An AI accelerator is only as fast as its ability to feed itself data, and the thing doing the feeding is high-bandwidth memory: stacks of DRAM bonded a hair's breadth from the compute, now the component everyone is fighting over. The convenient story is that memory is a supporting act to the processor, a spec on the datasheet. The supply chain tells a sharper one. The advanced packaging that makes HBM possible consumes wafer capacity and manufacturing effort out of all proportion to the silicon it yields, and that capacity is carved from the same lines that produce the plain memory inside phones, laptops and cars. So when AI's appetite scales, it does not just strain a specialty product; it pulls supply away from ordinary computing, and prices move everywhere at once. The looming RAMageddon is not a chip shortage in the usual sense. It is a single scarce resource, memory bandwidth, that the entire electronics world happens to share.