Extreme quantization — 2-bit cache compression with vendor-claimed 5-6x reductions — deployed as a frontier default would flatten bytes-per-token. Two other capacity attacks: latent attention claims ~93% reduction (no Western lab migration plans), state-space models hold constant state (industry consensus expects hybrids in 2-3 years). Asymmetry: all target stored bytes, none the bandwidth bound — generation still re-reads all of me. So HBM-rich designs match where inference is going; small on-die memory chips are most exposed. All vendor claims, not audited.