Buying GPU capacity looks like buying cloud. At scale it is closer to signing a lease: a large sum, a long term and a counterparty you will depend on for years. This checklist covers what to settle before you sign.
1. Know the workload
- Training or inference. Training needs a large cluster on one fast network. Inference needs capacity close to users, often in smaller blocks in more places.
- Size. How many GPUs you need now, and how many in 12 months.
- The chip. H100, H200, B200 or the next generation. The newest chip is not always the cheapest per unit of work. Price the work, not the chip.
2. Specify the cluster, not just the chip
- The network between GPUs. Ask what it is, how fast it is, and whether every GPU can talk to every other at full speed.
- Storage and data transfer. Ask how fast the storage is and what it costs to move data in and out.
- How you get it. Bare metal, virtual machines or a managed service. Each one moves work between you and the provider.
CCompute · the cluster, end to endExample data
GPUs1,024 H100 SXM, 128 nodes
NetworkNon-blocking, every GPU to every GPU
StorageShared filesystem, priced per GiB-month
Power1.6 MW firm to the hall
CoolingLiquid to the rack
The cluster is the whole hall, not just the chip. Example data.
3. Choose the term
On-demand capacity is flexible and dear. A reservation of one to three years is cheaper per hour, but you carry the risk if prices fall. On 8 October 2026 published H100 rates ran from $1.99 to $9.80 a GPU-hour across providers that state one, so the term you choose matters as much as the provider. How GPU deals actually close covers what moves the price.
4. Pin down the place and the date
- Region. Where the data may sit by law, and how far it is from your users.
- Delivery. Is the capacity live, or still being built? Capacity in a hall that is not yet powered is a promise. Ask which data centre, what power it has and when it hands over.
- Acceptance. Agree a test the cluster must pass before you start paying.
5. Read the contract
- Service levels. What counts as down, how it is measured and what you get back.
- Failures. How fast a failed GPU or node is replaced, and at whose cost.
- Money up front. How much is prepaid or held as a deposit, who holds it, and what happens to it if the provider fails.
- Exit. How you leave early, whether you can pass the contract to someone else, and whether you can resell capacity you do not use.
- Price changes. Whether the rate can move during the term, and on what basis.
CCompute · who owns whatExample data
LayerOwnerYour question
GPUsGPU cloudAre they financed?
BuildingData centre operatorWhen does it hand over?
PowerUtility contract holderIs the supply firm?
Often three different companies. Example data.
6. Check the counterparty
- Ownership. Who owns the GPUs, who owns the data centre and who holds the power contract. These are often three different companies.
- Financing. If the GPUs are financed, a lender may have rights over them that come before yours.
- Track record. Whether the provider has delivered a cluster of this size before, on time.
- Financial health. Whether the provider can carry your contract for its full term.
The checklist on one page
- What will the GPUs do: training or inference?
- How many GPUs now, and in 12 months?
- Which chip gives the lowest cost for the work?
- What network joins the GPUs, and at what speed?
- What do storage and data transfer cost?
- What term fits the risk you want to carry?
- Where must the data sit?
- Is the capacity live, and if not, when?
- What test must the cluster pass before payment starts?
- What happens to prepaid money if the provider fails?
- Who owns the GPUs, the building and the power?
- Is the quote in line with what buyers of your size pay?