Hugging Face Inference Endpoints as orchestrator - Per-minute billing, managed model hosting

Fetched 2026-10-04. Hugging Face Inference Endpoints is a managed model hosting platform. Key structural facts: (1) Hourly rates shown but actual cost calculated by the minute - finer granularity than per-hour clouds but coarser than per-second. (2) Runs on AWS and GCP infrastructure but charges its own rates (2-8x the underlying cloud provider's per-GPU rate). (3) GPU instances: AWS T4 x1 $0.5/hr, L4 x1 $0.8/hr, A10G x1 $1/hr, L40S x1 $1.8/hr, A100 x1 $2.5/hr, H200 x1 $5/hr. GCP: T4 x1 $0.5/hr, L4 x1 $0.7/hr, A100 x1 $3.6/hr, H100 x1 $10/hr. (4) No Azure GPU instances listed. (5) Autoscaling with min/max replicas. (6) Pricing formula: instance hourly rate * ((hours * min replicas) + (scale-up hrs * additional replicas)). (7) Requires HF subscription + credit card. (8) Some instance types need quota requests. The managed-service premium is the clearest published example in the market - HF charges 2-8x the underlying cloud provider's rate for the same hardware.