Direct, marketplace, or reseller: which GPU cloud fits your workload

A direct operator owns the racks and runs the network. When you sign a contract with one, the GPU under the hood is theirs, the IP address that serves your traffic is announced from their own autonomous system, and the on-call engineer who fixes a hardware fault works for the company on your invoice. This sounds obvious until you realise how many GPU offerings do not look like this at all.

The tells that a provider is a direct operator are usually visible from the outside. The same brand owns a registered ASN, that ASN announces meaningful IPv4 and IPv6 prefixes, and the operator is present at one or more named physical facilities. Capacity scales smoothly in small increments rather than in obvious hyperscaler-sized blocks. Support tickets are answered with reference to a specific rack and PSU rather than a generic case number. Pricing is typically the highest of the three models, because the single accountable counterparty is paying for everything end to end.

The workloads that fit this model are the ones that punish hand-offs. A continuous training run, a regulated inference fleet, anything that needs a clear SLA with one phone number behind it.

A marketplace is not a GPU cloud, it is a way to find one. The marketplace aggregates capacity from many independent operators, normalises the catalogue, and presents a single interface for booking. When you launch a job, the marketplace routes it to whichever underlying operator has the matching SKU available, often at the lowest current price. The contract is with the marketplace, the metal is with the operator.

The strengths of this model are breadth, speed, and price discovery. A marketplace gives you a hundred candidate operators on tap, which is the right shape for spot pricing and for jobs that can absorb interruption. The weakness is that the underlying counterparty changes from one booking to the next, which is uncomfortable for anyone whose compliance regime cares which entity actually touches the data.

Marketplaces are a great fit for batch training, hyperparameter sweeps, ephemeral notebooks, and anything that benefits from elastic burst across many regions. They are a poor fit for production inference that needs a stable address and a named SLA.

A reseller is a company that buys hyperscaler capacity wholesale and packages it under its own brand. The GPU is in an AWS, GCP, or Azure region, the network and the storage stack come from the hyperscaler, and the reseller adds its own portal, its own pricing, and often a specialised software layer (Kubernetes operators, ML platforms, fine-tuning workflows) on top.

The strengths of this model are hyperscaler-grade reliability and a friendlier commercial process. Many resellers will negotiate terms a small customer could not get directly from the hyperscaler, and many of them publish honest documentation of which hyperscaler is underneath each region. The weakness is that the resold capacity inherits the hyperscaler's pricing dynamics, including egress fees and committed-use complexity, while adding a margin layer on top.

The tells that a provider is a reseller are subtle. The advertised IP prefixes resolve to hyperscaler ranges, the regions track hyperscaler region names, and the headline pricing is often noticeably better than the equivalent hyperscaler list price (because the reseller is buying at a discount and passing some of it on). When in doubt, ask the provider directly which hyperscaler region a given workload would run in. Serious resellers will tell you.

The directory labels each provider as direct, marketplace, reseller, or pending. The label is derived from independent signals rather than self-declared. A provider is labelled direct only when its own brand owns an ASN that announces meaningful prefixes and is present at one or more independently identified physical datacenters. A provider is labelled reseller when its IP allocations resolve to hyperscaler space or its own site states the underlying capacity source. Marketplace is reserved for providers whose business model is clearly aggregation of independent operators. The pending label is used honestly when none of the three signals is strong enough to make a confident call.

This matters because the label changes how the rest of the profile should be read. A direct operator with two facilities and a small IP allocation is a regional player with real metal. A reseller with the same headline numbers is a packaging brand on top of someone else's network. Both can be the right answer for the right workload, but the diligence path is different.