GPU as a service means you rent GPU compute over the network and pay for what you use instead of buying the card. That is the whole definition. What you get behind it ranges from a container that scales to zero to a physical server nobody else touches. The label hides the difference, and the difference decides your bill.
I split the market into four shapes. Every provider in the provider index sells one or more of them. Prices below are not quoted here. See the GPU pricing chart for the on-demand hour.
1. GPU virtual machines and instances
A VM or pod with one or more GPUs, SSH access and per-second or per-hour billing. You pick the model, the count and the disk. You pay while it is running, idle or not. This is the default for fine-tuning, experiments and long jobs, and it is what most of the catalog is.
Examples in the index: Lambda, Hyperstack, Vast.ai and the hyperscalers. The trap is idle time. A GPU you forgot to stop is a GPU you paid for.
2. Serverless GPU
Serverless GPU runs your code or container on demand and scales to zero when no one calls it. You pay per second of execution, not per hour of reservation. That is a good deal for spiky inference, webhooks and low-traffic APIs, and a bad one for steady load or long training.
- Good for: bursty inference, demos, batch jobs that finish, pay-per-request endpoints.
- Watch for: cold starts while the model loads, concurrency limits and per-second rates that cost more than a reserved hour at steady load.
- Not for: multi-day training, anything that needs a fast interconnect, workloads that never go idle.
RunPod offers serverless workers alongside pods. Koyeb runs GPUs like a platform. Marketplaces such as Vast.ai have no native scale-to-zero. If cold starts matter, ask about image caching and model preloading before you commit.
3. Bare metal GPU servers
Bare metal means a whole physical server, no hypervisor, no noisy neighbors. You get the full NVLink fabric, local NVMe and, on clusters, InfiniBand. It is the shape for large training runs and for teams that need predictable performance and compliance isolation.
The cost is commitment. Provisioning takes longer, contracts tend to be longer, and the price per hour is often a quote. Specialist clouds such as CoreWeave and Crusoe sell this shape. Do not rank a quoted cluster against an on-demand hour. They are different products.
4. Managed inference and APIs
Some providers rent the model, not the GPU. You call an API and pay per token. That is not GPU rental, it is a different category, and Together AI straddles both. If you need a specific model on a specific card, rent the GPU. If you need tokens, buy tokens.
Which GPUaaS shape fits your job
- Spiky inference with idle gaps: serverless GPU.
- Fine-tunes with checkpoints, experiments, notebooks: GPU VM or pod.
- Multi-GPU training you can restart: VM with NVLink, or a marketplace if you can checkpoint.
- Long pretraining, reserved capacity, compliance: bare metal or a reserved cluster.
- Just need tokens: a managed inference API.
Whatever the shape, compare the same GPU, on-demand, in stock. Spot, waitlists and quotes are not the same hour. See on-demand vs spot GPU cloud and how we rank an hour.
FAQ: GPU as a service
What is GPU as a service?
Renting GPU compute over the network on a pay-for-use basis instead of buying hardware. It covers VMs, containers, serverless and bare metal.
Is serverless GPU cheaper than a GPU VM?
For spiky or low traffic, yes, because you stop paying when idle. For steady load, a GPU VM or reserved hour is usually cheaper per unit of work.
What is a bare metal GPU server?
A physical GPU server dedicated to you, with no hypervisor. It gives full performance and the real interconnect, usually with longer commitments.
Which GPU cloud providers offer serverless?
Several, including RunPod and Koyeb in this index. Confirm cold-start behavior and concurrency on the provider page before you build on it.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.