Replicate vs raw cloud GPU pricing - Per-second rates vs per-hour cloud rates

Fetched 2026-10-04. Replicate per-second rates vs raw cloud providers: H100 $5.49/hr (Replicate) vs CoreWeave H100 $6.16/GPU-hr vs AWS p5 $6.88/GPU-hr vs Together H100 $3.99/GPU-hr. A100 80GB $5.04/hr (Replicate) vs CoreWeave A100 $2.70/GPU-hr vs AWS p4d $4.10/GPU-hr. L40S $3.51/hr (Replicate) vs CoreWeave L40S $2.25/GPU-hr vs Beam L40S $0.76/hr. H200 $5.49/hr (Replicate) vs CoreWeave H200 $6.31/GPU-hr vs Together H200 $5.99/GPU-hr. Replicate is priced at a premium to raw cloud providers, reflecting the model hosting and API layer. The per-second billing is advantageous for short inference jobs (no minimum hour billing), but for sustained inference the per-hour rate is what matters. Replicate's H100 at $5.49/hr is competitive with Together's dedicated inference ($5.49) but above Together's GPU Clusters on-demand ($3.99). Multi-GPU tiers require committed spend, making Replicate unsuitable for small-scale multi-GPU inference without a contract.