The operational inference costs of artificial intelligence models decrease rapidly over time