Rent an RTX 5090 in the Cloud by the Minute (2026 Prices, No Card Needed)

The RTX 5090 is the card most people building on open models actually want in 2026: 32 GB of GDDR7, enough to run a quantised 32B model or fine-tune a 7B with a comfortable batch size, at a fraction of what an H100 costs. The problem is not the hardware. It is that most places renting it want a Visa card, a monthly commitment, or both.

This post is a plain answer to three questions: what a 5090 costs per hour today, why per-minute billing matters more than the headline rate for most people, and how to pay if your bank keeps declining international AI vendors.

RTX 5090 cloud prices, 6 September 2026

These are the live rates on cloudgpu.app at the time of writing. They move with supplier stock, so treat the table as a snapshot and check the pricing page before you plan a budget.

Card	VRAM	Price / hour	Billing
RTX 3090	24 GB	$0.21	per minute
RTX 4090	24 GB	$0.33	per minute
RTX 4090 48G	48 GB	$0.65	per minute
RTX 5090	32 GB	$0.59	per minute
A100 80G	80 GB	$1.49	per minute
RTX PRO 6000 96G	96 GB	$1.49	per minute
H800	80 GB	$1.79	per minute

For comparison, mainstream US providers list the RTX 4090 between $0.34 and $0.74 per hour and the 5090 between $0.69 and $0.89 per hour on their on-demand tiers as of this month. The gap is not magic. Our supply comes from GPU suppliers in China and Southeast Asia where consumer cards are cheap and plentiful, and we take a thin margin on top.

Why per-minute billing is the number that matters

Most 5090 workloads are short. A LoRA fine-tune on a 7B model is 20 to 60 minutes. A batch of 500 Flux images is 15 minutes. A model evaluation run is 40 minutes. On hourly billing you pay for the full hour every time; on per-minute billing you pay for what you used.

A concrete example from our own logs this week: a customer ran four RTX 5090 sessions totalling 2 hours 47 minutes. On hourly rounding that is 4 hours, or $2.36. Per minute it was $1.64. Over a month of daily experiments the difference is a two-digit dollar amount, which is real money when the whole budget is $20.

How it works on cloudgpu.app: when you click Deploy, one hour is pre-authorised from your balance. When you stop the machine, the unused minutes come back. If your balance runs low you get an email six hours and one hour before it hits zero, and the machine is stopped automatically at zero so you never go into debt.

Paying without a US credit card

This is the reason most of our users are here. If you are in Vietnam, Indonesia, Pakistan, Turkey, Argentina, Nigeria or a dozen other places, international card payments to AI vendors are unreliable at best.

• USDT (TRC-20). Send USDT to your personal deposit address on the billing page. The balance is credited after network confirmation, usually within a few minutes. What you send is what lands in your balance; we absorb the network fee. Typical first top-up is $5 to $20.
• PayPal. Added in September 2026 for people who would rather not touch crypto.

There is no subscription, no minimum balance and no expiry on unused balance.

What you get with the machine

Every 5090 rental comes with:

• Root SSH access with a per-instance password.
• A template of your choice: plain CUDA + PyTorch, Ollama, vLLM, SGLang, ComfyUI, Stable Diffusion WebUI or LLaMA-Factory.
• For templates that expose an API (Ollama, vLLM, SGLang, ComfyUI), a public HTTPS endpoint on our domain, for example https://5a18f8b218351513-i.cloudgpu.app, so you can call the model from your laptop or your app without setting up an SSH tunnel. The address is a random 16-character id and is only reachable while your instance runs.
• JupyterLab on the supplier’s own URL for file management.

The endpoint detail matters more than it sounds. Most cheap GPU suppliers only expose JupyterLab. Without the tunnel, “Ollama on a rented 5090” means “Ollama you can only reach through ssh -L”. We wrote up how the tunnel works in Run Ollama on a cloud GPU with a public API in two minutes.

What a 5090 is good for, and when to pick something else

Pick the 5090 for:

• Running 14B to 32B open models at 4-bit for a chatbot or agent backend. Qwen3 32B and DeepSeek-R1 distill 32B fit with room for context.
• Fine-tuning 7B to 8B models with LoRA or QLoRA. Full fine-tunes of 7B models are possible at small batch sizes.
• Image and video generation. Flux.1 dev at full precision, SDXL, Wan 2.x video at 480p.
• Anything where you are compute-bound rather than memory-bound. Blackwell’s FP8 and FP4 paths are noticeably faster than the 4090 for inference.

Pick something else when:

• You need more than 32 GB in one card. The RTX 4090 48G ($0.65/h) or RTX PRO 6000 96G ($1.49/h) are the next steps up. An A100 80G at $1.49/h is the right call for training runs that need NVLink-class memory bandwidth.
• Your job is tiny. A 3090 at $0.21/h runs a 7B model fine and costs a third as much.
• You need the machine in a specific jurisdiction. Our 5090 inventory is in China; check the region column on the pricing page before putting regulated data on it.

Rules of the house

Our supplier accounts are ours, and their acceptable-use policy becomes ours. Two things get an instance destroyed and an account banned, no appeal: cryptocurrency mining and anything illegal. A watchdog process runs on every machine and reports back; we also check from outside. This is not a marketplace for anonymous compute, and we say so up front so nobody wastes a top-up finding out.

Try it

1. Sign up at cloudgpu.app (email or Google). New accounts get $1.00 of trial credit: $0.50 at signup and $0.50 once you verify your email. That is about 100 minutes on a 5090 or five hours on a 3090.
2. Top up $5 to $20 with USDT or PayPal on the billing page.
3. Go to Deploy, pick a template, pick RTX 5090, click Deploy.
4. Two minutes later: SSH details and, for API templates, your https://…-i.cloudgpu.app endpoint are on the dashboard.
5. Stop the machine when you are done. Unused minutes come back.

If a card you need is missing from the list, tell us on the contact page. The inventory follows demand.