Cloud GPU: what you rent, how it works and what it costs paid in crypto
A cloud GPU is a graphics processor in someone else's data center that you use over the internet and pay for by the hour; at Petaflare that payment is in crypto only. Behind that simple idea are very different products and prices.
A cloud GPU is an NVIDIA or AMD accelerator in a provider's data center that you rent remotely, usually by the hour or minute, instead of buying the card. You connect to a machine that has the GPU attached, over SSH, a notebook or an API, and run training, inference or rendering on it. What varies between providers is whether you get a whole card or a slice of one, what is included in the price, and how much the same GPU costs: on 19 November 2026 one H100 rented for anywhere from $1.33 to $12.29 an hour.
How a cloud GPU works
The provider owns the hardware
Servers with one to eight GPUs sit in a data center, with power, cooling and networking paid by the provider.
You get a machine with the GPU attached
Usually a virtual machine or container with the card passed through, sometimes a whole physical server.
You connect remotely
Over SSH, a Jupyter notebook or an API, then install your code or start from a ready image with the drivers and CUDA.
You pay for the time you use
By the second, minute or hour while the machine runs. Stopping it stops the GPU charge, though storage may still cost money.
What you actually rent
Two offers that both say H100 can be very different products. Three things decide what you get.
| Question | Options | What it changes |
|---|---|---|
| Whole card or a slice? | Whole GPU · MIG partition · shared vGPU | A MIG slice of an A100 or H100 has a fraction of the memory and compute; a shared vGPU time-slices the card between users |
| What kind of machine? | Virtual machine · container · bare-metal server | Bare metal gives full control and no hypervisor overhead; containers start fastest but limit what you can install |
| Which form factor? | SXM (on HGX boards) · PCIe card | SXM cards are faster and linked by NVLink; PCIe cards draw less power and talk over the PCIe bus |
Petaflare rents whole GPUs only, 1 to 8 per machine, and shows the form factor once the machine runs. How renting works here.
Three ways to pay for a cloud GPU
- On demand. A fixed hourly price, and nobody takes the GPU back while you pay. Example: an H100 at $2.39 an hour here, $4.39 on RunPod's Secure Cloud.
- Spot, preemptible or interruptible. Spare capacity sold cheaply, which the provider can reclaim with little notice. Example: Azure's 8-GPU H100 machine costs $98.32 an hour on demand and about $18.17 as spot, $2.27 per GPU. Good for jobs that save checkpoints and can restart.
- Reserved or committed. A discount for paying for months or years in advance, common at the big clouds for steady workloads.
Whatever the model, check what else is billed: storage while a machine is stopped, data leaving the cloud, or a minimum billing period. The cloud GPU pricing page shows how our blocks of GPU-hours work.
Most clouds bill a card or an invoice. Petaflare takes crypto only: you buy a block of GPU-hours in Bitcoin, USDT, USDC, Ether, Solana, Litecoin or Monero, from $25.50 for 75 RTX 4090 hours, with no card and no ID. See how paying for a GPU with crypto works.
Who rents cloud GPUs, and what one H100 costs
Three families of providers rent the same card at very different prices. One H100, per GPU-hour, as listed on 19 November 2026 unless stated.
| Provider | Family | H100 offer | Per GPU-hour |
|---|---|---|---|
| Microsoft Azure | Public cloud | ND96isr H100 v5, 8 GPUs, East US | $12.29 |
| Google Cloud | Public cloud | a3-highgpu-8g, 8 GPUs, Iowa (price read on 29 October 2026) | $11.06 |
| AWS | Public cloud | p5.4xlarge or p5.48xlarge, US East | $6.88 |
| Lambda | GPU cloud | 1 × H100 SXM | $4.89 |
| RunPod Secure Cloud | GPU cloud | 1 × H100 SXM · PCIe | $4.39 · $2.89 |
| Petaflare | GPU cloud | 1 to 8 × H100, SXM or PCIe | $2.39 |
| RunPod Community Cloud | Marketplace of hosts | 1 × H100 PCIe | $1.99 |
| Vast.ai | Marketplace of hosts | H100 SXM, median of the day | $1.33 |
Sources: Azure Retail Prices API, Google Cloud GPU pricing, AWS EC2 on-demand pricing, Lambda, RunPod, Vast.ai public price feed. Public-cloud prices are for whole machines with CPUs, memory and local disk, divided by the number of GPUs.
What separates the three families
Public clouds (AWS, Azure, Google Cloud) charge the most per GPU, but bundle the GPU with large CPUs, memory and fast networking, and connect it to databases, storage and identity tools. They suit companies already running there, and they often require quota requests before you can launch big GPU machines.
GPU clouds focus on GPUs alone and price them per GPU-hour. Setup is simpler and prices are a half to a fifth of the big clouds, with fewer surrounding services.
Marketplaces connect you with independent hosts. They have the lowest prices and the widest choice of cards, but quality, uptime and terms vary from host to host. A side-by-side of named providers is on GPU cloud providers compared.
Cloud GPU or your own card
Divide what a card costs you by the hourly price of renting it. At our prices, $2,000 spent on an RTX 4090 buys about 5,880 hours of one in the cloud, and $25,000 spent on an H100 buys about 10,460 hours: roughly eight months and fourteen months of running non-stop. Owning also means a server, power, cooling and repairs.
The cloud wins for work that comes in bursts, for trying a card before buying it, and for jobs that need eight GPUs for a few days. Your own card wins when it will be busy most of the time for years, or when data must never leave your building. To scale up later, a whole GPU server gives you up to eight cards in one machine. When you are ready, you can rent a GPU by the hour here.
When a cloud GPU is the wrong choice
- Small models. Many classical machine-learning models and small LLMs run well on a laptop or a CPU server.
- Constant, predictable load for years. Buying or reserving capacity is usually cheaper than paying by the hour.
- Strict data rules. Some regulated data cannot leave your own infrastructure.
- Learning on a budget. Free cloud GPU options such as notebook services can be enough to start, within their limits.
Cloud GPU questions
How does a cloud-based GPU work?
The provider runs servers with GPUs in its data center and gives you remote access to a machine with one or more of them attached. You connect over SSH or a notebook, run your workload, and pay for the time the machine runs.
Which cloud GPU provider is best?
It depends on what you need: public clouds for integration with other cloud services, GPU clouds for the lowest fixed price per GPU, marketplaces for the lowest price at any cost. Our provider comparison lists them with prices.
Does Google Cloud provide GPUs?
Yes. Google Cloud rents NVIDIA GPUs such as the L4, A100, H100, H200 and B200 in machine types like A2, A3 and A4. Their per-GPU prices are compared on Google Cloud GPU pricing.
Is a cloud GPU slower than a local one?
The GPU itself runs at full speed. What can be slower is moving data to it over the internet, so upload datasets once and keep them on the machine's disk.
Sources
- NVIDIA Multi-Instance GPU (MIG) user guideread 19 November 2026
- Azure Retail Prices APIread 19 November 2026
- AWS EC2 on-demand pricingread 19 November 2026
- Google Cloud GPU pricingread 29 October 2026
- RunPod pricingread 19 November 2026
- Vast.ai public GPU price feedread 19 November 2026