Multi-year GPU lease terms explained

Published 24 October 2026

A multi-year GPU lease is a reserved-capacity contract in which a buyer commits to a dedicated GPU cluster for a fixed term, usually 1 to 5 years, in exchange for a lower hourly rate than on-demand pricing. The commitment, the prepayment and the exit terms matter as much as the headline price per GPU-hour.

This guide covers how the terms work, what changes between a 1-year and a 5-year commitment, and which clauses to negotiate before signing.

How a reserved GPU contract is structured

Most reserved GPU contracts combine five elements:

Prices are usually quoted per GPU-hour so offers can be compared, but billing is often monthly for the whole reserved cluster, whether you use it or not.

Why choose a 3 to 5 year term

Longer terms usually earn a lower hourly rate, because the operator can finance hardware against your committed revenue. For large clusters, a multi-year commitment is often the only way to secure a specific allocation and delivery date.

The trade-off is exposure to change. New GPU generations arrive roughly every one to two years, and market rental prices for older models have fallen as newer ones ship. A 5-year contract signed today will still be billing when today's newest GPU is two generations old.

Choose a longer term when:

• your workload needs a guaranteed, dedicated cluster rather than whatever capacity is free;
• you can use the full cluster for most of the term, for training, fine-tuning or steady inference;
• the price difference against a shorter term is large enough to cover the risk of using older hardware later.

How prepayment works

Prepayment is the share of total contract value paid before or at delivery. Operators use it to fund hardware purchases, so larger prepayments can buy a lower rate or an earlier delivery slot.

Example: a cluster of 1,024 GPUs at USD 2.00 per GPU-hour costs about USD 17.9 million per year (1,024 × 8,760 hours × 2.00). Over 3 years that is about USD 53.8 million. A 30% prepayment would be about USD 16.1 million upfront, and a 50% prepayment about USD 26.9 million.

Before agreeing to a prepayment, confirm:

1. Who receives it. It should be paid to the entity that delivers the service and signs your contract.
2. What protects it. Ask for refund terms if delivery is late or fails acceptance tests. Some buyers request escrow, a parent guarantee or a letter of credit.
3. How it is applied. Check whether prepayment is credited against the first months or spread across the term, and whether unused credit is refundable.

Our quote comparison guide works through how prepayment changes the true cost of competing offers.

Clauses to negotiate before signing

• Delivery date and remedies. A firm date, plus credits or a right to cancel if it slips.
• Acceptance testing. Burn-in and performance tests, such as NCCL bandwidth and multi-node training runs, before billing starts.
• Node replacement. Maximum time to replace a failed node or GPU, and credits when it is exceeded.
• Uptime credits. How availability is measured and what you receive when it is missed.
• Hardware refresh or upgrade rights. An option to move part of the commitment to newer GPUs during the term.
• Assignment and resale. Whether you can sublet or transfer unused capacity.
• Termination. What you owe if you exit early, and what happens if the operator fails.
• Price adjustments. Whether power, data center or currency costs can change the rate during the term.

The GPU cluster rental checklist lists the technical requirements to fix before you ask for quotes.

Public reference prices

Published rental prices from 17 providers, checked on 24 October 2026 on each provider's own pricing page or official pricing API, show how much a reserved term moves the rate. Prices are USD per GPU-hour, with each provider counted once.

Reserved figures use each provider's lowest published committed rate, with terms from two weeks to five years. Marketplace and spot prices are excluded.

• Multi-year commitments cut list prices most at the large clouds. Azure's 3-year H100 reservation is 56% below pay-as-you-go (12.29 to 5.40). Google Cloud's 3-year commitment on A3 High is 56% below on-demand (11.06 to 4.86).
• Short reservations save less. Together AI lists H100 at 3.99 on demand and 3.19 for a 91–180 day reservation, 20% lower.
• A commitment is not always cheaper. Lambda's reserved H100 clusters (5.54–6.16) cost more than its self-serve 8-GPU on-demand instance (3.99).
• Large-cloud on-demand H100 runs 1.5–3× the median. AWS lists 6.88, Oracle 10.00, Google Cloud 11.06 and Azure 12.29.
• Multi-year cluster rates are mostly negotiated. CoreWeave, Nebius, Crusoe and Lambda quote long-term reserved pricing through sales rather than on a price page.

Sources: Lambda · CoreWeave · Nebius · Together AI · Crusoe · RunPod · Hyperstack · DigitalOcean · Paperspace · Verda · Jarvislabs · Voltage Park · Oracle Cloud · Google Cloud · AWS · AWS Capacity Blocks · Azure. AWS and Azure figures come from their public pricing APIs, which feed those pages. Rates change often; check the provider before relying on a figure.

Related questions

Is a GPU lease the same as buying GPUs?

No. A lease or reserved contract gives you dedicated use of hardware the operator owns, hosts and maintains. Buying means you own the servers and are responsible for data center space, power, networking and replacement.

Can I get a reserved GPU contract shorter than a year?

Yes. Many providers offer monthly or quarterly reservations, usually at a higher rate than multi-year terms.

How do I request multi-year GPU capacity through RackShore?

Send a capacity enquiry with the GPU model and quantity you need. Term, prepayment and price are discussed by email. RackShore introduces you to operators and resellers whose capacity may fit; the supplier contracts with and bills you directly.