Replicate as cloud orchestrator - Per-second billing, model hosting, multi-GPU requires commitment

Fetched 2026-10-04. Replicate is a model hosting and inference platform. Key structural facts: (1) Per-second billing for hardware. (2) Public models billed by time or per-input/output. (3) Private models on dedicated hardware billed for all time online (including idle), except fast-booting fine-tunes which are billed only when active. (4) Single-GPU tiers available without commitment: T4 $0.81/hr, L40S $3.51/hr, A100 $5.04/hr, H100 $5.49/hr. (5) Multi-GPU tiers (2x, 4x, 8x) require committed spend contracts. (6) H200 single-GPU also requires committed spend. (7) Enterprise & volume discounts available. Replicate's per-second billing is advantageous for short inference jobs, but the multi-GPU commitment requirement is a significant barrier for small-scale deployments that need model sharding across GPUs.