Independent reference. We are independent of every vendor listed. No affiliate links. No sponsored placements.

RFP template - Last verified June 2026

GPU cloud RFP template

Six sections of vendor questions covering the line items that decide a GPU procurement. Copy and adapt for your shortlist. Apply consistently so quotes are comparable.

1. Pricing

Publish per-GPU per-hour rates for every SKU we plan to use (H100 SXM, H100 PCIe, H200, A100 80GB, L40S, A10G as applicable).
Break out reservation tiers (1, 6, 12, 24, 36 month) with the percentage discount versus on-demand for each.
Confirm whether per-GPU rate includes host CPU, system RAM, local NVMe, and intra-node networking.

2. Networking

Confirm InfiniBand or RoCE topology (rail-optimised, fat-tree, dragonfly) and oversubscription ratio.
Confirm intra-cluster bandwidth per node and any topology constraints on placement.
Quote egress to public internet, to a peered cloud, and inter-region.

3. Capacity SLA

Capacity guarantee for the reservation term; what happens if you cannot deliver.
Mean-time-to-replace a failed node and historical observed failure rate per node-year.
Quote any liquidated damages on capacity miss.

4. Support tier

List included support tier and the upgrade path to 24x7 named-engineer support.
Confirm response and resolution SLAs for severity-1 incidents during a training run.
List named onboarding-engineering hours included in year 1.

5. Exit terms

Mid-term exit fee schedule.
Data-export support: bandwidth, fees, format for model checkpoints and training datasets.
Notice period for non-renewal and post-term data retention.

6. Hidden line items

Persistent storage rate card (block, object, file) per GB-month.
Snapshot rate card and lifecycle policy options.
Egress price list (per GB, per destination class).
MLOps and monitoring add-ons (Weights & Biases integration, in-product telemetry, log retention).
Idle and warm-pool billing on serverless tiers.

Last verified June 2026.