NAND prices are up 80% and the market hasn't noticed

In October 1984, a 28-year-old engineer named Eli Harari filed a patent for something he called System Flash. The idea was simple in hindsight and almost heretical at the time. Take the floating-gate transistor that EEPROM chips used to store a single byte at a time, and architect it so you could erase and rewrite entire blocks at once. Faster. Denser. Cheaper per bit than anything else on the market.

The established memory players laughed. Intel had core memory. Toshiba was building DRAM empires. Nobody wanted a slower, less reliable cousin of the chips they were already selling. Harari spent three years trying to raise money. He eventually scraped together $1.5 million from a Japanese venture group and founded a company called SunDisk in a Santa Clara office park. Within a decade his strange little flash chips were inside every digital camera on earth. By the mid-2000s they were inside every iPod, every USB stick, every phone. The name changed slightly along the way. The technology took over the world.

Flash memory is now the substrate of modern computing. Every photo, every model weight, every cached video, every database log. And for the last fifteen years, the people who built it have been treated by the market like a deeply unsexy commodity business. Boom, bust, boom, bust. Buy the dip, sell the rip, never pay more than 8 times earnings or you are a fool.

That framing is breaking right now, and almost nobody on the buy side has updated their model.

Here is what changed. In 2024 and early 2025 the memory industry consolidated more aggressively than at any point in its history. Kioxia and Western Digital merged their flash operations. Samsung pulled back capacity to defend pricing. Micron pivoted aggressively to HBM, the high-bandwidth memory that sits next to Nvidia GPUs, and started running its NAND fabs lean. SK Hynix did the same. Then the demand side detonated.

AI inference, the part of the AI economy that actually serves answers to users, is hilariously memory-hungry. A single frontier model loaded for inference can require terabytes of fast storage per server. Multiply that by the hyperscaler buildout. Multiply that again by the enterprise wave that started landing in Q4 2025. Then add the secondary effect almost nobody is modeling, which is that training runs are now generating so much synthetic data that storage demand from the training side is growing faster than compute demand. Anthropic, OpenAI and xAI are each sitting on data hoards measured in exabytes. They have to put it somewhere. SSDs are the only physically viable answer.

DRAM gets all the press because it sits next to the GPU. Flash is doing the actual work of holding the AI economy together. And flash supply, three years into a deliberate underinvestment cycle, cannot meet the demand curve that arrived in 2026. Contract NAND prices are up 47 percent year over year as of last month. Spot prices on enterprise SSDs in certain capacities are up over 80 percent. Hyperscalers are signing two and three year supply agreements with prepayment clauses, something that has not happened in memory since the 1990s.

This is not a cyclical boom. The supply response that would normally crash prices in 18 months is not coming. Building a new NAND fab takes four years and roughly $20 billion. The Japanese and Korean producers are explicitly telling investors they will not add capacity speculatively this cycle. They learned in 2018 and again in 2022 what happens when you do. So we are entering a period, probably 2026 through 2028, where the memory industry runs structurally tight for the first time in two decades.

The market sees this and is buying Micron. Fine. Micron is a great business. Up 180 percent in the last year. Most of the move is behind you.

The market is not yet seeing the company that just spun out of a 36-year corporate marriage, that owns the second largest NAND fab footprint on the planet, that trades at a fraction of Micron's multiple, and that happens to be the direct corporate descendant of Eli Harari's SunDisk. Same patents. Same Yokkaichi joint fab with Kioxia. Same architecture lineage. Newly independent, newly capitalized, and walking into the tightest flash market since the company was founded.

It started trading as a standalone business in February. Most of Wall Street still has it filed under the wrong category.