A major investment bank called us "the new memory" — third-largest AI-server bill after GPU and memory, with AI demand seen growing ~4.3x by 2030 against ~10% annual capacity growth. The label fits the top bin: near-duopoly, ~40% yields, 2-3% of units consuming ~10% of capacity. But we're under 0.1% of rack cost versus memory's ~25%, so buyers don't defend the oligopoly. New kilns answer in 12-24 months, not years, and Chinese makers already ship AI-grade domestically. Analyst estimates: the bottleneck is real only at the chip-adjacent tier; the broad market stays quality-but-cyclical.