The Memory Rollercoaster Is Coming for AI
The memory business has a pattern. It goes like this:
Demand surges. Prices skyrocket. The big three — Samsung, SK Hynix, Micron — report numbers that make other industries look like lemonade stands. Everyone races to build new fabs. Then, three years later when those fabs come online, demand has normalized, supply floods the market, prices collapse, and everyone spends the next two years bleeding red ink and apologizing to shareholders.
It’s happened in 1996, 2001, 2008, 2012, 2018, and 2022. Like clockwork. The chip industry calls it the “silicon cycle,” but it’s really just the most expensive game of musical chairs ever invented.
Right now, we’re in the euphoria phase. SK Hynix and Micron have tripled their revenue in a year. Samsung has doubled. The AI datacenter boom is devouring every stick of HBM, DDR5, and NAND flash the industry can produce. You can’t even buy a budget smartphone anymore without paying a premium because there’s no DRAM left for the cheap stuff. The South Korean government just announced a $576 billion investment led by SK Hynix and Samsung to build even more capacity. Micron is throwing another $3 billion at U.S. supply chain resilience.
And none of it will be online for at least three years.
That’s the crux of it, and it’s where the story gets interesting. An IDC report cited by The Register warns that memory prices won’t ease until at least 2028. New fabs take years to stand up — the clean rooms, the lithography tools, the ultra-pure water systems, the yield ramp. You can’t shortcut a wafer fab. Physics doesn’t care about your Series C.
So here’s the bet the AI industry is making, whether it admits it or not: memory prices stay high long enough that AI companies figure out how to make money before the VC taps run dry.
The objection you’re thinking of
“Yeah but this time is different. AI demand is structural, not cyclical.”
I’ve heard this before. People said the internet was structural in 1999. They were right — the internet was structural. Amazon, Google, and Salesforce all survived. But Cisco, whose routers were the “picks and shovels” of that boom, saw its stock drop 80% when overcapacity hit. The technology was real. The cycle was still undefeated.
The memory cycle doesn’t care whether the underlying demand is “real.” It cares about the lag between investment and production. Three years is a long time in an industry where VC-funded startups burn cash every quarter. If the music stops before the new fabs come online, the AI companies holding the bag won’t be the ones selling chips.
The second objection
“But memory makers learned their lesson — they’re being disciplined this time.”
They always think they are. Then the boom stretches into year two, the revenue numbers make CFOs dizzy, and suddenly “strategic capacity expansion” turns into “everyone build everything.” The $576 billion Korean investment is a bet that demand will keep growing at current rates. If AI training demand plateaus — which it will, eventually, because training runs on budgets, not infinity — that much capacity becomes a problem.
The memory industry has a collective action problem. Individually, each company is making a rational bet. Collectively, they’re building the next glut.
The third objection
“AI companies will figure out profitability before then.”
Some will. OpenAI, Google, Microsoft — the ones with real products and real customers. But the margin math is brutal right now. High memory prices are inflating the cost of inference infrastructure. The cost per token isn’t falling as fast as the optimists projected, partly because the hardware underneath it isn’t getting cheaper. If you’re an AI startup paying inflated prices for HBM-equipped hardware, your unit economics are getting squeezed from both sides — high CapEx and uncertain demand.
And here’s the thing: I’d change my mind on this if the data shifted. If inference costs start dropping faster than expected, or if someone figures out a memory-efficient architecture that bypasses HBM entirely, the calculus changes. But right now, the data says memory prices stay high through 2028, and the typical VC fund doesn’t have that kind of patience.
Where I land
The memory industry’s boom-bust cycle is older than most of the companies buying its products. It has survived the PC revolution, the internet revolution, the mobile revolution, and the cloud revolution. Each time, it looked different. Each time, it ended the same way.
AI might be the biggest boom yet. It might also produce the gnarliest bust. Not because AI is fake — but because the people building the picks and shovels are making the same mistakes they always make, just at a scale that would make your eyes water.
The fabs are coming. The question is who’s still standing at the table when they arrive.
Sources: The Register — Memory makers are slaves to the boom-bust rollercoaster, IDC report via The Register, Samsung profit surge