The Great Memory Robbery

For forty years, the computer business kept one promise better than any industry has ever kept any promise: memory got cheaper. Every year brought more bits for fewer dollars, so reliably that the whole culture of computing leaned on the assumption. Software bloated because RAM would cost less next year. Photographs went uncompressed because storage would cost less next year. Then, late in 2025, the promise snapped. A 32-gigabyte kit of DDR5 desktop memory, the sensible middle of the market, sold for under ninety dollars in early 2025. By December, listings touched five hundred. By the second of September, 2026, the builders’ forums had settled on 375 to 400 dollars as the going rate, and the daily trackers showed the price still climbing this week, up ten percent in seven days. One index of the broad market moved from 445 percent above its old baseline to 486 percent in a single month this summer. The entire digital world has been repriced, and the repricing took about ten months.

And it reaches everything, because memory is in everything. Sony’s PlayStation 5 climbed to $649. Microsoft raised the Xbox Series X to $799.99 on the first of August. Nintendo lifted the Switch 2 from $449.99 to $499.99 on the first of September and, unusually for that secretive company, named the culprit in its announcement: memory costs. A terabyte solid-state drive that sold near $45 in late 2025 sells near $90 now. Handheld gaming PCs brush $2,000, and Valve’s new Steam Machine arrived at $1,049, reportedly hundreds above its internal target. Lenovo, Dell, HP, and Asus have warned of 15 to 20 percent increases on finished machines. In a smartphone, memory once amounted to 10 or 15 percent of the parts bill; this year it runs 30 to 40, which is why mid-range phones are quietly shipping with less RAM than last year’s models, reversing a decade of upgrades. A core commodity has tripled and quadrupled while demand from ordinary buyers stayed flat, and that combination should be impossible.

How does a commodity triple while its customers buy no more of it than before? Start with the sellers. Three companies, Samsung, SK Hynix, and Micron, make more than nine of every ten memory chips on earth, past 95 percent by recent tallies. Memory is the most concentrated major commodity humanity trades in; OPEC at its 1970s height pumped about half the world’s oil. When three sellers hold a market and all three conclude that scarcity pays better than volume, the price becomes whatever they care to charge. All three reached that conclusion together, because all three carry the same scar.

The scar is 2023. Memory has always boomed and busted, and the bust that bottomed in 2023 was the worst in the industry’s books. Warehouses drowned in unsold chips, prices fell below the cost of making them, and the losses came in on freight cars: Micron dropped $5.8 billion in its fiscal 2023, SK Hynix posted an operating loss near eight trillion won, Samsung’s semiconductor division bled about fifteen trillion, and Samsung cut production, on some lines by as much as half, to stanch it. The survivors emerged converted. Gone was the old religion of market share, in which each company raced to add capacity and flood the channel; in its place came discipline as doctrine: build less than the world wants, and let the shortage do the selling. Wall Street canonized the doctrine at once. So when demand came roaring back, the three met it with a price list.

Then the demand that returned wore a face nobody had priced: AI, which hungers for a different memory altogether. High Bandwidth Memory, HBM, is DRAM dies stacked eight and twelve high like a wedding cake, wired vertically through the silicon, mounted millimeters from the GPU so data can pour in at terabytes per second. Every serious AI accelerator on earth is wrapped in the stuff. And HBM is a glutton. Micron’s chief executive, Sanjay Mehrotra, put a number on the appetite in March 2024, and his company restated it at the Hot Chips conference this August: producing a gigabyte of HBM consumes roughly three times the wafer supply of a gigabyte of ordinary DDR5, and the next generation, HBM4, widens the gap. The stacking is brutal on yields, since one bad die condemns the whole cake, so HBM yields run about two-thirds those of conventional parts. TrendForce projects that in 2026 HBM will absorb 22 percent of the big three’s wafer input while producing 9 percent of their bits, and in 2027, 30 percent of the wafers for 13 percent of the bits. Every wafer that becomes HBM is three wafers that never become the memory in your machine.

Behind the glutton stands a buyer with a checkbook this industry had never seen. On the first of October, 2025, in the presidential office in Seoul, Sam Altman of OpenAI sat across from Samsung’s executive chairman Jay Y. Lee and SK’s chairman Chey Tae-won, President Lee Jae-myung presiding, and signed letters of intent that read like an annexation: the two Korean giants would scale toward 900,000 DRAM wafer starts per month for OpenAI’s Stargate build-out, the data-center program of as much as $500 billion backed by Oracle and SoftBank. Global DRAM capacity runs a little over two million wafer starts a month, so one customer reached for about 40 percent of the planet’s supply of the material your computer thinks with. The wafers ship undiced, raw disks of silicon by the crate, memory bought wholesale the way refineries buy crude. The letters carry no delivery timetable, which in a shortage is its own weapon: the demand hangs over the market whether or not a single wafer ships. Dell cannot outbid that. Neither can Nintendo. Neither can you.

The quarterly numbers since then read like storm bulletins. TrendForce’s January survey projected first-quarter 2026 contract prices for conventional DRAM rising 55 to 60 percent; its February revision tore that up and printed 90 to 95, the largest quarterly increase ever recorded, with PC memory crossing 100 percent in some contract categories. The second quarter was forecast at 58 to 63 percent and the market blew through it; phone memory, LPDDR5X, jumped 89 percent in the quarter, a single 12-gigabyte part going from $77 to $146. NAND flash, the substance of SSDs, rose 55 to 60 percent and kept going. On the spot market tracked by Tom’s Hardware, a 16-gigabit DDR5 chip that cost $6.84 in September 2025 hit $27.20 by December and $42.50 by the tenth of August. Gartner coined a word for the condition, memflation, and its research director Ranjit Atwal spoke for the whole supply chain: “The speed at which the memory pricing has increased has shocked everybody.” The sellers are not shocked. Samsung’s first-quarter operating profit multiplied nearly ninefold year over year. SK Hynix crossed fifty trillion won in revenue for the first time, with operating margins reported above seventy percent. Micron had sold out its 2026 supply before 2026 was half over.

And the raid runs wider than HBM. An AI server is stuffed with ordinary memory too, a terabyte and more of server DDR5 riding alongside the accelerators, so the hyperscalers are emptying both shelves at once. A 64-gigabyte server module that cost about $450 at contract in late 2025 crossed $900 by the first quarter of this year, and the big buyers responded by signing multi-year supply agreements at those prices, locking the shortage in as policy. Channel inventories, the buffer between the fabs and your shopping cart, have thinned to two or three weeks. The market has no cushion left; every tremor at the factory now reaches retail within days.

Follow who pays. A hyperscaler buys memory by the exabyte, books it as capital expenditure, and depreciates it against revenue projections written in billions. A sixteen-year-old building a first PC pays retail, in cash, after a summer of mowing lawns. The shortage is a regressive tax collected at the parts counter. Independent memory brands, the companies that buy chips and assemble the sticks, have been warned their allocations could fall to 30 percent of this year’s supply in 2027, which means fewer choices and dead brands. Analysts project the sub-$500 laptop, the machine schools and first-job workers actually buy, could largely vanish by 2028, because memory now swallows up to 35 percent of a PC’s parts cost, up from a historical 15 to 18. Even the escape hatches are burning. DDR4, the older and cheaper standard, has more than doubled, because the big three are winding its production down to free the lines, and Raspberry Pi, maker of the cheapest real computers on earth, raised its prices and named memory as the reason.

Now to the question beneath the question: is this simply what technology prices do? Do they only rise from here? History howls no. Memory is the most violently cyclical commodity electronics has ever produced, and every boom in its four-decade record has died the same death. In 1985, Japanese producers dumped chips so aggressively that Intel, the company that created the DRAM business, abandoned it forever. In 1996, 16-megabit parts lost about 80 percent of their value inside a year. The dot-com collapse of 2001 gutted prices again. January 2009 killed Qimonda, Germany’s memory champion, outright. February 2012 took down Elpida, Japan’s last DRAM maker, in what was then the largest manufacturing bankruptcy in Japanese history; Micron bought the bones. Prices halved again through 2019. And 2023 delivered the bust that taught the survivors their discipline. Gluts follow booms in this industry the way hangovers follow weddings.

I lived through one of these panics from the inside. On the fourth of July, 1993, an explosion tore through Sumitomo Chemical’s plant in Niihama, Japan, a facility producing on the order of 60 percent of the world’s cresol epoxy resin, the black compound that seals memory chips into their packages. RAM prices doubled within weeks. I was writing computer books that year, and the trade press wrote obituaries for cheap memory, confident the age of the affordable megabyte had closed. Three years later came the 1996 glut, and the obituaries needed obituaries. The long arc dwarfs every one of these panics: a megabyte of RAM cost around $300 in 1985; by 2020 it cost a third of a cent, a fall on the order of a hundred-thousandfold, the steepest sustained price collapse of any manufactured good in human history. Even now, at the crest of the worst spike ever recorded, memory costs a small fraction of what it cost when the doomsayers of 1993 declared the party over. Prices for this technology fall down a long staircase for decades, and every few years they leap up several steps and terrify everyone standing on them.

Two honest cautions before anyone books the glut. This cycle differs from its ancestors in one structural way: the old busts began when consumer demand sagged, and this boom’s buyer is an industry spending borrowed hundreds of billions on faith, which could keep spending longer than any consumer ever did, or stop all at once. And the analysts charting the far side warn that the floor will reset, with projections clustering 30 to 50 percent above the old lows, meaning the eighty-dollar kit may be a museum piece for good. But cartel arithmetic has a known solvent, and it is at work in Hefei. ChangXin Memory Technologies, CXMT, China’s one proven DRAM maker, blacklisted by Washington and suing the Pentagon over it, unveiled homegrown DDR5-8000 chips last November, a generation behind the big three’s densities, and real, shipping, and cheap. Through 2026 the bulletins kept landing: production capacity reported to approach Micron’s this year, DDR5 yields past 90 percent by August, risk production of LPDDR6 and of HBM3E, PC makers fighting over its allocation, Gigabyte qualifying the chips on consumer motherboards, and Apple, of all companies, reportedly asking Washington for permission to buy. Cartels hold at three members. Four is where they leak.

So when does relief arrive by the calendar? Every construction schedule involved is public. The 2027 wave brings Micron’s new Singapore plant, a retooled Micron fab in Taiwan late that year, and SK Hynix’s $13 billion packaging complex at Cheongju, the largest HBM assembly site on earth, and every one of those is purpose-built to feed the accelerators, none of it aimed at the memory in your machine. Then 2028 adds SK Hynix’s fab at West Lafayette, Indiana, and a new Samsung line at Pyeongtaek, while Micron’s Boise expansion is a 2027-at-the-earliest story and its Clay, New York megafab, the largest planned addition anywhere, broke ground in January 2026 with first chips due around 2030. Forecasters have placed their bets accordingly. Intel’s chief executive, Lip-Bu Tan, told a February summit there is “no relief until 2028.” Counterpoint Research marks the fourth quarter of 2027 as the earliest point where the supply and demand curves could cross. Gartner’s memflation report says the second half of 2027; IDC expects settling by mid-2027 and calls the whole affair a structural reset of the industry; Samsung and SK Hynix themselves warn of tightness running into 2028; AMD told the Computex crowd in June that DDR5 pricing will stay abnormal into 2028. The honest window: a first crack late in 2027, breathable air through 2028 and 2029, and, if AI spending pauses in the season those delayed fabs all arrive, the classic overshoot, and a crash worth waiting for.

Which brings the question home to the checkout page: buy the new computer now, or wait six months for a better deal? The options deserve honest weighing. Waiting six months, to next March, is not effective because every needle still points uphill: the daily trackers showed DDR5 up another ten percent in the first week of September, contract increases are already booked into the winter quarters, manufacturers’ spring costs are still working their way onto retail shelves, and the first new supply of any kind is a late-2027 event. Even when memory breaks, it breaks slowly; the 2022 collapse took more than a year to travel from peak to trough. Six months buys a routine processor refresh wrapped in costlier memory, March’s prices for September’s machine. Waiting two to three years is effective because that horizon holds the actual relief: the 2028 fabs, the CXMT ramp, a possible AI digestion, and the historical certainty of the glut. Buying now, if you genuinely need a machine, is effective because today’s hateful price is the lowest one on the visible calendar, with two disciplines attached: buy the capacity the work requires and no more, since future-proofing at fifteen dollars a gigabyte is the most expensive insurance in computing, and buy upgradeable where you can, socketed memory over soldered, so the far side of the cycle can be purchased cheaply when it comes.

My recommendation splits on need, as every honest one must. If the machine on your desk still does the work, hold it, clean it, add nothing, and plan the purchase for 2028 territory, past the far ridge of this cycle. If the machine is failing or the work has outgrown it, buy now, this month, sized to the job, and stop watching the trackers afterward, because they will keep rising and reading them will feel like watching your house burn from across the street. The one seat nobody should choose is the middle one, because the six-month wait collects all of the price and none of the relief. In this market the discount runs on a calendar of years, and no one is printing a six-month coupon.

The machines that caused all this were built to remember everything: every book, every forum post, every photograph of every dinner, held in stacked silicon so a model can be asked any question and answer inside a second. That ambition repriced remembering itself. And the bill lands, this afternoon and every afternoon, at a parts counter somewhere in America, where a teenager who saved all summer for thirty-two gigabytes does the arithmetic of the robbed, and slides one of the two sticks back across the glass.