Latest / Logosofia no Dia a Dia
GPU Idle Waste and Serverless Green Computing
Everyone talks about the cost savings of serverless GPU platforms, but the environmental angle gets far less attention. In this episode, we break down the surprising physics of GPU idle draw — an H100 sitting idle still pulls 100-150 watts — and why the relationship between utilization and power consumption is far from linear. We explore how serverless platforms like Modal pack workloads densely across time zones, achieving 80-90% utilization versus the 30-60% typical of dedicated instances. The result: energy per useful teraflop can drop by more than half, with embodied carbon savings from…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.