DRAM

DRAM is the ordinary fast working memory that sits in almost every computer and server — the near-term scratchpad a processor reads from and writes to while it runs. In the AI chain it looks mundane next to its exotic cousin, high-bandwidth memory, but the two are made on the same lines by the same handful of firms. That shared manufacturing base is the whole point: capacity is fungible, so a wafer that could have become commodity DRAM can instead be steered toward the far more lucrative stacked memory that accelerators demand.

That fungibility is why the model treats DRAM as a coupled market rather than a sleepy commodity. When makers divert capacity to feed AI, conventional DRAM tightens and prices firm across the wider server and PC base — a squeeze felt far from any GPU. The exposed names are a concentrated set: Samsung, SK hynix and Micron anchor the leading edge, with Nanya, Winbond and China's CXMT filling out the field. Their pricing power rises and falls with how aggressively the industry reallocates the same silicon between two markets.