How to Rent a Cloud GPU: Step by Step

Renting a cloud GPU takes about ten minutes: sign up with a provider, add a payment method, pick a GPU and an image, and launch. The cheapest GPU on our index right now is $0.10/hr for a RTX A4000 at TensorDock, and a datacenter-class H100 starts at $1.99/hr on-demand. Those numbers come from the live dataset and change daily. The ten minutes of clicking is the easy part; the part that decides what you actually pay is the three choices before it: which GPU, which pricing kind, and which provider. Here is the whole process in order.

Step 1: Pick the GPU for your workload, not the biggest one

The most expensive mistake in GPU rental is made before you touch a provider: renting more card than the job needs. Match the GPU to the workload first. Our use-case pages walk through the picks for LLM inference, fine-tuning, Stable Diffusion, and more, and the full GPU list shows every card we track with live prices.

The single number that matters most is VRAM: your model either fits in memory or it does not. A quantized 7B model runs on a 24GB consumer card at $0.35/hr; a 70B model needs 80GB-class hardware. Work out the memory requirement first, then rent the smallest card that clears it. The VRAM buckets sort cards by memory size so you can shop at exactly the capacity you need.

Step 2: Pick the pricing kind

Every provider sells the same silicon under two or three different pricing models, and the gap between them is often larger than the gap between providers.

On-demand (secure) is the default: dedicated datacenter hardware, yours until you shut it down, at the highest rate. It is the right choice for anything that cannot tolerate interruption, like a live endpoint or a long training run without checkpointing.

Spot and community tiers are the discount aisle. A spot instance rents spare capacity that the provider can reclaim with short notice; a community tier pools GPUs from independent hosts rather than a certified datacenter. Both price the same card well below the secure rate. They are the right choice when your job can checkpoint and resume: batch inference, rendering, experimentation. Read the secure vs community trade-off before choosing.

Serverless flips the model: you pay per second of actual execution and nothing while idle, in exchange for a higher per-second rate and cold starts. For spiky or occasional workloads it usually beats holding an instance open. The serverless pricing page compares those rates across providers.

Step 3: Pick a provider

With a GPU and a pricing kind chosen, the provider question mostly answers itself: sort by price and check the details. Our provider directory lists everyone we track, and every price table on this site sorts by price with no sponsorship or referral influence on ordering; that is an iron rule of the index. What separates providers at similar prices: billing granularity, egress policy, region availability, and whether the tier is datacenter or community hardware. The provider pages spell each of those out.

Step 4: Price the whole bill, not the GPU-hour

The advertised rate is the compute price. The invoice adds more:

• Egress. Moving data out of the network is billed per gigabyte on the big hyperscalers and can rival the compute cost on data-heavy jobs. Several GPU-focused clouds charge nothing for egress, which matters more than a few cents of hourly rate if you move terabytes. See egress fees.
• Storage. Persistent volumes bill separately from compute and keep billing while the instance is stopped. Checkpoints and datasets add up over a month.
• Idle time. A stopped instance may stop billing for compute but not for the disk attached to it.

The fees page breaks down what each provider charges beyond the sticker rate. Fold these in and you get the effective cost per hour, the number our tables actually sort on.

Step 5: Launch, then verify the billing granularity

Launching is the short part: pick an image or template (most providers offer prebuilt PyTorch and CUDA images), attach your SSH key, and start the instance. Before you leave it running, confirm two things on the billing page. First, the granularity: per-second billing charges you for exactly what you use, while per-hour billing rounds a 61-minute session up to two hours. RunPod bills per second; Lambda bills per minute on on-demand; others round to the hour. Second, the shutdown behavior: know the difference between stop (disk keeps billing) and terminate (everything gone) before you walk away.

If you are renting regularly, the rent-vs-buy calculator shows the break-even point where owning the card beats renting it, computed on live prices. And if you own idle GPUs yourself, the earnings calculator runs the same math from the other side.

Common first-timer mistakes

• Paying on-demand rates for interruptible work. Batch jobs that checkpoint can run on spot or community tiers at a fraction of the secure rate. On-demand is for workloads that cannot go down, not for everything by default.
• Ignoring egress until the invoice arrives. The compute math looks great until you export a few terabytes of outputs from a provider that bills per gigabyte. Check the egress policy before you upload your dataset.
• Renting more VRAM than the model needs. An 80GB card for a model that fits in 24GB roughly triples your bill for zero benefit. Check the VRAM buckets and quantize before you size up.
• Missing per-second vs per-hour billing. On short, bursty sessions, hourly rounding can double the effective rate. Granularity matters more the shorter your jobs are.
• Leaving instances running. The classic. Set a spend alert if the provider offers one, and know whether stop actually stops the billing.

FAQ

How much does it cost to rent a cloud GPU?

The floor of our index today is $0.10/hr for a RTX A4000, a datacenter H100 starts at $1.99/hr on-demand, and an A100 at $0.89/hr. Rates change daily; the live table shows every current price with its source and timestamp.

Do I need a long-term contract?

No. On-demand rental is the default across the industry: start and stop whenever you want, pay for what you use. Reserved terms exist and cut the hourly rate, but they bill whether you use the hours or not, so commit only after you know your utilization. See min-commit.

Can I rent a GPU for just an hour?

Yes, and often for much less than an hour. Providers with per-second billing charge a 90-second job as 90 seconds. For occasional short jobs, serverless is usually the cheapest structure because you pay nothing while idle.

What do I actually get when I rent one?

Typically a virtual machine or container with the GPU attached: SSH access, a prebuilt CUDA or PyTorch image, and root on the box for the life of the rental. Serverless platforms differ: you deploy code or a container and never see a machine at all.