I ran 35 controlled energy tests on NVIDIA A100 and H100 GPUs on RunPod.
Standard monitoring tools missed something critical.
When nvidia-smi reports 0% GPU utilization, data center operators assume the GPU is idle. Billing stops. Cost tracking stops.
The power draw does not stop.
On A100 SXM, I consistently measured 67–146W while utilization reported 0%. I call it ghost power.
Key findings
- A100 idle floor: 67.1W — never drops lower
- Peak ghost power: 146.7W at 0% utilization — sustained 11 minutes
- H100 SXM: zero ghost power detected across 11 tests
- FP16 draws 60% more power than FP32 at same matrix size
- Power capping blocked at hypervisor level on RunPod — tenants cannot fix this
- $58.70 per GPU per year in pure idle waste
- At 1 million GPUs globally — $58M/year invisible waste
Live API & Interactive Docs
Base API URL: https://ai-gpu-brain-v3.onrender.com
Try it yourself (curl)
bash
curl https://ai-gpu-brain-v3.onrender.com/detect/a100/13