HBM is eating consumer RAM, and the bill lands on cheap phones

Simon Willison points to a David Oks essay arguing that AI's memory hunger is starting to reprice consumer electronics. The argument turns on wafer math. Memory makers split a fixed pool of fabrication capacity between HBM, DDR, and LPDDR, and producing a gigabyte of HBM eats more than three times the wafer capacity of a gigabyte of the others. HBM took about 2 percent of that wafer pool in 2025, and is on track for 20 percent by the end of 2026 as data centers absorb everything they can get.

The shortfall does not land evenly. DDR and LPDDR prices have to clear against AI-side demand, and the segment that cannot absorb the increase is the sub-$100 smartphone, which is the on-ramp to internet access across most of Africa and South Asia. Hacker News retitled the Oks post to capture the broader point: this is a consumer electronics story, not only a phone story, because anything that uses memory now competes for the same wafers.

Willison calls Oks's piece the clearest he has read on why memory-bearing devices will keep getting more expensive. Memory vendors are deliberately holding capacity tight to avoid a glut, so the squeeze is a strategic choice, not a temporary mismatch between supply and demand.

Why it matters

If you sell hardware or model an addressable market in emerging economies, AI infrastructure is now a line item on your bill of materials. The price floor on a usable smartphone is rising for several years, and the people priced out are exactly the next billion users your product roadmap probably depends on.