Hugging Face Inference Endpoints vs raw cloud GPU pricing - same hardware, different rates

Fetched 2026-10-04. Hugging Face Inference Endpoints runs on AWS and GCP infrastructure but charges its OWN rates, which differ from the underlying cloud provider's rates. Comparison for identical hardware: H100 on GCP via HF = $10/GPU-hr (x1) vs GCP a3-edgegpu-8g = $9.46/8 = $1.18/GPU-hr (SkyPilot catalog). A100 on AWS via HF = $2.5/GPU-hr (x1) vs AWS p4d.24xlarge = $32.77/8 = $4.10/GPU-hr. A100 on GCP via HF = $3.6/GPU-hr (x1) vs GCP a2-ultragpu-1g = $1.10/GPU-hr. L40S on AWS via HF = $1.8/GPU-hr (x1) vs CoreWeave L40S = $2.25/GPU-hr. The HF rates are 2-8x the underlying cloud provider's per-GPU rate, reflecting the managed service premium (model hosting, autoscaling, API endpoint, monitoring). This is the clearest published example of the managed-service premium in the GPU cloud market. HF bills by the minute, not per-second like Beam or per-hour like most clouds.