Choose an exact H100 variant first
H100 is a family name. The rows can differ in exact variant and rental terms, so a family-wide floor is an entry point for research rather than a recommendation for every workload. Open the exact-variant listing and check device memory, software precision support and any required GPU interconnect.
Separate fit from the rental model
Measure memory at the intended batch size, sequence length and precision. For a multi-GPU plan, confirm the sold configuration and its topology; adding device memory does not by itself prove that the application can use it as one pool. Compare on-demand offers for interruption-sensitive work and test checkpoint recovery before choosing spot.
Build a shortlist you can actually launch
Use the table’s pricing model and reported availability, then confirm the exact configuration at checkout. Run a bounded representative job and compare billed cost per completed task. Do not use a lower family-level hourly rate to infer faster completion or equivalent deployment terms.
Limitations
Family-level prices combine different exact variants. Collected rates and reported availability can change, and hardware specifications do not establish measured workload performance.