Buy GPUs, commit to reserved capacity or rent on demand: which fits an IT services firm adding AI work?
Quick answer, Nov 18, 2026As of 18 November 2026, match each GPU commitment to the client commitment behind it. Rent on demand while client work could stop. Reserve a block for fixed-date projects. Commit for one or three years only to load clients have signed for that long. Buy once a client signs take-or-pay terms as long as the loan.
An IT services firm that adds AI work for its clients needs GPU capacity before it knows how long those clients will stay. Each way of getting that capacity answers one question differently: who pays for the hours nobody uses. The terms below were read on 18 November 2026 from the cloud providers' own documentation and pricing pages, CoreWeave's 10-Q for the quarter ended 30 September 2026 and a lender guide. No prices are quoted here; the sources set out structure, not rates.
Four ways to get capacity, and who carries the idle hours
| Path | What you commit to | Who pays for unused hours | Source |
|---|---|---|---|
| Rent on demand | Nothing beyond the hours you run; CoreWeave describes on-demand instances as the choice "for additional workloads when you're not looking for long-term capacity commitments", and its filing calls on-demand "pay-as-you-go" | The provider | CoreWeave pricing, "On-demand GPU instances"; CoreWeave 10-Q, Part II, Item 1A, "We have a limited history selling access to our platform…" |
| Reserve a fixed block | AWS Capacity Blocks run "one to 14 days or a multiple of 7 days, up to 182 days"; the price "is charged up front at the time the reservation is made", and "Capacity Block cancellations aren't allowed" | You, paid in advance; "There is no additional charge for unused time" | How Amazon EC2 Capacity Blocks work; Capacity Blocks pricing and billing; Capacity Blocks for ML, "Considerations" |
| Commit for a term | Google resource-based commitments, which cover GPUs, run "either 1 year or 3 years"; an AWS Savings Plan is "a commitment for one or three years"; CoreWeave's committed contracts "are take-or-pay" and "typically include a prepayment" | You, for the whole term: Google bills "regardless of whether or not you use those resources" and "You can't cancel a commitment after its purchase" | Google Cloud, CUDs for Compute Engine, "Resource-based CUDs"; AWS, "Returning a purchased Savings Plan"; CoreWeave 10-Q, Part II, Item 1A |
| Buy the hardware | A loan or lease payment every month; "A traditional loan needs 2-3 years of take-or-pay" from a customer | You, plus what the hardware is worth when the term ends | GPU Financing 101, "1. Get a strong offtake commitment" |
The middle rows vary in how they can be unwound. An AWS future-dated Capacity Reservation lets you cancel, "but a cancellation charge might apply depending on when you cancel", and during the commitment "you can't decrease the instance count below the committed instance count" (Concepts for Amazon EC2 Capacity Reservations, "Start date and time"). Some providers do not publish reserved terms at all: Lambda says "Contact us for reserved capacity" (Lambda, "AI cloud pricing"). Read the order form before you count on an exit.
Which one fits your situation
- A pilot, or client work that can stop on short notice. Rent on demand. The provider keeps the risk of idle hardware, and you pay the higher rate for that.
- A fine-tuning or migration project with fixed dates. A reserved block, sized to the statement of work. You pay for the block whether or not the client's data arrives on time, so tie the start date to a client deliverable.
- A base load of clients on signed annual or multi-year terms. A one- or three-year commitment, sized to that signed base and no higher. Serve growth above it on demand until new contracts are signed.
- One client signs a take-or-pay contract for two to three years. Owning becomes financeable, and at that point you are running a GPU cluster as well as a services firm. Read buy GPUs or rent them for whether to own at all, and a GPU loan, a lease or vendor financing for how to pay for it.
Where the mismatch sits
The risk sits in the gap between what your clients sign and what you sign upstream. CoreWeave reports that committed contracts were 98% of its revenue in the quarter ended 30 September 2026, and that the weighted-average prepayment across its active contracts was 15% to 25% of total contract value at the end of 2023, 2024 and 2025 (10-Q, Part II, Item 1A). CoreWeave sells almost entirely on commitment, so a services firm that commits upstream takes on a take-or-pay obligation that its own client agreements may not match.
Illustrative only: a client signs a 12-month managed AI contract that it can end on 90 days' notice, and you buy a one-year commitment to serve it. If the client gives notice in month three, its contract ends in month six, and you pay for months seven to twelve with no revenue behind them.
Three client terms close most of that gap: a minimum term at least as long as your commitment, an early-termination fee that covers the commitment still owed, and a prepayment like the one your provider asks of you.
Where Frontier Desk fits
In the Cybernetic Age an IT services firm sells compute as part of its work, and the capacity contract behind it belongs in the firm's financial plan. Frontier Desk is an advisor: it helps firms set the capacity they commit to against the contracts they hold, and it brings in lenders once a client contract can carry owned hardware. Lenders read the contract first (how lenders underwrite a GPU cluster loan), the kinds of lender are in who lends against GPUs, and what they hold if a customer weakens is in what a GPU lender holds if offtake weakens. Start with the GPU financing tool on /compute or a free first look.
Questions
Can my capacity commitment help me borrow against GPUs later? No. A commitment to a cloud is a cost. What a lender counts is the revenue your clients have signed for: "Lenders finance contracted revenue. On-demand and marketplace rentals don't count" (/compute).
If a client leaves, can I move the committed capacity to another client? Only as far as the terms allow, so check where the commitment can be used. A Google resource-based commitment covers resources "only in the region that was specified by the commitment". Whether it covers projects beyond the one that bought it depends on the billing account's CUD sharing setting, which is on by default for billing accounts created on or after 16 June 2026 (CUDs for Compute Engine, "Location- and project-level scopes"). An AWS Capacity Block "can't be moved or split" (Capacity Blocks for ML, "Considerations").
Can I return a commitment I bought by mistake? Only within narrow limits. AWS lets you return a Savings Plan "purchased in the last seven days and in the same calendar month", subject to a size limit and a quota (AWS, "Returning a purchased Savings Plan"). Google says "You can't cancel a commitment after its purchase" (CUDs for Compute Engine, "Resource-based CUDs"). Treat a signed commitment as final.
What if I sell the firm with a commitment still running? A buyer reads the commitment alongside your client contracts, and the closer the two match, the less there is to explain. How AI changes what a services firm is worth covers what buyers check.