Rent an RTX 4090 with crypto from $0.34 an hour
The cheapest card we rent, paid only in crypto, and still one of the most useful: 24 GB of GDDR6X for image generation, ComfyUI workflows, rendering and LLMs up to about 14B parameters.
You can rent an RTX 4090 on Petaflare for $0.34 per GPU-hour, $0.32 on the 1,200-hour Project block and $0.31 on the 4,000-hour Fleet block. The smallest purchase is 75 GPU-hours for $25.50. Each card is a whole RTX 4090 with 24 GB of GDDR6X, 1 to 8 per machine, with root SSH and a PyTorch, CUDA or Docker image. In 2026 it still handles SDXL, Flux in FP8, ComfyUI pipelines and 8B models in BF16; for bigger models, see the 32 GB RTX 5090.
RTX 4090 specs
Ada Lovelace, the same generation as the L40S, in a consumer card. Its fourth-generation Tensor Cores support FP8, and it keeps RT cores and two AV1 video encoders.
- Architecture
- Ada Lovelace
- Memory
- 24 GB GDDR6X
- Memory bandwidth
- 1,008 GB/s
- BF16 Tensor (dense)
- not published as a Tensor figure for GeForce
- FP8 Tensor (dense)
- supported (4th-gen Tensor Cores)
- Board power
- 450 W
- GPU link
- PCIe Gen4, no NVLink
- Form factor
- PCIe
- CUDA cores
- 16,384 CUDA cores
CUDA cores, memory and board power from NVIDIA's RTX 4090 page. Memory bandwidth is worked out from its 21 Gbps GDDR6X on a 384-bit bus.
RTX 4090 cloud prices
One RTX 4090, as listed on each provider's pricing page on 19 November 2026.
| Provider | Price type | Per GPU-hour |
|---|---|---|
| Petaflare | Fixed, on demand | $0.34 · $0.31 on Fleet |
| RunPod Secure Cloud | Fixed, on demand | $0.89 |
| RunPod Community Cloud | Third-party hosts | $0.34 |
| Salad | High priority · lowest priority (interruptible) | $0.44 · $0.16 |
| Vast.ai | Marketplace, median of the day | $0.36 |
Sources: RunPod, Salad, Vast.ai public price feed (18 November median, cited with attribution). Read on 19 November 2026. Every card is on the full price list.
Reading those RTX 4090 prices
Our $0.34 matches RunPod's Community Cloud and sits below the Vast.ai median and Salad's high-priority tier, at less than 40% of RunPod's Secure Cloud price. Salad's lowest tier is cheaper at $0.16, but those machines can be taken back at any time, which suits batch jobs that can restart. Ours are not interruptible: nobody takes the card back while you have hours.
Seventy-five RTX 4090 hours cost $25.50, the cheapest block on the site, and 4,000 hours cost $1,240, or $0.31 an hour. Monero keeps the payment private and is credited after 10 confirmations, about 20 minutes. Crypto is the only way to pay and no card or ID is asked: see the Monero GPU server page.
Is the RTX 4090 still good in 2026?
For most image and small-model work, yes. What has changed since its 2022 launch is model size, not the card's speed, so the answer depends on whether your job fits in 24 GB:
- Stable Diffusion 1.5 and SDXL: comfortably, with batch generation and LoRA training. See the Stable Diffusion GPU guide for the memory per model.
- Flux.1: the 12-billion-parameter model takes about 24 GB in BF16, too much with its text encoder, but runs well with FP8 weights.
- ComfyUI pipelines: most workflows fit; heavy video workflows need FP8 models or offloading. Our ComfyUI setup guide shows how to run it on a rented card.
- LLMs: 8B models in BF16, 14B in 8-bit, around 30B in 4-bit with short contexts. Beyond that you need more memory.
- Rendering: Blender Cycles with OptiX, Redshift and Octane run well thanks to its RT cores; see cloud GPU rendering.
What fits in 24 GB
Weights take 2 bytes per parameter in BF16, 1 in FP8 or 8-bit and about 0.5 in 4-bit, before activations and cache.
| Job | Memory needed | RTX 4090 (24 GB) |
|---|---|---|
| SDXL image generation | About 8 to 12 GB | Yes, with batches |
| Flux.1 dev in BF16 with its text encoder | More than 24 GB | No: use FP8 weights |
| Flux.1 dev in FP8 | About 12 GB plus the encoder | Yes |
| Llama 3.1 8B in BF16 | About 16 GB plus cache | Yes |
| A 14B model in 8-bit | About 14 GB plus cache | Yes |
| A 32B model in 4-bit | About 18 to 20 GB | Yes, short contexts |
| A 70B model in 4-bit | About 35 to 40 GB | No: L40S, A100 or H100 |
Rules of thumb; measure your own job on a Burst block before buying more hours.
RTX 4090 or RTX 5090
The RTX 5090 costs $0.49 an hour, 44% more, for 32 GB instead of 24 GB, about 78% more memory bandwidth (1,792 GB/s against 1,008 GB/s) and FP4 support. Pick the 5090 when a model does not fit in 24 GB or when memory speed limits you, as in LLM serving. Stay on the 4090 when the job fits: for SDXL batches or most renders, the cheaper card often gives more images or frames per dollar.
Rent or buy an RTX 4090?
Divide the price of a card by the hourly rate. If an RTX 4090 cost you $2,000, that money buys about 5,880 hours here at $0.34, or 6,450 hours at the Fleet rate: roughly eight to nine months of one card running day and night. Owning also means a host machine, power and cooling. Renting makes sense for projects that come and go, for testing before you buy, or when you need several cards for a short time.
RTX 4090 rental questions
How much does it cost to rent an RTX 4090?
$0.34 per GPU-hour on Petaflare, $0.32 on the 1,200-hour block and $0.31 on the 4,000-hour block. The smallest block is 75 GPU-hours for $25.50.
Is the RTX 4090 good for Stable Diffusion?
Yes. SDXL runs comfortably in 24 GB, including LoRA training; Flux needs FP8 weights to fit with its text encoder.
Can I run an LLM on a rented RTX 4090?
Models up to about 8B in BF16, 14B in 8-bit or 30B in 4-bit. For 70B models, rent a card with 48 GB or more.
Can I mine cryptocurrency on it?
No. Mining is not allowed on any Petaflare instance; the acceptable use policy lists every rule.
Sources
- NVIDIA GeForce RTX 4090, specificationsread 19 November 2026
- RunPod pricingread 19 November 2026
- Salad pricingread 19 November 2026
- Vast.ai public GPU price feedread 19 November 2026