Buy GPUs or rent them? What the decision turns on
Quick answer, Nov 2, 2026As of 2 November 2026, rent while your demand is still moving and buy once a customer has signed for the capacity. Rented capacity is published by the hour and stops costing you the hour you stop using it. An owned cluster carries a fixed payment every month, and lenders will not underwrite on-demand rates.
What each path publishes today
| Path | What the public figures say | Source |
|---|---|---|
| Rent on demand, priced per GPU | NVIDIA H100 SXM at $3.99 and B200 SXM6 at $6.69 per GPU an hour on an eight-GPU instance, plus tax | Lambda |
| Rent on demand, priced per node | HGX H100 at $49.24 an hour and HGX B200 at $68.80 for the eight-GPU node; the same nodes at $19.71 and $34.11 on spot | CoreWeave |
| Rent on a commitment | CoreWeave puts reserved capacity at "up to 60% off" and quotes it through sales; Lambda says "contact us for reserved capacity at our lowest prices" | CoreWeave, Lambda |
| Lease the hardware | A lessor buys the servers and rents them to you, so it asks for less cash up front than a loan | /compute |
| Buy with a loan | An advance of 50 to 80 percent of equipment cost, a debt-service coverage covenant of 1.15 to 1.25 times, and a rate of 7 to 15 percent set by the customer behind it | GPU Financing 101, American Compute, updated June 2026 |
| Buy with onchain credit | Up to 80 percent of hardware value against at least 20 percent of your own, over three to five years | USD.AI |
| What the hardware is assumed to earn over | CoreWeave depreciates technology equipment over six years, having raised the estimate from five effective 1 January 2023 | CoreWeave 10-K, FY2025 |
Every figure above was read from the page it links to on 2 November 2026.
The question that settles it
Has a customer signed for the capacity, and for how long?
Renting is the market price of flexibility. You pay the published rate while you use the machines and nothing when you stop, and someone else carries the hardware when a new generation lands. Owning turns a variable cost into a fixed one: the payment arrives every month whether the cluster is busy or idle.
A lender reads it the same way: "Contracted revenue is fixed by the offtake agreement; on-demand and spot rates swing with the market, which a lender cannot underwrite" (GPU Financing 101). That guide says a traditional loan wants two to three years of signed commitment behind it, and nine months at the shortest under a residual-value-insured structure.
The contract that makes owning financeable is the same thing that makes owning worth doing. Without one you take market risk on a depreciating asset against a payment that does not move.
What owning asks for that renting does not
- Cash in the deal. Lenders look for 30 to 50 percent of equipment cost from you, your investors or a customer deposit, plus about six months of operating costs in the bank, and closer to half from a first-time operator (GPU Financing 101).
- A site. A purchase quote with the full bill of materials and a colocation quote from a Tier III data center turn an order into a project a lender will read (/compute).
- A view on what the hardware is worth at the end. Six years is the life CoreWeave books. Where an onchain loan carries residual-value cover, USD.AI says it "may cover part of a shortfall on qualifying server sales" (USD.AI).
- A term that fits the contract. Lenders rarely go past three years, and a contract under two years is priced like no contract at all. How lenders underwrite a GPU cluster loan covers the order they read a project in.
When renting is the right answer, and what sits between
Rent while the workload is still being designed, while you need a generation of hardware for weeks rather than years, and while your own customers buy month to month.
Two paths sit between renting and owning. Reserved capacity commits you for a term and the cloud discounts the rate, with the hardware staying on its balance sheet. An operating lease puts the servers in your racks on a lessor's balance sheet, asks for less cash up front than a loan, and ends with you returning them, extending, or buying them at market value (/compute). Renting now and buying later is a normal sequence: lenders finance a record of smaller clusters before a large one.
What Frontier Desk does
The GPU financing tool on /compute reads a project the way a lender would and names the kinds of financing that fit it, with the lenders that do each kind linked to their own pages. Who lends against GPUs names them by kind.
Questions
Is it cheaper over six years to own than to rent? There is no public answer. On-demand rates are published; what you would pay for hardware, power and colocation is quoted privately, and the result moves with how many hours a year the cluster earns. A general break-even figure is an extrapolation.
Can I finance a cluster I plan to rent out on demand? Rarely on a term loan. Lenders discount utilization assumptions and read the signed contract first (GPU Financing 101).
My customer will only sign for a year. What then? Rent or lease that year. A one-year commitment is shorter than the term a lender will write against it, so buying on it leaves the refinancing risk with you.