Start with the job, not the listing
A low advertised GPU-hour rate is only one input to a compute decision. First define the work: training or inference, memory needs, data volume, completion deadline and an acceptable result. Specify whether interruption is tolerable and whether the workload needs a tightly connected cluster. Two offers with the same accelerator label can support materially different jobs.
Write down the software environment, storage requirements and security constraints before searching. Identify a representative benchmark or acceptance run that the supplier can support. Without workload evidence, a cheaper rate cannot establish a cheaper completed job. Record unknowns rather than treating nominal hardware specifications as measured performance.
Distinguish booked time from productive time
Rental charges follow contract terms, not necessarily productive GPU use. Reserved capacity may be billed while data is loading, software is being configured or the application is waiting. Some services offer interruption-sensitive capacity; others require a fixed reservation window. Confirm when billing starts, what stopping an instance changes and whether unused time is refundable.
AWS documents a distinct Capacity Blocks reservation model, including cancellation restrictions. Vast.ai describes separate rental types and cost components. These sources show why billing rules matter; they do not establish that either service fits your workload or has capacity available for your dates.
Build the complete cost boundary
Request the compute unit explicitly: GPU-hour, node-hour or a whole-cluster commitment. Add storage, data transfer, networking, support, setup and applicable taxes where they are not included. Confirm minimum terms, deposits, payment timing and exit costs. Keep one-time deployment effort separate from recurring charges so the comparison can be reused for a longer or shorter project.
Public pricing pages, including Lambda's, provide indications rather than an allocation or binding project quote. Ask for a dated operator-issued offer with hardware, location, quantity, term, inclusions and expiry. Preserve the original billing units instead of silently treating a node price as a GPU price.
Test a range before committing
Calculate a base case and a downside case using your own billable hours and expected useful output. Change utilization, setup time and workload duration separately. A fixed commitment can look attractive at high use and expensive when demand slips. Include engineering effort and migration constraints, but avoid assigning arbitrary monetary benefits just to make one option win.
Before acceptance, agree how delivery and performance will be checked, who resolves a failed run and what remedies the contract provides. Compare keeping the incumbent, renting less or delaying a commitment as real alternatives. A browser calculation is a planning aid; the final decision needs workload evidence and confirmed commercial terms.
Primary sources & further reading
Vast.ai: instance pricing and billingAWS: Capacity Blocks for MLLambda: public GPU cloud pricingIndependent source information is not a ComputeBrokers partnership, live allocation or approval. Confirm current terms before acting.