Replicate multi-GPU tiers require committed spend contracts

Fetched 2026-10-04. Important structural finding: Replicate's multi-GPU tiers (2x, 4x, 8x) for A100, H100, H200, and L40S all carry the note 'Additional Multi-GPU capacity is available with committed spend contracts.' This means the published per-hour rates for multi-GPU configurations are LIST prices, not walk-in rates - you must negotiate a committed spend contract to actually access them. Single-GPU tiers (1x A100, 1x H100, 1x L40S, 1x T4) are available without commitment. This is a significant barrier for small-scale inference deployments that need multi-GPU inference (e.g. for large model sharding). The H200 single-GPU tier also requires committed spend. Enterprise & volume discounts available via /enterprise page.