I have a lambda that runs inference on a model for a feature that is seldomly used.

Lambda is far more economic given the memory requirements, but we recently moved to Rust+Candle to shave off ~300ms on cold starts as the lag could be really jarring.