GPU cloud billing granularity comparison - Per-second, per-minute, per-hour, per-millisecond

DERIVED ANALYSIS from multiple records in this repo. Billing granularity varies significantly across providers and has a direct cost impact for short or intermittent workloads. Per-millisecond: Beam (serverless GPU tasks). Per-second: Beam (serverless), Replicate, GPU.ai, Vast.ai. Per-minute: Hugging Face Inference Endpoints. Per-hour: AWS, GCP, Azure, CoreWeave, Together AI, Paperspace, Anyscale, Lambda, Runpod, Fireworks, Baseten. The practical impact: for a 5-minute inference job, per-hour billing charges 1 hour ($3.99 for H100 at Together), per-minute billing charges 5 minutes ($0.42 for H200 at HF), per-second billing charges 300 seconds ($0.48 for H100 at Beam serverless). For short jobs, finer granularity is cheaper. For sustained jobs, the per-hour rate is what matters. Beam's per-millisecond billing is the finest granularity available, but it is only for serverless workloads (not sustained inference). The billing granularity decision should be driven by the workload pattern: sporadic/short jobs favor per-second or per-millisecond, sustained/long jobs favor per-hour.