Google's 2x Energy Efficiency Claim Is Real — But Here's What They're Not Measuring

This is a submission for the Google I/O Writing Challenge What I Found Building a Real Benchmark My project — the AI GPU Energy Optimizer — measures something the industry largely ignores: what GPUs consume when they're doing nothing. We call it ghost power. On an NVIDIA A100 SXM running on RunPod infrastructure, I measured: Idle floor: 67W — the baseline you pay for just having the GPU allocated Ghost power: up to 146W at 0% compute utilization — power draw with no workload run...