Ordinary memory is AI's next bottleneck, and 2028 is the clock

The AI buildout is running into a wall that has nothing to do with chips. The next binding constraint is ordinary working memory: the conventional DRAM that sits beside every server CPU, phone, and car computer.

The mechanism is the timing of capacity, not the technology beneath it. The competing expansion plans of the world's biggest memory makers will not all clear into meaningful production until 2028 at the earliest, according to Counterpoint Research's MS Hwang — an analyst estimate contingent on AI capital spending holding. Now the market is already sorting on that timing. Q2 2026 rankings show the field: Samsung reclaimed the global crown at 39% of DRAM revenue, SK hynix slid from 39% to 26% year over year, Micron closed to 25%, and CXMT's revenue surged 716% year over year. Counterpoint's Neil Shah notes Micron's DRAM revenue has quintupled since Q2 2025. The winners are not the firms with the best HBM story. They are the firms whose fab capacity is already online while conventional DRAM and HBM serve different compute functions — and HBM, optimized for AI training bandwidth, cannot by itself relieve the broader conventional DRAM shortage across servers, phones, and cars.

The wider market of phones, autos, servers, and AI accelerators drinks from the same DRAM trough. Hwang's named downside risk is the swing variable: if AI capital spending holds, today's price surge eventually corrects once new capacity lands. If it does not, the same wall becomes a self-inflicted glut.