How to compare GPU rental quotes

Published 1 October 2026 · Updated 22 October 2026

Compare the full allocation, billable time, included services and payment obligations before choosing a GPU rental quote. An hourly rate is only one input. Missing fabric, a later delivery date or a large advance payment can change the decision even when the GPU name is the same.

Put each quote on the same basis

A quote may use GPU-hours, server-hours, node-months or a cluster total. Record GPUs per billable unit, currency, taxes and minimum charge. For monthly billing, use the supplier's actual definition of a month. For hourly billing:

Compute charge = allocated GPUs × billable hours × rate per GPU-hour.

This calculates compute only. Add separately charged services and fees once, and identify what remains unpriced. A fixed allocation may be billed continuously even while your workload is idle.

A hypothetical three-year comparison

All rates and payment structures below are invented for arithmetic. They are not market benchmarks, supplier quotes, discounts available through RackShore or offers of capacity. The allocation is illustrative, not a buyer's order.

Assume 128 nodes with eight GPUs each: 1,024 GPUs, continuously billed for an illustrative 26,280-hour term. This uses three 365-day years and 36 equal 730-hour billing periods, ignoring leap days and actual calendar-month variation. Assume identical hardware, connected fabric, delivery, service scope and currency for this calculation. Real offers must substantiate those assumptions.

The billable allocation is 1,024 × 26,280 = 26,910,720 GPU-hours.

Each row assumes a non-cancellable three-year compute commitment. Monthly payment in C does not imply a right to cancel monthly. B's advance is part of its $59,203,584 total; adding it again would overstate the compute obligation. No refundable security deposit or setup fee is assumed. Storage, transfer, support, taxes and other extras are excluded from every row, so these are not all-in service totals.

A has the lowest nominal compute total in this invented scenario, but also the largest exposure before delivery. The table does not put a value on financing, supplier default, refund rights or lost access to prepaid cash. Those differences need a separate cash-flow and risk assessment; a lower nominal total alone is insufficient.

Utilization changes economics, not the contracted bill

At an assumed 70% productive utilization, the allocation yields 18,837,504 productive GPU-hours. Compute cost per productive GPU-hour is about $2.86, $3.14 and $3.43 respectively. The contracted compute totals stay the same in this example.

Productive GPU-hours still cannot compare different GPU platforms on their own. Measure spend and elapsed time to complete the same workload and quality target, including data loading, checkpoints, interruptions and retries.

Separate prepayment, deposits and fees

For every upfront amount, ask whether it is credited compute, a refundable deposit, a reservation charge or a non-refundable fee. Record when it is due, what triggers invoicing, any expiry of credits, refund conditions and treatment of delayed or failed delivery. Put those terms beside the cash amount rather than hiding them in a blended hourly rate.

Distinguish a price commitment from a capacity commitment. For example, AWS documentation says Savings Plans do not reserve capacity. That specific product illustrates why the actual capacity provision must be checked; it does not describe every supplier's contract. AWS Savings Plans documentation

Add the service and contract differences

Use the same comparison rows for every supplier:

• GPU variant, exposed memory, tenancy, node count and connected allocation
• Region, delivery milestones, acceptance tests and billing commencement
• Inter-node fabric, topology, oversubscription and storage performance
• Persistent storage, snapshots, data ingress and egress charges
• Setup, support, orchestration, licenses, taxes and any disclosed brokerage fees
• Maintenance, interruption rights, replacements and service-credit process
• Full committed spend, invoice timing, renewal, resizing and exit terms

Treat an unconfirmed item as an open question. Do not assume it is included or assign it a zero price. Compare confirmed offers separately from estimates and future capacity proposals. The cluster rental checklist provides a broader technical brief.

Request comparable offers

Send suppliers the same requirements and keep exceptions beside each response. Identify who operates the infrastructure and who signs, bills and supports the service. Use the GPU broker evaluation guide to check the introduction process as well as the quote.

RackShore makes buyer-supplier introductions. The supplier confirms the offer and handles the buyer contract, billing and infrastructure support. Sharing your allocation, region, timing and term starts a discussion; it does not reserve capacity.