The same H100 costs $1.73/hr on one provider and $6.16 on another. Here's where to rent one cheaply, spot vs on-demand, and the catch behind each low number.
This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
The H100 is the GPU everyone actually wants to rent — 80GB of fast memory, enough to run a 70B model or fine-tune seriously — and the price you pay for it varies more than any other card. As of our latest capture on 22 Sep 2026, the cheapest H100 rentals are about $1.73/hour on Vast.ai's spot marketplace and $1.99/hour on RunPod on-demand, while the enterprise neoclouds charge $3.85 to $6.16 for the identical chip. That's a 2–3.5x spread for the same silicon. Here's where to actually rent an H100 cheaply, and the catch behind each low number. This is a spoke of our cloud-GPU pricing pillar.
The cheapest H100 rentals right now
The Vast.ai and RunPod figures below come straight from our own price tracker, captured 22 Sep 2026; Vast is a spot marketplace, so I show the cheapest live offer (the median runs higher). The neocloud rates are each provider's published on-demand pricing, cross-checked in the pillar. One number to hold onto: the cheapest reliable on-demand H100 is under $2/hour — anything much above that is buying you enterprise guarantees, not speed.
H100 rental · $/GPU-hour · captured 22 Sep 2026
Vast.ai (SXM)
Provider
$1.73
Rate (from)
Spot
Type
Interruptible; variable hosts
Hyperstack
Provider
$1.95
Rate (from)
On-demand
Type
Managed; per-minute billing
RunPod (PCIe)
Provider
$1.99
Rate (from)
On-demand
Type
Stable; the safe default
RunPod (SXM)
Provider
$2.69
Rate (from)
On-demand
Type
Faster interconnect
Nebius
Provider
$3.85
Rate (from)
On-demand
Type
Enterprise SLA
Lambda
Provider
$3.99
Rate (from)
On-demand
Type
1x available; training-grade
CoreWeave
Provider
$6.16
Rate (from)
On-demand
Type
Platinum-rated; contracted scale
Provider
Rate (from)
Type
The catch
Vast.ai (SXM)
$1.73
Spot
Interruptible; variable hosts
Hyperstack
$1.95
On-demand
Managed; per-minute billing
RunPod (PCIe)
$1.99
On-demand
Stable; the safe default
RunPod (SXM)
$2.69
On-demand
Faster interconnect
Nebius
$3.85
On-demand
Enterprise SLA
Lambda
$3.99
On-demand
1x available; training-grade
CoreWeave
$6.16
On-demand
Platinum-rated; contracted scale
Spot vs on-demand: the real cost lever
The single biggest factor in your H100 bill is spot versus on-demand. Vast.ai's spot market puts an H100 in reach for ~$1.73/hour because you're renting spare capacity from independent hosts — brilliant for interruptible inference or an experiment you can checkpoint, but the instance can be reclaimed when demand spikes. RunPod on-demand at ~$1.99/hour is only cents more and won't disappear mid-job, which makes it the sensible default for most people. Save the spot discount for work that can survive an interruption, and pay the small on-demand premium for anything you'd hate to lose.
Referral link
Rent the cheapest H100 on Vast.ai
Vast.ai's spot marketplace has the lowest H100 rate — from about $1.73/hr as of 22 Sep 2026. Best for interruptible inference and experiments; checkpoint your work so a reclaim doesn't cost you a run.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
What an H100 actually costs per job
An hourly rate only matters next to what you get done in the hour. An H100's 80GB comfortably runs a 70B model quantised, or hosts a smaller model at high throughput. At RunPod's ~$1.99/hour, an evening of inference (say four hours poking at a 70B) costs about $8; a QLoRA fine-tune of an 8B model in a couple of hours is around $4. Those are illustrative, not quotes — your real bill depends on utilisation and how long the box sits idle. Use our cost-to-run tool to match a specific model to the cheapest card that fits it, and shut the instance down the moment you're done.
Want an H100 that won't get reclaimed? RunPod on-demand, ~$1.99/hr.Referral link
As of 22 Sep 2026, the cheapest H100 is about $1.73/hour on Vast.ai's spot marketplace, followed by roughly $1.95–$1.99/hour on-demand from Hyperstack and RunPod. Enterprise neoclouds (Nebius, Lambda, CoreWeave) charge $3.85–$6.16 for the same chip because you're paying for guaranteed capacity and SLAs. For a personal or small-team workload, the sub-$2/hour options win easily; check the live figure before renting because spot prices move constantly.
Is a spot H100 safe to use?
For interruptible work, yes — inference, experiments, anything you can checkpoint. The risk is that a spot instance can be reclaimed with little warning, so you can lose an in-progress job. For a long training run or anything you can't afford to restart, pay the small premium for an on-demand instance (RunPod at ~$1.99/hour) instead of the cheapest spot listing.
Why is CoreWeave's H100 three times the price of Vast.ai's?
They're selling different products. Vast.ai is a marketplace of independent hosts competing on price with variable reliability; CoreWeave sells contracted, guaranteed capacity with enterprise SLAs and a Platinum reliability rating. The chip is identical — the premium buys certainty at scale, which matters to a company that can't tolerate an instance vanishing, and is dead weight for one engineer running weekend jobs.
How much does it cost to run a 70B model on a rented H100?
At RunPod's ~$1.99/hour on-demand rate, a few hours of inference on a quantised 70B model costs under $10, and you only pay while the instance is running. The key to keeping it cheap is shutting the box down when you're idle — an H100 left running overnight by accident is the most common way people overspend.
The cheapest H100 is under $2/hour if you rent from the marketplace tier — Vast.ai spot for the lowest rate, RunPod on-demand for stability. Compare the live cheapest offer on our cloud-GPU compare page, read the pricing pillar for the full market, and see the cheapest cloud GPU overall if you don't specifically need an H100, or the rent-vs-buy math before you commit to hardware.
