Marketplace or managed GPU cloud

Two ways the same card reaches you

Choose a marketplace when price and card choice matter more than consistency, and a managed GPU cloud when you need the same machine, network and support every time. A marketplace lists hosts owned by many different operators, and the rate is set by those owners and by demand. A managed cloud owns or leases its fleet, sets one price for each instance type, and answers for the machine when it fails. They differ most in who answers when a machine fails.

Both kinds sit in the GPU cloud category, and the compute layer lists 108 live products in all. The category name does not tell you which model a provider follows, so the first job is to read how its pricing page describes the hosts.

The reason it matters is what happens when something goes wrong. On a marketplace you deal with a host you may never hear from directly, through the marketplace's own rules. On a managed cloud you deal with the provider that owns the machine. Know which of those two conversations you are signing up for before the first launch.

Signs you are looking at a marketplace

The listing shows individual hosts rather than instance types. Each host has its own rate, its own location and its own disk and network figures, and some marketplaces show a reliability score or uptime history per host. Two listings for the same card can sit side by side at different prices because different people own them.

Consumer cards can appear here. For the RTX 4090 24GB, the RTX 4090 page shows sellers with a listed rate: 14; cheapest per GPU-hour: $0.12; dearest: $0.92. Each rate links to the page it was read from, which shows whether the seller lists hosts or instance types.

Prices move. A rate you saw this morning may not be the rate when you launch, and an interruptible listing can be outbid. Read the rate at the moment you start the instance, not from a bookmark.

Signs you are looking at a managed cloud

The page lists a small set of instance types, each with a fixed card count, CPU, memory and disk, and one price per type per region. There is a support plan, a status page, documentation for networking and storage, and usually reserved terms for those who want them.

The trade is that the price covers more than the card. You are paying for someone to replace a failed host, for a network that behaves the same tomorrow, and for an account team if you are large enough to have one. For several nodes linked by a fast interconnect for training, ask a managed provider; a marketplace listing describes one host, and linking hosts is something to ask about rather than assume.

Ask for the instance types in writing, with the card, the CPU, the memory and the disk for each. That list is what you will set against a marketplace host's listing, so keep the date you received it beside it.

Questions that separate them for your job

What happens when the host dies mid-run? On a marketplace you may be moved to a different owner's machine with a different disk; on a managed cloud ask whether the provider replaces the instance and whether the local disk survives. Either way, keep data you cannot lose on storage that is not tied to one host.

Who can see your data? A marketplace host is someone else's machine, and the provider's terms will say what isolation is offered. Read that section before the price section when customer data is involved.

How much setup can you absorb? A managed cloud often gives you images, drivers and volumes that work together. A marketplace often gives you a container on a host and leaves the rest to you, which is fine for a quick fine-tune and tiring for a larger team that has to keep many machines alike.

Moving between the two later

Many teams start on one model and move to the other, so build for the move from the first day. Keep the environment in a container image you control, keep data in object storage you can reach from anywhere, and keep the scripts that launch an instance in version control. Then a change of provider is a change of a few lines and a data copy, not a rebuild.

Watch for things that tie you to one side. A managed cloud's own volume format, its private network or its job scheduler can make leaving expensive. A marketplace host with an unusual driver version can make a container that works there fail elsewhere. Write each one down as you adopt it.

Comparing a rate from each

Put both on the same unit before you compare: price per GPU per hour, on-demand, for the exact card. A managed cloud that quotes per instance needs dividing by the card count, and a marketplace host with a lower rate but slower disk can take longer to finish the same job. The rental cost calculator works from the published hourly rate and adds nothing for those differences.

Using both is common. Prototype on a marketplace where a consumer card at a low rate is enough, then move the work that has to stay up onto a managed provider once you know its shape. Keep your container and your data portable so that the move costs a day and not a month.