Maybe, but inference costs can come down too with more purpose built hardware and continual optimization and quantization strategies.