Inside its August forecast, Gartner put 2026 memory revenue at $837 billion — about 54% of the entire semiconductor market. DRAM revenue is projected to grow 246.6% and NAND flash 371.9% in a single year. Memory has gone from a commodity line item to the largest single category in the industry.
Almost none of that is driven by the devices most engineers work on. It is high-bandwidth memory for AI accelerators, and Gartner analyst Shrish Pant expects HBM demand to keep supporting memory revenue through 2027 and beyond. HBM competes for the same fabs, the same advanced packaging and the same test capacity as ordinary DRAM, so scarcity propagates downward.
The practical effect reaches designs that have nothing to do with AI. A product using a modest DDR4 device, an eMMC part or a serial flash can find that its memory line has quietly become one of the most expensive and least predictable items on the bill of materials, sometimes with lead times measured in quarters.
Two habits help. Keep the memory interface generic enough that a different density or vendor can be dropped in without a board respin, and avoid designing right up to the capacity limit of a part — headroom is what lets you substitute when the exact device you specified is unavailable.
Source: Gartner