A100 GPU rental by the hour, compared per model

The A100 is the value tier of the datacenter GPU market: the previous generation's flagship, still serving inference and mid-scale training at roughly half the H100 tier's rate - $0.84/hr at the floor in our feed. 15 providers document A100-class capacity here, and the honest framing is "more supply than the H100 tier, thinner pricing discovery than the commodity tier."

Why the A100 tier still exists at all

Post-H100-generation, the A100's story could have ended; instead it became the value tier - and the reason is supply behavior [1][2]. The installed A100 fleet is large, its owners rent it out freely, and H100 demand concentrated on training-scale work. That leaves the A100 serving exactly the workloads where its VRAM-to-rate ratio wins: inference and mid-scale training. This page prices that tier honestly - variant, fabric, camps, checklist - because "cheap datacenter GPU" is a claim that needs all four answers.

The variant question, A100 edition

The A100 ships in 40 GB and 80 GB variants that price differently and fit differently [1]:

• 40 GB - the mid-model sweet spot for 13-30B-class inference at FP16, and quantized fits above it.
• 80 GB - the fit for 30-70B models without heavy quantization, and for longer-context serving where KV cache eats the budget.
• MIG slices - legitimate fractional products; ask which the listing is, per the h100 hub's variant question.

The workload tool sizes the model first; the variant is the second line of the A100 budget - the same structure as every tier on this site, with the memory split being the variant that matters most here.

The flat feed line, like the H100's, is price discovery: a documented floor from a single camp, with the spread yet to develop. The gpu-prices page is where that spread shows up.

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

The multi-GPU and fabric question

The A100's fabric story is quieter than the H100 tier's but real: NVLink pairs and multi-GPU nodes exist on both variants, and the same rule from the h100 hub applies - fewer well-connected cards beat many poorly-connected ones at equal budget [1]:

• NVLink pairs - the common multi-card A100 configuration; fine for model-parallel inference.
• PCIe listings - cheaper; wrong for anything fabric-bound.
• The honest test - a fabric-limited job on cheap cards loses to fewer connected cards; the workload tool sizes the model before the fabric question matters.

Where the A100 wins - and where it does not

The honest tier comparison, priced per finished job rather than per hour - and the boundary cases are where the llm-inference hub VRAM tiers do the deciding:

• Wins: inference and fine-tuning at mid-model scale, where its VRAM-to-rate ratio beats the H100 tier and its reliability beats commodity cards [2].
• Loses: interconnect-bound training - the H100 tier's NVSwitch fabrics price that gap; the h100 hub carries the fabric arithmetic.
• Loses on per-VRAM price only rarely: quantized commodity cards close the gap on small models - the rtx-4090 hub prices that boundary.
• The tie-breaker is reliability per finished job. A100-class hosts report fewer silent-failure patterns than commodity marketplaces, which matters on long batch jobs even at a higher rate - the compute hub verification ladder prices that difference.

The operational checklist before scale

Same loop as every tier, at value-tier stakes [1][2]:

1. Fit first - the workload tool with your real context length; fit decides the tier.
2. Name the variant - 40 or 80 GB, whole card or MIG slice.
3. Price the crossover - per-token versus dedicated, per the llm-inference hub replay method [1][2].
4. Verify the tier - the attestation ladder applies at double stakes on datacenter cards.
5. Rent short - the flat feed line is discovery; a monthly re-price is the position's maintenance cost.

1. Fit first - the workload tool with your real context length [5].
2. Name the variant - 40 or 80 GB, whole card or MIG slice.
3. Price the crossover - per-token versus dedicated, per the llm-inference hub replay method [1][2].
4. Verify the tier - the [attestation ladder](#) applies at double stakes on datacenter cards.
5. Rent short - the flat feed line is discovery; a monthly re-price is the position's maintenance cost.

15 providers make the live table; 87% of all rows carry facts we verified with dates. The A100's honest role in this market: the tier where mid-model workloads get datacenter reliability without datacenter-tier pricing - and where the compute hub's interruption discipline is the discount's price. A closing arithmetic note: entry prices across this population start around $0.03, and the cheapest hub effective-price rule applies at datacenter stakes - the cheapest A100 headline and the cheapest finished-job rate are different numbers, and the fabric decides which is which.