The headline price is the smallest piece of the bill
An H100 quoted at two dollars an hour and an H100 quoted at six dollars an hour are not the same product. The cheap quote almost always comes from a marketplace passing through spot capacity from an underused operator, with no committed availability and no support relationship beyond the platform. The expensive quote usually comes from a direct operator who is also paying for the network, the building, the on-call engineer, and the contract that guarantees the GPU will be there next month. Neither is wrong, but they belong to different procurement decisions.
The first piece of small print to look at is what the per-hour price actually covers. Many providers price the GPU alone and bill the host CPU, the memory, the network attached storage, and the egress separately. A few providers price the whole machine flat, which looks more expensive at the headline but often lands cheaper at the invoice. The only way to compare is to write down a real month of work and to ask both providers to price it.
Storage and egress are where surprises live
A training workload typically reads a large dataset many times, writes checkpoints continuously, and pushes the final model artefact out to another system. Each of those is metered, and each provider meters them differently. Persistent storage on fast tiers runs from cheap on a direct operator with its own filesystem to genuinely punitive on resold hyperscaler capacity. Egress fees can quietly add a thousand dollars to a model release that involved nothing more dramatic than uploading the final weights.
The diligence trick is to ask for a worked example. A serious provider will quote, in writing, the all-in cost of training a model of your typical shape on their stack. The number they produce, and how long it takes them to produce it, is itself part of the diligence.
Commitment terms swing the price by a factor of two
Spot, on-demand, monthly committed, annual committed, and multi-year reserved each move the per-hour price around by a meaningful amount. A one-year reservation typically halves the headline price relative to on-demand. A three-year reservation can halve it again, at the cost of substantial inflexibility. The right choice depends on how confident you are in the workload shape, and how willing you are to be wrong about it.
The procurement-grade move is to estimate the minimum committed term you can defend (the workload you will absolutely run no matter what) and to layer on-demand or spot on top for the variable component. Almost nobody picks the cheapest headline rate end to end, because almost no real workload is flat.
What the directory shows, and what it does not
viabandwidth lists pricing where a provider publishes it on its own website, and dates the capture. We do not synthesise prices, infer them from competitors, or republish numbers scraped from price-aggregator sites. When the directory shows a number, it is what the provider quoted in public, on the date shown.
What the directory cannot show is your specific quoted price. The number you will actually pay depends on the contract you negotiate, the term you commit to, and the workload pattern you describe. The published price is a starting point for the conversation, not a substitute for it.