Rent an RTX 4090 48GB in the Cloud: Run a 70B Model on One Card for $0.65/hour (2026)

Most open models people actually want to run in 2026 sit in an awkward place: too big for a 24 GB card, not worth an 80 GB A100. A 70B model at 4-bit is about 43 GB. A 32B model at 8-bit is about 35 GB. Flux.1 dev and the newer video models at full precision want 30 GB or more.

The RTX 4090 48GB is the cheapest single card that closes that gap. Same chip as the regular 4090, twice the memory. On cloudgpu.app it is $0.65 per hour, billed per minute (20 October 2026).

This post answers the practical questions: what fits in 48 GB, what does not, what it costs, and when a different card is the better buy.

What fits in 48 GB

Weights below are the Ollama / llama.cpp Q4_K_M or Q8_0 file sizes. KV cache is what is left for context.

Model	Precision	Weights	Fits in 48 GB?	Context left (single user)
Llama 3.3 70B	Q4_K_M	~43 GB	Yes, tight	~8k tokens
DeepSeek-R1-Distill-Llama-70B	Q4_K_M	~43 GB	Yes, tight	~8k tokens
Qwen3 32B / QwQ 32B	Q8_0	~35 GB	Yes	32k+ tokens
Qwen3 32B	Q4_K_M	~20 GB	Yes, roomy	64k+ tokens, several users
Gemma 3 27B	Q8_0	~29 GB	Yes	32k+ tokens
Flux.1 dev	BF16	~24 GB + text encoders	Yes, no offloading	n/a
Qwen2.5 72B	Q4_K_M	~47 GB	No, nothing left for context	—
Any 70B	FP16 / BF16	~140 GB	No	—

The rule of thumb: 48 GB is the right size when you need one model a step bigger than 24 GB allows, for one user or a small team. It is not a training card for 70B and it is not a serving card for many concurrent users.

If you want the full arithmetic behind these numbers, we wrote it up in How much VRAM do you need to run an LLM.

Price per hour and per day

Card	VRAM	Price / hour	8-hour day	Best for
RTX 4090	24 GB	$0.33	$2.64	Models up to ~14B at 8-bit, 32B at 4-bit with short context
RTX 5090	32 GB	$0.59	$4.72	32B at 4-bit with 16k–32k context, fastest inference per card
RTX 4090 48G	48 GB	$0.65	$5.20	70B at 4-bit, 32B at 8-bit, Flux full precision
A100 80G	80 GB	$1.49	$11.92	70B with long context, multi-user serving, bigger fine-tunes

Billing is per minute. When you click Deploy, one hour is pre-authorised from your balance; when you stop the machine the unused minutes come back. A 25-minute test of a 70B model costs about $0.27.

How to run Llama 3.3 70B on it in about two minutes

1. Sign up at cloudgpu.app with email or Google.
2. Top up on the billing page. USDT (TRC-20) is credited within a minute; the card checkout also offers Google Pay and Apple Pay.
3. On Deploy, pick the Ollama template and RTX 4090 48G, then click Deploy.
4. When the instance is ready (about 2 minutes), the dashboard shows SSH details and a public HTTPS endpoint like https://<random-id>-i.cloudgpu.app.
5. Pull the model and call it from your laptop:

# on the instance
ollama pull llama3.3:70b

# from anywhere
curl https://<random-id>-i.cloudgpu.app/api/generate \
  -d '{"model": "llama3.3:70b", "prompt": "Explain KV cache in two sentences."}'

The endpoint is only reachable while your instance runs. More detail in Run Ollama on a cloud GPU with a public API.

Need an OpenAI-compatible API? Ollama already serves one on the same endpoint at /v1/chat/completions, so the OpenAI SDK works with base_url="https://<random-id>-i.cloudgpu.app/v1". If you prefer vLLM, you have root over SSH and can pip install vllm yourself; for a 70B model on one 48 GB card, use an AWQ or GPTQ 4-bit checkpoint and set --max-model-len 8192 so the KV cache fits.

When to pick a different card

• Your model fits in 24 GB. Take the regular RTX 4090 at $0.33/h. Same compute, half the price.
• You want the fastest tokens per second on a 32B model. The RTX 5090 at $0.59/h is newer silicon with FP8/FP4 paths and is quicker than either 4090 when memory is not the limit.
• You need long context on a 70B model, several users at once, or a full fine-tune. That is an 80 GB card. Check the pricing page for A100 80G and A800 80G availability.
• You need the machine in a specific country. Check the region before putting regulated data on it.

Rules of the house

Cryptocurrency mining and anything illegal get the instance destroyed and the account banned. Everything else people normally do on a GPU (inference, fine-tuning, image and video generation, notebooks, evaluation runs) is fine.

Try it

Pick RTX 4090 48G on the Deploy page. If it shows out of stock, it comes back on its own when supply returns, and the regular 4090 or the 5090 is usually available in the meantime. Questions about a specific model or workload: contact us and we answer within a day.