Rent an RTX 4090 48G by the minute

The RTX 4090 48G is the same chip as the RTX 4090 with double the VRAM: 48 GB on one card. That extra memory is the point. Models and pipelines that do not fit in 24 GB can run on a single GPU, without splitting across cards. On CloudGPU you rent it by the minute and destroy it when you are done.

See the current RTX 4090 48G rate · Deploy an RTX 4090 48G

The card

RTX 4090 48G
Chip Same as the RTX 4090 (Ada)
VRAM 48 GB
Memory bandwidth 1008 GB/s
FP16 165 TFLOPS
CUDA compute capability 8.9 (RTX 4090 chip)
Billing Per minute, 1-minute minimum
Rate On cloudgpu.app/pricing

What 48 GB lets you run

Which model fits depends on its size, quantization and context length; see how much VRAM an LLM needs. The context window needs memory too: longer contexts use more VRAM for the KV cache.

Related guides: Llama 3.3 70B, Qwen2.5 72B, Qwen3 32B, Rent a GPU for ComfyUI and the blog post Rent an RTX 4090 48G in the Cloud, Billed by the Minute (2026).

If 32 GB is enough, the RTX 5090 has more memory bandwidth; if you need 80 GB, see the A100 80G.

How it works

  1. Sign up with an email address or a Google account.
  2. Add balance. Visa or Mastercard, Alipay (in most countries) or Google Pay / Apple Pay in one checkout, USDT on TRC-20, or bank transfer for business customers. See payments.
  3. Deploy. On Deploy, choose a template (Ollama, JupyterLab + Ollama, ComfyUI, LLaMA-Factory) or Raw GPU, and the RTX 4090 48G.
  4. Work. About 2 minutes later the instance row shows root SSH and JupyterLab; API templates also get a public HTTPS URL. For vLLM, see vLLM on a rented GPU.
  5. Destroy. Copy off what you want to keep, then destroy the instance. Billing stops and unused minutes go back to your balance.

Billing in plain words

  • Billed per minute of GPU time, with a 1-minute minimum.
  • While an instance runs, each started hour is pre-charged from your balance; unused minutes go back to your balance when you destroy the instance.
  • No subscription and no monthly fee; the balance does not expire.
  • Persistent storage across restarts is still on the roadmap, so plan large downloads into each session and copy your outputs off before you destroy.

FAQ

What is an RTX 4090 48G? It is the same chip as the RTX 4090 with double the VRAM: 48 GB instead of 24 GB, so larger models fit on one card.

Which models fit in 48 GB? Which model fits depends on its size, quantization and context length; see the guide on how much VRAM an LLM needs. Leave memory for the context window; longer contexts need more VRAM.

How is it billed? Per minute of GPU time, with a 1-minute minimum. Each started hour is pre-charged from your prepaid balance, and unused minutes go back to your balance when you destroy the instance. The current rate is on the pricing page.

How fast does it start? An instance usually takes about 2 minutes from Deploy to a running machine.

Do I get root access? Yes. Every instance comes with root SSH and JupyterLab.

Ready to start?

Deploy an RTX 4090 48G · See the current rate

Sources

All read on 24 October 2026 (Beijing time, UTC+8).