A recurring question from clients this year: our product uses a small serial flash and a modest DDR device, so why has the memory line doubled? Nobody here is building AI hardware.
The mechanism is capacity, not demand for your specific part. High-bandwidth memory for accelerators is built on the same wafer capacity, the same advanced packaging lines and the same test equipment as ordinary DRAM and NAND. When a supplier can sell that capacity into HBM at a much higher margin, everything else is deprioritised. Older, smaller-density parts are the first to be squeezed, because they occupy the same line while earning the least.
The result is that scarcity flows downhill. A 512 Mbit DDR3 device or a 64 Mbit serial flash — parts that were pennies and always available — can become the longest lead item in a design, and occasionally get end-of-lifed rather than re-ramped.
Three practical defences, all cheap if applied early. Design the memory interface so a different density or a second vendor drops in without a respin — same footprint, compatible command set, address lines routed for the larger part. Keep meaningful headroom rather than fitting the firmware exactly to the part you specified, because headroom is what makes substitution possible. And check the manufacturer’s product longevity commitment before the schematic is frozen, not after the first shortage.