GPU compute is access to processing capacity.
When you rent GPU compute, you pay for access to machines that perform computational work. You usually do not own the GPUs. You receive the right to use a specified resource under an agreement that defines its duration, price, and access conditions.
GPUs perform many mathematical operations in parallel. That makes them useful for AI training, fine-tuning, inference, and other parallel workloads. The right resource depends on what you are trying to run.
Know the unit you are comparing.
A GPU-hour is one GPU made available for one hour. A node-hour is one whole server made available for an hour. A node may contain multiple GPUs, so a node-hour price and a GPU-hour price are not directly comparable.
For example, eight GPUs at an illustrative $2 per GPU-hour cost $16 per hour before other charges. Over 24 hours, that is $384. These figures explain the calculation; they are not a Corgi quote. A fixed reservation may remain payable even during idle periods.
A model name is only part of the specification.
Two offers with the same GPU family can differ in memory, hardware variant, networking, CPU, storage, and delivery conditions. Those differences can change whether a workload fits, how quickly it runs, and what it costs to finish.
- GPU model, hardware variant, and memory per GPU.
- GPU count and the number of GPUs in each server.
- Interconnect within a server and networking between servers.
- CPU, system memory, storage, and data-transfer allowances.
- Region, access method, software environment, and support.
A collection of GPUs is not always a cluster.
A multi-node training workload often needs machines configured to communicate with each other efficiently. The same number of GPUs scattered across unrelated locations is not equivalent to a connected training cluster.
Describe the workload and its networking requirements in the request. If you need a specific scheduler, container setup, or bare-metal environment, include that as well.
Compare the cost of completed work.
The cheapest hourly rate does not always produce the lowest cost. A slower run, data movement, downtime, or unused committed hours can outweigh a low headline price.
Before agreeing to an offer, check the full reservation cost, how additional charges are calculated, when access is delivered, and what remedies apply if the service does not meet the agreement.
Further reading
NVIDIA: accelerated computingPrime Intellect: multi-node cluster deploymentThese sources provide technical or competitor examples. Their product terms and capabilities are not Corgi commitments.