Lambda memory tuned for cost, not for vibe.

The 1024 MB default was a 2019 decision made when most of the functions were small Node.js handlers. The fleet had grown — some had become Python data-heavy workers (under-memoried at 1024), some had stayed lightweight (over-memoried at 1024). Nobody had revisited.

Lambda Power Tuning ran each function against a Step Functions state machine that tried multiple memory configurations against representative payloads. The output was a recommended memory size per function balancing cost and duration. Most functions came in lower than 1024; a few came in much higher.

Compute bill dropped 41% across the fleet. p99 latency improved 14% on the previously-starved functions. Cold start impact was function-by-function — some improved with the larger memory (and faster CPU); some got slightly worse and stayed at the higher tier anyway.