The AI boom broke the memory market, and the bust could be brutal

The AI boom has broken the memory market’s oldest rule. Prices that should be falling are soaring instead, and the AI memory crunch will not ease until 2028. The hangover, when it comes, could be brutal.

Memory is meant to be boring. For decades DRAM and NAND flash have behaved like any commodity, sliding through predictable booms and busts as new factories flood the market and crater prices. That cycle has now snapped.

As The Register lays out, the AI build-out has devoured every spare chip, and the result is an AI memory crunch with no quick way out.

The cycle that broke

By the old rules, 2025 and 2026 should have been a down-year. Prices should have drifted lower as supply caught up. Instead they climbed. GPU servers need vast amounts of high-bandwidth memory, DDR5 and NAND, and that demand has swallowed the lot.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!

The knock-on has reached the shelf, pushing up the cost of consumer electronics and, as The Register puts it, killing the cheap smartphone. For the makers it is a windfall. SK Hynix and Micron have tripled their revenues in a year, and Samsung has roughly doubled its own.

A fix measured in years

The obvious answer is more factories, and the money is moving. In June, South Korean President Lee Jae Myung announced a $576bn drive led by SK Hynix and Samsung. Micron said last week it would spend up to $3bn shoring up its US supply chain, with more going to Singapore, Taiwan and Japan.

None of it arrives soon.

A new memory fab is one of the hardest things in industry to build, from permits and ultra-pure water systems to lithography tools that take months to tune. Anything started today needs at least three years to come online and longer to reach full capacity. The analyst firm IDC does not expect relief until 2028.

Everyone downstream pays

Until then, the cost lands on everyone below the memory giants. Chipmakers are already redesigning around the shortage. Samsung is readying a budget solid-state drive that strips out its onboard DRAM entirely, TechRadar reports, borrowing a slice of system memory to keep the price down.

The AI companies feel it too.

Every model developer renting AI infrastructure is paying more for it, which is poison for margins already stretched thin. The labs have spent four years and hundreds of billions in venture capital. They still have to turn the cost per token into a profit, and dearer memory makes that sum harder.

The bust nobody is pricing in

Here is the trap. Memory vendors finance their giant fabs on the strength of a boom. They know full well that when the new capacity finally switches on, it can flood the market and collapse prices. That is the cycle, and AI has only made the swing wider.

The whole edifice rests on one assumption, that demand for AI keeps climbing. If it falls short just as the new fabs ramp, the makers face what The Register calls a bust to end all busts. So the real race is not memory against demand. It is whether the fabs come online before the AI bubble deflates and the music stops.

For Europe, watching from the sidelines of a supply chain it barely controls, the squeeze is a reminder of how exposed it is. The region is racing to build its own AI datacentres. Yet a handful of firms in Korea, Taiwan and the United States ration and price the parts to fill them.

The RAMpocalypse is a good time to be a memory maker. It is a nervous one for everybody else, and the reckoning has only been deferred, not cancelled.