How to compare GPU cloud quotes
The narrative method behind this checklist.
← Back to Compute College
Compute College
GPU Cloud Quote Comparison Checklist
Eleven concrete fields to compare on every GPU cloud quote before signing.
Plain-English definition
A GPU cloud quote comparison checklist is a fixed list of fields recorded on every quote so two or more offers can be read against each other without confusion. The checklist used here is eleven fields, grouped into accelerator, commercial, capacity, and quote-quality categories. It is a rubric, not a score; no field is given a universal weight because the right weight depends on the workload.
Memory trick: Read every quote as eleven facts in a row, not as one number with a logo next to it. The number is the last column, not the first.
Why it matters
Without a fixed comparison rubric, the lowest-rate quote tends to win attention even when its access terms, networking, or interruption risk would produce a higher completed-workload cost. A checklist forces every offer to expose the same eleven facts, which lets a buyer compare like-for-like and an analyst record offers consistently across providers.
- Accelerator fields name what is being priced, so a buyer is not comparing an H100 SXM cluster against a single L40S box.
- Commercial fields name what the buyer commits to, including minimum spend, term length, change rights, and the rate paid for capacity used outside the commitment.
- Capacity and reliability fields expose access terms that determine whether the quoted hours can actually be delivered when the workload needs them.
- Quote-quality fields make the comparison auditable: who quoted, on what date, and with what assumptions about region and configuration.
Simple example
Quote A offers H100 SXM at $4 per GPU-hour with a 24-month commitment in US-East. It includes 3,200 Gbps InfiniBand, a 99.5% availability SLA, 50 TB of storage, and $0.05 per GB egress. Quote B charges $5 month to month in US-West, with 200 Gbps Ethernet, metered storage and egress, and $2.50 spot capacity without an SLA. For a deadline-bound run on 128 connected GPUs, the buyer must ask whether B can deliver an uninterrupted cluster on time and whether A provides enough flexibility to justify the commitment.
- Quote A is cheaper on rate but binds capacity and region; the commitment is only a saving if utilization stays above the break-even monthly hours.
- Quote B is more flexible but spot risk and lower networking can slow or restart a 128-GPU job, raising effective completed-workload cost.
- A side-by-side row that says "$4 vs $5" hides everything that actually differs; the eleven-field row does not.
- Numbers in the example are illustrative; a real comparison requires verified quote text and a date for each row.
Example figures are illustrative calculations, not current quoted market prices.
Common mistake
Do not compare quotes by rate alone or by a custom score that hides its weights. A single rate or a black-box score collapses eleven different facts into one number and lets the cheapest-looking row win regardless of whether it can do the job. If a score is used, the rubric and its weights must be visible alongside the result.
Practical takeaway
What you can do with this
Record the eleven fields on every quote. Use the checklist below as the column header for a side-by-side comparison sheet, and keep the source attribution and date on every row.
- 1. GPU generation: chip model, board form factor (SXM, PCIe), memory capacity, and connectivity (NVLink, InfiniBand). H100 SXM is not the same offering as H100 PCIe.
- 2. Hourly and effective monthly cost: published rate per GPU-hour and the implied monthly cost at full utilization, plus a row at the expected utilization the buyer can sustain.
- 3. Minimum commitment: term length, minimum monthly spend, prepay structure, and whether unused capacity can be reclaimed or sold back.
- 4. Region and latency: physical region, available zones, end-user latency budget, and whether the buyer can change region during the term.
- 5. Capacity and quota: connected cluster size available now, lead time for the required block, soft quotas, and what triggers a manual approval.
- 6. Networking: interconnect type and bandwidth between GPUs in a node and between nodes, plus any topology constraints on a multi-node job.
- 7. Storage and egress: included high-speed storage, metered overflow, and per-GB egress charges for moving model weights, data, and checkpoints out of the region.
- 8. SLA and support: availability target, remedies if missed, support tier included, response-time commitment, and named escalation path.
- 9. Interruption risk: spot or interruptible eligibility, expected eviction rate by region, checkpointing assumptions, and the contract treatment of forced interrupts.
- 10. Flexibility: rights to pause, change GPU type, change quantity, extend or shorten the term, transfer capacity to another project, and renegotiate before renewal.
- 11. Quote source and date: the named source contact, quote document or page URL, the effective date, and any expiry on the quoted terms.
Decision check: a quote can only be chosen when the eleven-field row is filled in and the gaps named. An empty field is a question, not a permission.
Compute College
Turn the lesson into a number
Use the GPU-Hour Cost Calculator, AI Training Cost Calculator, or Model Serving Cost Calculator.
Compute College track
Buyers & Operators
Step 3 of 8: GPU cloud quote comparison checklist