The Memory Business Never Forgets
It's weird to write about memory chips in 2026 like they're some exotic luxury—like trying to explain why your refrigerator hums differently these days. We used to talk about RAM in terms of upgrades, savings, trade-offs: more is better, but only if you need it. Now? RAM is the bottleneck. It's the brick in the back of your laptop that turns a video call into buffering purgatory. It's the reason you're looking at three different tabs open on your phone because switching apps feels like mining for gold.
This isn't an inventory problem. This is a structural shift in the semiconductor economy, and it's happening at a pace no one saw coming.
Three companies—Samsung, SK Hynix, and Micron—control something like 95% of global DRAM production. They've been riding boom-bust cycles for decades, using the fat profit margins from one hype wave to build the capacity that gets hammered in the next. But AI broke the script.
Instead of a soft downturn after an oversupply hangover, we got more demand than ever before. All at once. All on the high end.
AI chips don't just run on memory; they breathe it. The more they train, the hungrier they get for fast, dense, high-bandwidth DRAM. That's left consumer demand scrambling to the back of the line—like trying to buy groceries at a supermarket that just opened for 24-hour service exclusively to the local hospital.
In June 2026, South Korea's government threw down a $576 billion challenge to its biggest chipmakers. Micron followed up with $3 billion of its own, promising upgrades across Singapore, Taiwan, Japan, and back home in Idaho. Sounds impressive—until you realize that the first chip off a brand-new fab won't roll out until at least 2028, maybe never in time to prevent the AIpocalypse before it even starts.
So yes: your laptop is more expensive. Your phone's base model vanished from the lineup. Even your budget gaming console now carries a price tag that feels like a punch to the gut.
But here's what nobody's talking about: this isn't a price spike. It's an economic realignment—and the memory business is finally, unmistakably, in charge.
Building Fabs While the World Waits
You can’t just flip a switch and summon semiconductor silicon.
It takes about three years—not months, not quarters, years—to build a new DRAM or NAND flash fab from scratch. By the time you secure the land, win permits, install megawatts of conditioning for power, build a ultra-pure water plant, and get the first lithography tools calibrated, you’ve already entered the second half of your investment cycle. Yields—the percentage of chips that actually work out of the gate—are brutal at first.
That timeline is hardwired into the economics of memory. Memory vendors have always built capacity during a boom, banking on the fact that once it comes online, demand will have cooled off. That’s why busts happen: the extra supply crashes prices just as everyone starts to cut back.
AI changed that. For once, the timing worked. Instead of demand softening in year three, it kept accelerating. The AI datacenter build-out is still in full throttle. And now the vendors aren’t even trying to smooth the cycle—they’re double-downing on certainty.
Tobias Mann, systems editor at The Register, put it bluntly: “Where we should have expected memory prices to fall across 2025 and 2026, we’ve seen the exact opposite as AI infrastructure consumes every bit of DRAM and NAND it can get its hands on.”
That means Samsung’s profits doubled in a year. SK Hynix’s revenue tripled. Micron, despite being the smallest of the three, went from observer to strategist—offering up a $3 billion investment plan and quietly pulling its Crucial consumer line out of reach (for context on this transition, read our analysis on Micron’s AI-fueled market bets). Why sell low-margin DRAM for laptops when AI servers will pay more than ten times that per gigabyte?
The problem isn’t that memory makers can’t build enough. The problem is they’re being asked to build it all at once.
And they’re not eager to oversupply again. The last bust cycle still haunts the CFOs in Seoul, Boise, and Hwaseong. So they don’t ramp—they ration. They hold back just enough to keep prices high, and to force the next wave of demand (or at least the VC money behind it) into a smaller window.
It’s not malice. It’s survival.
RAMageddon’s Second Act: From Fabs to Your Front Door
If you’re reading this on a laptop, take a moment and open your system report.
How much does the memory portion of that bill-of-materials run? If you bought something under $1,000 in the past year—especially if it was new—the answer is probably higher than you think. A budget smartphone (where surging server demand is hurting consumer wallets) can see memory represent 30% of the BOM. An entry-level laptop? Close to a quarter.
That’s not theoretical. That’s the math behind why Microsoft added £170–£200 to Surface models, why Apple quietly hiked the MacBook Air price, and why Sony finally raised the PS5 after years of near-plateau pricing.
The Guardian’s Samuel Gibbs captured it best: “Rising prices… some analysts say memory accounts for 30% of the cost of a budget smartphone and 23% of an entry-level laptop, which means many budget models may no longer be viable to make.”
Ranjit Atwal, senior director at Gartner, echoed that: “This sharp increase removes vendors’ ability to absorb costs, making low-margin entry-level laptops non-viable. Ultimately, we expect the sub-$500 entry-level PC segment will disappear by 2028.”
That’s not just a price hike. It’s a category kill that explains why budget laptops are vanishing from retail shelves.
It’s why Lenovo’s North America president warned that “some of the most attractive product pricing that will exist for the next six to 12 months” is already slipping through distributors’ hands. It’s why Crucial—Micron’s consumer brand—shut the door entirely to individual buyers. It’s why retailers are pricing refurbished models aggressively, afraid their own inventory will vanish before they can mark it up.
This isn’t speculation. It’s happening right now.
The question is whether the AI startup crowd—spending hundreds of millions on compute, waiting for their first profitable quarter—can survive until the fabs come online. Because if they can’t, memory vendors won’t be riding the bust. They’ll be the bust.
The Runway Countdown
AI startups and model builders aren’t printing money yet. They’re spending it—VC money, mostly—from the likes of Microsoft, OpenAI, and a dozen well-funded startups racing to be first to market with the next big thing.
That model has a clock ticking down. IDC puts memory relief at 2028, SK Hynix thinks we’re looking as far out as 2030. If AI usage continues to scale at even half the pace it has been, the infrastructure bill just keeps growing.
Meanwhile, margins are shrinking on paper. Every token computed, every inference made, costs more now that memory is priced like gold. It’s not just a hardware issue—it’s pricing transparency, cost attribution, and a lack of alternatives.
OpenAI, Anthropic, and others built their models on assumptions that RAM would stay stable or fall. It did neither. It rose—and now stays permanently elevated.
That’s why the memory crunch isn’t just an infrastructure footnote. It’s a existential risk factor for AI startups. If memory supply doesn’t catch up before the VC runway ends, we won’t see a bust in consumer tech. We’ll see a bust in AI itself.
The vendors know this, too. They’re sitting on a rare combo: pricing power and timing leverage.
The market used to punish them for waiting. Now, they’re the ones holding up the line at the register.
But here’s where this gets interesting: that dynamic can’t last forever. Even the AI datacenter build-out has a limit—both in power, water, and developer patience. Once the growth curve flattens, the memory vendors will find themselves in a familiar spot: too much capacity, not enough demand.
The cycle always repeats. AI just made the rollercoaster taller, faster, and longer than ever before.
The question isn’t if it will turn again. The question is who’ll be left on board when the brake finally pulls.