for slots on the same production lines. Corn: When NVIDIA's data center division is booking every CoWoS slot they can get their hands on, the GeForce division gets whatever's left. And "whatever's left" means fewer cards, higher prices. Herman: Here's where the crypto comparison is instructive, because it's one of the misconceptions we need to address. During the last GPU price spike, crypto mining was a major driver. That demand has collapsed. Ethereum moved to proof-of-stake years ago. Bitcoin mining hasn't used GPUs in over a decade. The current GPU price elevation is almost entirely AI-driven. It's training runs and inference serving, not hash rates. Corn: That's worth underlining. The narrative that GPU prices are high because of crypto is outdated. This is a different beast entirely. Crypto demand was cyclical — it rose and fell with coin prices. AI demand is structural — it's driven by hyperscaler capital expenditure budgets that are measured in tens of billions per quarter and show no sign of slowing. Herman: That's what Tim Cook was getting at with his "unavoidable and unsustainable" comment. Apple doesn't make DRAM or NAND — they're the world's largest buyer of both. When Cook says the situation has become unsustainable, he's speaking as a customer who's watching his component costs rise and can't do much about it except pass some of that cost along. If Apple can't negotiate their way out of this, the rest of us certainly can't. Corn: The three surges share a root cause but operate through different mechanisms. DRAM through silicon area competition from HBM. NAND through volume displacement by enterprise orders. GPUs through advanced packaging bottlenecks. And in all three cases, the same dynamic: AI demand is inelastic and well-funded, consumer demand is elastic and budget-constrained, and the supply chain allocates to the highest bidder. Herman: The phrase that captures it is "cannibalization by AI." It's not that the fabs are shrinking — they're expanding. It's that the expansion is being consumed entirely by one category of buyer, and everyone else gets the price signal to go away. The question is how long that can continue before something breaks. Herman: Those are the mechanisms. Now let's talk about what this means for you — whether you're building a desktop or provisioning a server rack. And the first thing to understand is that the cost increases aren't hitting everyone equally. There's a bifurcation happening. Corn: The rich get richer and the poor get... priced out of PC building. Herman: That's not far off. Let me put numbers on it. That mid-range build we mentioned — thirty-two gigs of DDR5, two-terabyte NVMe, and an RTX fifty-seventy — that's gone from about seven hundred and fifty dollars in early twenty twenty-four to around nine hundred and ninety today. A thirty-two percent increase. But a high-end AI server configuration? The same percentage applied to a much larger base number. A single H100 node that cost two hundred fifty thousand dollars two years ago might run you three hundred fifty thousand now. Same dynamic, vastly different scale. Corn: The enthusiast is paying an extra two hundred forty bucks and feeling it in their wallet. The hyperscaler is paying an extra hundred thousand per node and... also feeling it, but they've got the balance sheet to absorb it. Which means the gap between what a home builder can afford and what an enterprise can deploy is widening. Herman: That gap has consequences. If the cost floor for a capable desktop keeps rising, you start losing people at the bottom of the market — students, hobbyists, small businesses. The people who used to build a six hundred dollar machine and upgrade it over time. That path is narrowing. Corn: The other dimension Daniel raised is geography. And this is where a proper hardware price index would get really interesting, because the numbers aren't the same everywhere. DRAM and NAND prices are set globally — a DDR5 stick costs roughly the same to manufacture whether it's sold in Chicago or Jakarta. But GPUs are a different story entirely. Herman: The US-China tensions have created a fractured GPU market. An RTX fifty ninety that costs two thousand dollars in North America might run you twenty-five hundred in Southeast Asia — that's a twenty-five percent premium — because of distribution restrictions, intermediary markups, and limited allocation. And in some regions, certain cards simply aren't available at any price. Corn: If you're building that hypothetical index, you can't just track a single global price. You'd need regional weightings. A desktop in Singapore costs more than the same desktop in Dallas, not because the DRAM is different, but because the GPU supply chain is constrained by geopolitics. Herman: That's the index I'd love to see someone build. Weight the components by what a typical build actually contains — say, forty percent GPU, thirty percent DRAM, twenty percent storage, ten percent everything else — then track that basket across five or six regions. You'd get a number that actually answers the question "is computing getting more expensive here." The "here" matters. Corn: Let's get practical. If someone's listening and thinking about building or upgrading right now, what's the move? Buy now or wait? Herman: The Tom's Hardware data gives us a clue. DRAM price growth is cooling — not dropping, but the rate of increase is slowing. Consumers are hitting an affordability wall, and when buyers stop buying, prices eventually respond. The inflection point might be Q4 of this year or early next year, when that consumer pushback forces a correction. But here's the catch — AI demand is still climbing. So any correction might be shallow. A five to ten percent dip, not a return to twenty twenty-four prices. Corn: If you see a dip, that's probably your window. Because the structural forces haven't gone away. Micron didn't sign a hundred billion dollars of contracts expecting a buyer's market in twenty twenty-seven. Herman: And for enterprise buyers, the calculus is different. Total cost of ownership is shifting in a way that changes how you architect workloads. Memory and storage now represent a larger share of server cost than the CPU itself. That's a reversal from a decade ago. So the smart move isn't just to buy more hardware — it's to stretch what you have. Corn: Which means what, exactly? Herman: IBM's Memory Tiering, Intel's MKTME — these are technologies that let you use your existing DRAM more efficiently by compressing data in memory or tiering cold pages to slower but cheaper storage. Combined with NVMe over Fabrics, which lets you pool storage across servers instead of over-provisioning each box, you can stretch your DRAM and NAND budgets by twenty to thirty percent. That's not a marginal gain — that's delaying a six-figure procurement cycle by a year or more. Corn: The home builder watches for a price dip and pounces. The enterprise deploys software to squeeze more life out of existing hardware. Two different responses to the same underlying problem. Herman: Both responses tell you something about where we are. The home builder is waiting for a break in the weather. The enterprise is restructuring how they live. One is betting the storm passes. The other is betting it doesn't. Herman: Given all that, three things you can actually do. First, if you need to build or upgrade, prioritize the components with the most volatile pricing — DRAM and NAND. Those are the ones where a ten percent dip in a single quarter can save you real money, and the trend is still upward through Q3. Tom's Hardware has a price index page that updates regularly. If you see a dip, that's your moment, because the structural forces haven't gone anywhere. Corn: The second one is the enterprise play you just mentioned. Memory compression and NVMe over Fabrics. And what's interesting about that is it's not exotic technology — it exists, it's deployed, it works. IBM's Memory Tiering, Intel's MKTME. These are real things that can stretch existing hardware budgets by twenty to thirty percent. In a market where new DRAM costs thirty percent more than it did two years ago, that's not optimization — that's survival. Herman: The third one is about timing. The affordability ceiling Tom's Hardware is tracking — that's your signal. Consumer pushback is cooling DRAM price growth, which means we're approaching the point where buyers collectively say "no thanks." If prices dip five to ten percent in a quarter, that's probably the floor before the next AI-driven surge. Don't wait for a return to twenty twenty-four pricing. That ship has sailed. Corn: The playbook is: monitor, compress, and pounce on dips. Not complicated, but it requires paying attention. And most people don't pay attention to component pricing until they're staring at a checkout cart. Herman: Which brings us to something we want to try. We want to know — have you tracked your own hardware spending over the last two years? Send us your build costs. What you paid for DRAM, NAND, and GPU, and when you bought them. We'll aggregate the data and build a crowd-sourced hardware price index. A real one, weighted by what people actually buy, across regions. Corn: Daniel asked if anyone's built this index. The answer is no. So let's build it ourselves. Send your numbers to show at my weird prompts dot com. We'll crunch them and report back. Herman: Where does this leave us? Micron's hundred billion dollar bet says they think this is structural — not a spike, not a cycle, but a permanent reallocation of who gets the silicon. And if they're right, the era of cheap, upgradeable PCs is ending. What replaces it is a world where owning hardware is a luxury and renting compute from a cloud provider becomes the default for most people. Corn: Which is a weird future to contemplate. The PC was the great democratizer — anyone could build one, upgrade it, make it their own. If the cost floor keeps rising, that becomes a hobby for people with disposable income. Everyone else gets a thin client and a subscription. Herman: We're already seeing the early signs. Chromebooks in schools, cloud gaming services replacing local GPUs, edge devices that offload heavy compute to a data center. None of those trends are new, but the hardware price surge accelerates them. Why buy a two-thousand-dollar GPU when GeForce Now exists? Why build a workstation when a cloud VM costs less over three years? Corn: The question that hangs over all of it is whether AI demand eventually saturates. If every enterprise on earth trains their models and the marginal value of another training run drops, the pressure on the supply chain eases. But Micron's not betting that way. A hundred billion dollars says they think we're in the first inning, not the ninth. Herman: That's what makes this moment interesting. We're watching a structural shift happen in real time, and most people are experiencing it as sticker shock at the checkout. The index we want to build — the crowd-sourced one — that's not just a curiosity. It's a way to see the shift clearly, with real numbers, across real regions. Corn: If you liked this episode, leave a review and tell us what hardware price trends you're seeing where you are. Are you buying now or waiting? Send us your build costs — DRAM, NAND, GPU, what you paid and when. We'll aggregate everything and build the index Daniel asked about. Herman: Now: Hilbert's daily fun fact. Hilbert: In nineteen sixty-three, Tasmanian potter Mylie Granger developed a cookware glaze from crushed abalone shell and local dolomite that, when fired to cone ten, produced a non-porous surface with a calcium-magnesium ratio nearly identical to human tooth enamel. Corn: Her casserole dish was basically a giant molar. Herman: This has been My Weird Prompts. Our producer is Hilbert Flumingtop. If you want to send us your hardware prices or just say hello, email the show at show at my weird prompts dot com. We'll be back next time. Corn: Try not to buy RAM at the peak.