☁️ Cloud GPU Pricing: When Renting Beats Buying a Card
H100s span $1.49 to $6.98 an hour across providers. Here's the break-even maths against buying, and the billing details that decide your real bill.
Sam Whitaker · Developer & API Cost Writer
· 5 min read
⚡ TL;DR — quick answers
- How much does it cost to rent an H100 per hour?
- Published rates run from about $1.49/hr on marketplace providers up to roughly $6.98/hr on hyperscalers, with specialist GPU clouds landing between $2 and $3.50. That 4.7x spread is for nominally identical hardware, and the difference is reliability, support, region availability and whether you are on a spot tier that can be reclaimed.
- Is renting a GPU cheaper than buying one?
- It depends entirely on utilisation, and the maths is simple enough to do in one line. An RTX 4090 rents around $0.34/hr; a card costs roughly $1,800. That is about 5,300 hours before buying wins, which is around seven months of genuine 24/7 use. If your GPU would sit idle most of the week, renting is cheaper. If you are saturating it daily, buy.
- What is the catch with cheap spot GPU pricing?
- Preemption. Spot and community tiers can be reclaimed while your job is running, so anything without checkpointing loses its progress. They are excellent for batch inference, rendering queues and anything you can resume, and a poor fit for long unattended training. The advertised saving is real; it is paid for in reliability rather than being free.
Key takeaways
- H100 hourly rates span roughly $1.49 to $6.98 depending on provider, which is a 4.7x spread for nominally the same silicon.
- An RTX 4090 rents around $0.34/hr, so a $1,800 card needs roughly 5,300 hours of use to break even against renting.
- Spot and community tiers are dramatically cheaper but can be reclaimed mid-job. They suit batch work with checkpoints, not long unattended training runs.
- Per-minute billing and idle-time charges matter more than the headline rate for bursty workloads. Storage and egress are billed separately almost everywhere.
I read the pricing page before the landing page, and with GPU rental that habit pays immediately, because the headline hourly rate is the least interesting number on it. The same H100 spans roughly $1.49 to $6.98 an hour depending on where you rent it, which tells you the number is describing something other than the silicon.
By the numbers
- H100: about $1.49/hr (marketplace) up to $6.98/hr (hyperscaler); specialist clouds land near $2.89-$3.29
- A100 80GB: roughly $1.07/hr on-demand, dipping toward $0.60-$0.67/hr on spot or marketplace tiers
- RTX 4090: around $0.34/hr on community tiers
- Break-even on a ~$1,800 4090 at $0.34/hr: roughly 5,300 hours, about seven months of continuous use
The break-even line, written out
Here is the only calculation that matters, and it takes ten seconds. Card price divided by hourly rental rate equals hours-to-break-even. A 4090 at $1,800 against $0.34/hr gives about 5,300 hours. Run the GPU eight hours a day, every working day, and that is roughly two and a half years. Run it continuously and it is about seven months.
Now be honest about your actual utilisation. Most people buying a card for AI work use it in bursts: a heavy weekend, then a fortnight of nothing. At four hours a week you are looking at twenty-five years to break even, and the card will be obsolete long before that. The local versus cloud comparison walks through the non-cost reasons to own hardware anyway, and they are legitimate. Cost is simply not usually one of them at low utilisation.
So what: measure a fortnight of real GPU hours before buying anything.
What the 4.7x spread is actually buying
Nothing about the chip. The spread is reliability, region, support, and preemption risk. A marketplace provider matches you with someone's spare capacity, which is why it is cheap and why the host can disappear. A hyperscaler charges several times more for capacity that will still be there tomorrow, with an SLA and a support contact.
The practical read: pick by workload shape, not by price. Batch inference and rendering queues tolerate preemption fine, especially with checkpointing, so the cheap tiers are genuinely the right answer. A multi-day training run that loses everything on reclaim is a false economy at any hourly rate.
The billing details that mug you
Three things reliably cost more than people budget.
Idle time. Most providers bill a running pod whether or not it is computing. A job that finished at 2am on a pod you stopped at 9am cost you seven hours of nothing. Script the teardown rather than trusting yourself to remember.
Persistent storage. Volumes bill continuously, including while the pod is stopped. A stopped instance with a large dataset attached is a small recurring charge that quietly runs for months.
Egress. Getting results out is metered on most platforms. It rarely dominates, but it is real and it never appears in an hourly comparison table.
I script the whole lifecycle for anything I run more than twice: spin up, run, pull results, tear down. Not for elegance, but because every one of those charges comes from a step a human forgets at 2am.
So what: automate teardown, and check what your stopped pods are still billing you for.
Rent to find out, buy once you know
The pattern that works: rent first to discover your real utilisation and which card class you actually need, then buy if the maths turns over. Renting is a cheap experiment that answers a question a spec sheet cannot, which is how many hours you genuinely use. If you would like a view on which card to buy once that answer is in, the GPU buyer's guide covers the VRAM ceilings that decide it, and the same discipline about measuring before spending shows up in our cost-per-shot breakdown.
Matching the card to the job
Renting also lets you right-size, which is the saving people forget to count. Plenty of work does not need an H100 at all. Inference on a mid-sized model, most image generation and a great deal of fine-tuning run perfectly well on a 4090 at a tenth of the hourly rate. Buying commits you to one tier forever; renting means picking the cheapest card that clears the job.
The rule I use: memory first, then speed. Find the smallest card whose VRAM holds your model and workload comfortably, then check whether the speed is tolerable. Renting an H100 for something a 4090 fits is the most common way people overspend in this category, and it happens because the H100 is the card everyone talks about rather than the card most jobs need.
Multi-GPU is where the maths gets less friendly. Two cards rarely make a job twice as fast, because interconnect bandwidth and how well your framework parallelises both intervene. Before renting a 4x node, confirm your workload actually scales, or you will pay four times the rate for something closer to double the throughput.
So what: rent the smallest card that holds the model, and verify multi-GPU scaling before paying for it.
The invoice check, again
Every figure in this piece is a published rate from a comparison tracker, and trackers in this market disagree with each other and go stale within weeks. Regional pricing, committed-use discounts and promotional credits all move the real number. Treat the table above as a starting estimate to sanity-check against, then read your own first invoice carefully before you scale anything up.
Sources & further reading
Outside figures cited above. First-hand test results are our own and noted as such in the text.
Frequently asked questions
▸How much does it cost to rent an H100 per hour?
Published rates run from about $1.49/hr on marketplace providers up to roughly $6.98/hr on hyperscalers, with specialist GPU clouds landing between $2 and $3.50. That 4.7x spread is for nominally identical hardware, and the difference is reliability, support, region availability and whether you are on a spot tier that can be reclaimed.
▸Is renting a GPU cheaper than buying one?
It depends entirely on utilisation, and the maths is simple enough to do in one line. An RTX 4090 rents around $0.34/hr; a card costs roughly $1,800. That is about 5,300 hours before buying wins, which is around seven months of genuine 24/7 use. If your GPU would sit idle most of the week, renting is cheaper. If you are saturating it daily, buy.
▸What is the catch with cheap spot GPU pricing?
Preemption. Spot and community tiers can be reclaimed while your job is running, so anything without checkpointing loses its progress. They are excellent for batch inference, rendering queues and anything you can resume, and a poor fit for long unattended training. The advertised saving is real; it is paid for in reliability rather than being free.
▸What hidden costs should I budget for?
Storage and egress, mostly. Persistent volumes bill whether or not a pod is running, so a stopped instance with 500GB attached is still costing you daily. Egress fees apply when pulling results out. And idle time bills at the full rate on most providers, meaning a pod left running overnight after a job finished costs the same as one doing work.
The 5 best AI video finds, every week
New models, tested prompts, and what actually worked in our production — one short email a week. No spam, unsubscribe anytime.
Written by Sam Whitaker
Developer & API Cost Writer
Indie developer who reads the API docs before opening the UI and scripts every test he runs more than twice. Tracks cost-per-call and rate limits the way accountants track invoices.
Explore these topics
Every guide, comparison and prompt library we have on each.
Keep learning
Workflow & Craft · GPUs & Hardware
GuidesRun Stable Diffusion Locally: Setup That Works in 2026
SD1.5 vs SDXL VRAM math, the ComfyUI desktop install, your first workflow, where checkpoints come from, and the errors that eat everyone's first afternoon.
