H100 GPU rental by the hour, compared per model

The H100 is the datacenter tier of the GPU market - the card every training budget and serious inference deployment measures against. On this site's population it rents from $2.56/hr at the floor, and every claim about "cheap H100s" needs the three questions the sticker never answers: which card variant, what interconnect, and what verifies the work. This page is the datacenter-tier dossier: the variant question, the interconnect bill, the supply camps, and the verification standard - in the order a deployment actually meets them.

SXM versus PCIe: the variant question

The H100 ships in variants that price differently and behave differently [1]:

• SXM modules on NVSwitch fabrics - the multi-GPU configuration: training-scale jobs need the interconnect, and listings that do not name the variant are quoting a lottery.
• PCIe cards - cheaper, common on marketplaces, fine for single-card inference. The wrong choice for anything spanning cards.
• MIG slices - partitioned H100s at fractional rates; a legitimate product and a legitimate discount, as long as the listing says so.
• H200 adjacency - the H200's larger VRAM changes the fit for big models; where listings offer both, the llm-inference hub VRAM tiers decide which card the workload actually needs [2].

The workload tool sizes the model first; the variant question is the second line of every H100 budget.

The feed's flat line is itself information: at this young stage, the floor price is price discovery - one documented provider sets the floor, and the spread will develop as more hosts list. Watch it in the gpu-prices page rather than extrapolating from a flat week.

The multi-GPU bill: interconnect is the second line

Once a job spans cards, the fabric between them is part of the price, and the H100 tier is where that bill is biggest [1]:

• NVSwitch fabrics - SXM configurations give every card high-bandwidth access to every other card; training-scale jobs need it.
• PCIe between cards - the commodity fallback; fine until the model spans cards, then the fabric becomes the bottleneck.
• The honest sizing - fewer well-connected cards beat many poorly-connected ones at equal budget. The workload tool sizes the model; the fabric question is the second line of the answer.

This is also why H100 pricing resists simple per-hour comparison: two listings at the same rate can deliver materially different finished-job times, and the gpu-prices page cannot see the fabric from the price alone.

The operational checklist before scale

Same loop as every GPU tier, at datacenter stakes [1][2]:

1. Size the model in the workload tool - fit decides the tier before price does.
2. Name the variant - SXM, PCIe or MIG slice; an unnamed H100 is a lottery ticket.
3. Price the interconnect into the budget if the job spans cards.
4. Verify the tier you will run - attestation on the cheapest tier is the question that sorts the population.
5. Rent short. The flat feed line is discovery, not stability; a monthly re-price is the maintenance cost of any H100 position.

Where H100 supply lives, priced by camp

The compute hub splits the population into three camps; for H100 specifically:

• On-chain lease markets - documented rates with public utilization data [1]; the $2.56/hr median in our feed is the camp's pricing signal.
• Community marketplaces - widest hardware spread, host reputation, preemption pricing [2].
• Token-incentivized networks - supply subsidized by emissions; fine for burst, budget for the token loop [2].

Live from our database - caveats and sources in each linked review.

Verification, datacenter edition

The verification ladder applies with higher stakes at this tier - an H100-hour costs an order of magnitude more than a commodity hour and the jobs are longer:

The verification standard at the datacenter tier

The compute hub's verification ladder applies with higher stakes here, because an H100-hour costs an order of magnitude more than a 3090-hour and the jobs are longer:

1. TEE attestation where compliance lives [1].
2. Checkpoint sampling for training and batch inference [2].
3. Bonded reputation as the market default - hosts stake against failures [2].
4. A dashboard screenshot - not a verification method, at any tier.

The llm-inference hub maps the same ladder for inference workloads specifically, where sampled outputs make verification cheaper than on training runs.

21 providers document H100-class capacity on this page; 86% of all rows carry facts we verified with dates. The honest H100 posture at current pricing: rent short, verify the variant, price the interconnect - and let the a100 hub and the commodity rtx-4090 hub price the tiers below before assuming the datacenter tier is the right fit. The entry spread matters: at $0.26 median entry across the population, testing an operator costs a few GPU-hours, not a procurement cycle - the beginners hub minimum-position logic applies verbatim.