Rent an H100, the real bill

The short answer

The same H100 SKU is offered under very different commercial structures, from spot capacity on a marketplace to a reserved allocation from a direct operator, and the headline moves by a factor of two or more across them before storage, egress and term are even counted. viabandwidth tracks 224 providers that publish H100 availability, so the task is to size the term to your workload, read what each rate includes and shortlist on structure rather than on the lowest number.

01 · The billThe per-hour rate is the down payment

An H100 quoted low and an H100 quoted high are frequently not the same product. The low quote tends to come from a marketplace passing through spot capacity with no committed availability and the high quote from a direct operator who is also paying for the network, the building, the on-call engineer and the contract that keeps the card there next month. Neither is wrong, they answer different procurement questions.

The bill that lands at month end is the per-hour rate times the hours you run, plus the host CPU and memory if billed separately, plus persistent storage, plus egress every time you move data out. For a training job that reads a large dataset repeatedly and ships a final artefact elsewhere, those extra lines can rival the compute itself.

02 · SourcingWhere H100 rental capacity comes from

Direct operators run H100s on their own infrastructure and sell guaranteed access, which is why their rates are steadier and their availability is real. Marketplaces aggregate capacity from many operators and clear it at a spot price, which is where the cheapest H100 hours live and also where availability is least certain. Resellers sit between, renting from a larger cloud and pricing against their own commitments. The single most useful filter when shortlisting is operator type, because it tells you which promise sits behind the rate.

03 · GeographyWhere the H100 supply sits

Region shapes both latency and availability. The counts below are the number of operators that state H100 capacity in each country, so a residency or latency requirement is a filter rather than a dead end.

The United States leads but does not dominate, and the depth in Europe and India means an H100 shortlist can respect a data-residency or latency requirement without falling back to a single region.

The H100 in the wider accelerator lineup

The H100 is the most widely listed accelerator in the directory, which is why its supply is deep and its rental market is the most competitive. The chart below is the number of tracked providers listing each of the leading accelerators.

Because the H100 is the most contested card, a buyer has real negotiating room here that the scarcer B200 and rack-scale systems do not yet offer. If the workload fits an A100, the supply is nearly as deep and usually cheaper again.

04 · TermChoosing a commitment you can defend

• 01Reserve only what you will certainly run. A one-year or multi-year term brings the effective rate down substantially, so commit the H100 hours you will run regardless.
• 02Layer the variable part on top. Keep on-demand or spot for the work whose shape you cannot yet defend.
• 03Shortlist three or four, not the single cheapest. Ask each, in writing, for the all-in cost of a workload shaped like yours.
• 04Judge the answer, not just the number. How quickly and precisely an operator prices your workload is itself a signal about how it behaves when something breaks.