Ironically it’s because increasingly it’s 100 words generated by an LLM from a 10 word prompt. We’ve effectively found a way to keep entire datacenters humming to do nothing more than obfuscate signal in very expensive noise and the lossily try and extract it again.