Vibe Engines is free, interactive AI engineering & system design — runs in your browser. Pick a learning path
GPU Rental Price Reference
A cloud GPU price reference and cost estimator. Compare approximate on-demand hourly rates for common GPUs (T4, L4, A10G, A100, H100) and estimate what a training or inference job costs from the GPU count and hours. Rates are approximate and drift over time — always confirm current pricing with your provider.
⚠ Approximate on-demand rates (USD), mid-2026 ballpark, for quick comparison only. Real pricing varies by provider, region and commitment (spot / reserved) and changes often — confirm current pricing with your provider.
| GPU | Memory | ~$/hr | Typical use |
|---|---|---|---|
| T4 | 16 GB | $0.35 | Light inference |
| L4 | 24 GB | $0.70 | Efficient inference |
| A10G | 24 GB | $1.00 | Mid inference / small training |
| A100 40GB | 40 GB | $2.50 | Training / large-model inference |
| A100 80GB | 80 GB | $3.50 | Training / large context |
| H100 80GB | 80 GB | $5.00 | Large-scale training |
Estimate a job — A100 40GB
GPU time is the largest line in most AI budgets, and a rough sense of the rates keeps a plan honest. This reference lists approximate on-demand hourly prices for six common cloud GPUs and multiplies out a job, so you can see what a training run or an inference fleet costs before you rent anything.
What it computes
The table holds one approximate on-demand rate per GPU, in US dollars per GPU-hour, described in the widget as a mid-2026 ballpark for quick comparison only.
| GPU | Memory | ~$/hr | Typical use |
|---|---|---|---|
| T4 | 16 GB | $0.35 | Light inference |
| L4 | 24 GB | $0.70 | Efficient inference |
| A10G | 24 GB | $1.00 | Mid inference / small training |
| A100 40GB | 40 GB | $2.50 | Training / large-model inference |
| A100 80GB | 80 GB | $3.50 | Training / large context |
| H100 80GB | 80 GB | $5.00 | Large-scale training |
Click a row to select a GPU, then enter how many GPUs and how many hours. Two numbers come out:
job cost = rate × gpus × hours
monthly if 24×7 = rate × gpus × 730
730 = 8,760 hours in a year ÷ 12 monthsRates display to the cent and both totals are rounded to whole dollars.
A worked example
The calculator opens on the A100 40GB with 8 GPUs for 20 hours.
- GPU-hours:
8 × 20 = 160. - Job cost:
160 × $2.50 = $400. - Monthly if run 24×7:
$2.50 × 8 × 730 = $14,600.
Keeping 8 GPUs for 20 hours and switching the card gives:
| GPU | Job (8 × 20 h) | Monthly if 24×7 |
|---|---|---|
| T4 | $56 | $2,044 |
| L4 | $112 | $4,088 |
| A10G | $160 | $5,840 |
| A100 40GB | $400 | $14,600 |
| A100 80GB | $560 | $20,440 |
| H100 80GB | $800 | $29,200 |
The gap between the job and the month is the point: a 20-hour run is a few hundred dollars, while the same allocation left on is five figures.
Assumptions and limits
- On-demand only. Spot, preemptible and reserved prices are not modeled, and they can differ a lot from on-demand.
- Approximate and dated. The rates are rounded, mid-2026 ballpark figures. Providers and regions differ, and prices change often.
- GPU time only. Storage, data transfer, networking and the CPU and memory attached to an instance are not included. Providers also tend to sell GPUs in fixed instance shapes, so you may not be able to rent exactly the count you type.
- Hours you enter are hours you pay for. The tool does not know your throughput, so it cannot tell you how many hours a job needs, and it ignores billing granularity. Doubling the GPU count rarely halves the hours either, because multi-GPU jobs spend time communicating, so re-estimate the hours whenever you change the count.
How to use the result
- Budget a run before you start it. Estimate the hours, multiply out the cost, and add a margin for failed runs and debugging.
- Compare cost per job, not cost per hour. The H100 80GB rate is twice the A100 40GB rate, so it wins only if the job finishes in less than half the hours. The A100 80GB costs 1.4 times the A100 40GB for twice the memory.
- Size the card by memory first. Check that the model fits, then choose the cheapest card that meets your throughput target.
- Use the monthly figure for always-on work. It is the number to hold against a reserved or committed-use quote, and the reason to shut down idle allocations.
How it works
- Approximate on-demand hourly rates (USD, mid-2026 ballpark) for T4, L4, A10G, A100 and H100.
- Job cost = rate × GPU count × hours.
- Monthly if run 24×7 = rate × GPU count × 730 hours.
- Prices drift constantly, so verify with your provider before committing spend.
Frequently asked questions
How much does it cost to rent an A100 or H100 per hour?
In this reference, which uses approximate on-demand rates as a mid-2026 ballpark in US dollars, an A100 40GB is about $2.50 per GPU-hour, an A100 80GB about $3.50 and an H100 80GB about $5.00. The smaller cards are a T4 at about $0.35, an L4 at about $0.70 and an A10G at about $1.00. Real prices vary by provider, region and commitment, and they change often, so confirm the current rate before you commit spend.
How do I estimate the cost of a GPU job?
Multiply the hourly rate by the number of GPUs and the wall-clock hours the job holds them. The default example is eight A100 40GB GPUs for 20 hours: 2.50 times 8 times 20 is $400. The same eight GPUs left running around the clock for a month cost 2.50 times 8 times 730, about $14,600. Count every hour the GPUs are allocated, including setup, debugging and idle time, because an idle GPU bills like a busy one.
Why does the calculator use 730 hours for a month?
A year has 8,760 hours (365 days times 24), and dividing by 12 gives an average month of 730 hours. The monthly figure is rate times GPU count times 730, which is what a GPU costs if it is never released. It is the number to compare with a reserved or committed-use quote for steady workloads such as an always-on inference fleet, and it shows quickly how a temporary experiment becomes a large bill when nobody shuts it down.
Which cloud GPU should I rent?
The cheapest one that fits your model in memory and finishes the work in acceptable time. The memory column is the first filter. After that, compare cost per job rather than cost per hour: the H100 80GB costs twice the A100 40GB rate here, so it is cheaper only if your job finishes in less than half the hours. The A100 80GB costs 1.4 times the A100 40GB for twice the memory, which can avoid splitting a model across cards.
How accurate are these GPU prices?
They are rounded, approximate on-demand rates intended for quick comparison and planning, not billing. The reference describes them as a mid-2026 ballpark. Actual prices depend on the provider, the region, the commitment level (on-demand, reserved or spot) and current supply, and they move often. The ratios between cards are more useful than any single number, but always confirm the live price with your provider before a decision that commits real money.
How can I lower my GPU bill?
Keep utilization high, since GPUs bill by the hour whether busy or idle: batch inference requests and release machines that are not working. Match the card to the model instead of defaulting to the largest one, and quantize a model if that lets it fit on a cheaper card. Use spot or preemptible capacity for work that can survive interruption, and reserved capacity for steady long-running load. Check the provider for current discounts rather than assuming one.