A GPU marketplace sells you a machine: a GPU instance or a bare-metal server from one of many providers, found through one search and billed through one account. Managed GPU Kubernetes gives you a cluster: machines from those same providers joined into one Kubernetes environment, with drivers, networking, scheduling, and scaling already handled. Both make cloud GPUs easier to get. They solve different parts of the problem.
During a GPU shortage, the marketplace part gets most of the attention, because finding capacity is hard. But renting the GPU is the first day of work, not the last. This post explains what each model covers, and where teams usually end up needing both.
What a GPU marketplace gives you
GPU marketplaces come in a few forms. Aggregators put many GPU clouds behind one console or API: Shadeform, for example, describes itself as a GPU cloud marketplace across more than 30 clouds, with centralized billing and reserved capacity on request. Peer marketplaces such as Vast.ai let hosts rent out their GPUs and let renters pick offers by price, reliability, and location. Compute exchanges such as SF Compute let tenants reserve cluster capacity and sell idle time back to others.
What they share is price discovery and availability. Instead of checking a dozen providers for H100 or B200 capacity, you search once and launch where the price and location fit. For a one-off experiment, a fine-tuning run, or a researcher who needs a GPU this afternoon, that is often all you need.
What you still build after renting the GPU
The marketplace hands you a machine with a GPU. Everything between that machine and a production workload is yours:
| Job | GPU marketplace | Managed GPU Kubernetes |
|---|---|---|
| Find and rent GPU capacity | Yes | Yes, through providers you choose |
| NVIDIA driver and container toolkit | Depends on the provider’s image | Installed and kept current on every node |
| Cluster and control plane | You build it | Managed |
| Networking between machines from different providers | You build it | Encrypted overlay included |
| Scheduling jobs onto the right GPU | You script it | Kubernetes scheduler with GPU model labels |
| Scaling up and down with demand | Manual or your own automation | Node auto-provisioning |
| Sharing a GPU between small workloads | Not covered | Time slicing and MPS |
| Moving a workload to another provider | Rebuild on the new machine | Change a node selector |
| Monitoring, upgrades, and access control | Yours | Part of the platform |
The table explains a pattern many AI teams go through. They start on a marketplace, run a few training jobs, then find themselves maintaining scripts that install drivers, copy data, restart jobs, and track which machine runs what. At that point the problem is no longer finding a GPU. It is running a platform on top of rented machines.
When each model fits
A marketplace fits short, independent jobs: experiments, benchmarks, a single fine-tuning run, or burst capacity for a batch job that can run anywhere. You pay for the machine while you use it, and nothing needs to outlive the job.
Managed GPU Kubernetes fits work that has to keep running: training pipelines that use many GPUs across nodes, inference services that scale with traffic, and teams that share a pool of GPUs with quotas and priorities. It also fits when capacity comes from several providers at once, because a single cluster hides the differences between them from the workloads.
Many teams use both: a marketplace to find capacity, and a cluster to run on it.
Where Cloudfleet fits
Cloudfleet is not a GPU marketplace. It manages the cluster: the control plane, node auto-provisioning across providers, NVIDIA drivers and container toolkit, encrypted networking between locations, GPU sharing, and Kubernetes upgrades. GPU nodes from AWS and Google Cloud are added by node auto-provisioning when Pods request them, and servers from GPU clouds and your own data center join through the Cloudfleet API, Terraform, or a single CLI command. Your workloads pick a GPU with labels such as cfke.io/accelerator-name and cfke.io/provider, and one cluster can run GPUs from several providers side by side.
On the capacity side, the Cloudfleet team works closely with most infrastructure providers and helps with procurement and finding the right deal, from a few on-demand GPUs to reserved clusters. You pay the providers for compute directly, with no markup. For how to choose between those buying models, see GPU procurement in a shortage. The GPU Kubernetes page lists current H100, H200, B200, and B300 offerings across more than 20 providers, where to rent NVIDIA B300 covers the newest GPUs in detail, and multi-cloud Kubernetes explains how a cluster spans them. For running training and inference together on that capacity, see training and inference in one Kubernetes cluster.
Frequently asked questions
What is a GPU marketplace?
A service that lists GPU capacity from many providers or hosts in one place, so you can compare prices and availability and rent a GPU instance or server through one account. Some aggregate GPU clouds, some let individual hosts rent out GPUs, and some trade reserved cluster time.
What is GPU as a service?
Renting GPU compute on demand instead of buying hardware. It covers GPU instances from hyperscalers and specialized GPU clouds, marketplaces that aggregate them, and managed platforms that run workloads on rented GPUs.
What is cloud GPU computing?
Running GPU workloads on GPUs in a provider’s data center, billed by the hour or reserved for a term. Typical workloads are model training, inference, rendering, and scientific computing.
Is Cloudfleet a GPU marketplace?
No. Cloudfleet manages Kubernetes clusters that run on GPUs from the providers you choose, and the team helps with procurement through its relationships with most infrastructure providers. You pay providers directly for compute, with no markup.
Can one Kubernetes cluster use GPUs from different providers?
Yes, if the platform connects nodes across providers. In a Cloudfleet cluster, GPU nodes from AWS, Google Cloud, GPU clouds, and your own servers join one cluster over an encrypted network, and workloads choose a GPU model or provider with node labels.
