“Etched believes three things are fundamentally required to scale inference: building a chip with the most flops in a given power budget, achieving the lowest latency between chips, and producing the chip at the largest possible scale.”