Memory — Same Name as 2022, a Different Product: Why the AI Cycle Is Not the Old One

Seoul reopened on Monday after two days shut for Chuseok and closed 2.37% lower at 6,913.17. The memory makers led: Samsung Electronics fell 4.73% and SK Hynix 4.57%. In Tokyo the flash-memory maker Kioxia lost 4.37%. Micron, the largest US memory maker, reports on Wednesday. Every sell-off in these names now raises the same question: has memory put in its top, the way it always did before? We think that question is asked with the wrong picture in mind. Memory in 2026 carries the same name as memory in 2022, but it is a different product with a different use. It is still a cycle. It is not the old one.

Same name, different product

In 2022 memory was a consumer part. Most DRAM and flash went into phones and PCs, prices followed device sales, and a weak holiday season meant a glut: factories kept producing, buyers stopped buying, and prices collapsed within a few quarters. That is the cycle investors learned, and it is the cycle they still fear.

In 2026 the growth sits somewhere else. High-bandwidth memory (HBM) is stacked next to AI chips inside data-center servers, and a second layer of demand comes from the working memory an AI model needs to hold a long conversation. That working memory grows with the context window. According to Artificial Analysis, cited by the I/O Fund, OpenAI's GPT-3.5 Turbo handled 4,000 tokens in 2023; GPT-5.5 handles 922,000 - 230 times more in three years. The HBM market grew from roughly $4 billion in 2023 to $34.6 billion in 2025, and Micron expects $100 billion in 2027. Same word, "memory". Different buyer, different contract, different reason to buy. Our memory supercycle entry explains the shift from commodity to AI constraint.

The closest picture outside chips is a drug that finds a second use. Semaglutide was approved to treat diabetes and sold as Ozempic. Then it turned out to work for weight loss and was sold again as Wegovy. The molecule did not change; the number of people who wanted it did, and for years the makers could not produce enough. Anyone who valued it on the old diabetes market was looking at the wrong patient list. Memory is in the same place: the chip is still called DRAM, but the buyer is now an AI data center, not a phone maker.