Managed service premium quantification - Same hardware, different providers, different rates

DERIVED ANALYSIS from multiple records in this repo. This record quantifies the managed-service premium for identical hardware across providers. H100 SXM per-GPU-hour rates: Beam $1.83 (on-demand machine, includes CPU/RAM/NVMe); Together AI $3.99 (GPU Clusters on-demand); Together AI $5.49 (Dedicated Inference); CoreWeave $6.16 (on-demand); AWS p5 $6.88 (on-demand); Replicate $5.49 (per-second); HF GCP $10.0 (per-minute); GPU.ai $3.49 (secure tier). The spread from cheapest to most expensive is 5.5x ($1.83 to $10.0). The managed-service premium is the difference between the raw cloud rate and the managed provider rate. For H100, the raw cloud rate (GCP A3) is $1.18/GPU-hr, and the most expensive managed rate (HF GCP) is $10.0/GPU-hr - an 8.5x premium. The premium reflects: orchestration layer, API endpoint, monitoring, support, autoscaling, model hosting, and convenience. For inference deployments, the managed-service premium is the single largest cost decision: choosing between raw IaaS (cheapest, most operational overhead) and managed platforms (most expensive, least operational overhead).