AI engineers claim new algorithm reduces AI power consumption by 95%

Simple question: if true, would power consumption stay at 100% because we’d work the algorithm harder?

I had assumed the latency etc were based on what was desirable for the use case and hardware, rather than power consumption.