Quantization Is How You Buy Tenants
Precision as a capacity lever on a single chip — why bf16 and not fp16 on TPU, what FP8 calibration actually involves, and why weight quantization and KV cache quantization are two independent levers that multiply.