RTX 5090 rental with crypto: 32 GB of GDDR7 for $0.49 an hour
NVIDIA's Blackwell flagship for creators, rented here for crypto only: a third more memory than the RTX 4090 and almost twice its bandwidth, with FP4 Tensor Cores. Whole cards, one to eight per machine.
RTX 5090 rental on Petaflare costs $0.49 per GPU-hour, $0.47 on the 1,000-hour Project block and $0.44 on the 3,000-hour Fleet block; the smallest purchase is 60 GPU-hours for $29.40. Each card has 32 GB of GDDR7 at about 1.8 TB/s. That is enough for Flux in BF16, most video-generation models with FP8 weights and 32B language models in 4-bit with room for context. You get whole cards with root SSH, paid in crypto with no KYC.
RTX 5090 specs
Blackwell generation, released in 2025. Fifth-generation Tensor Cores add FP4, and the card keeps RT cores, three ninth-generation NVENC encoders and two NVDEC decoders.
- Architecture
- Blackwell
- Memory
- 32 GB GDDR7
- Memory bandwidth
- 1,792 GB/s
- BF16 Tensor (dense)
- not published as a Tensor figure for GeForce
- FP8 Tensor (dense)
- supported, plus FP4 (5th-gen Tensor Cores)
- Board power
- 575 W
- GPU link
- PCIe Gen5, no NVLink
- Form factor
- PCIe
- CUDA cores
- 21,760 CUDA cores
CUDA cores, memory, encoders and board power from NVIDIA's RTX 5090 page. Memory bandwidth is worked out from its 28 Gbps GDDR7 on a 512-bit bus, as NVIDIA's page does not list it.
RTX 5090 rental prices compared
One RTX 5090, as listed on each provider's pricing page on 19 November 2026.
| Provider | Price type | Per GPU-hour |
|---|---|---|
| Petaflare | Fixed, on demand | $0.49 · $0.44 on Fleet |
| RunPod Secure Cloud | Fixed, on demand | $1.39 |
| RunPod Community Cloud | Third-party hosts | $0.69 |
| Salad | High priority · lowest priority (interruptible) | $0.60 · $0.25 |
| Vast.ai | Marketplace, median of the day | $0.46 |
Sources: RunPod, Salad, Vast.ai public price feed (18 November median, cited with attribution). Read on 19 November 2026. All blocks are on what a GPU-hour costs.
Reading those RTX 5090 prices
Our $0.49 is about a third of RunPod's Secure Cloud price and below its Community Cloud and Salad's high-priority tier. The Vast.ai median is three cents lower, and our Fleet block at $0.44 goes under it. The $0.25 offers you will see are interruptible capacity, Salad's lowest priority or spot instances elsewhere: the machine can be taken back, which suits restartable batch jobs, not interactive work.
Sixty RTX 5090 hours cost $29.40 and 3,000 hours cost $1,320, or $0.44 an hour. Send USDT, USDC, Bitcoin, Ether, Solana, Litecoin or Monero; on a small block, Solana keeps the network fee to a fraction of a cent. We take nothing but crypto: here is the step-by-step payment guide.
What fits in 32 GB but not in the RTX 4090's 24 GB
Weights take 2 bytes per parameter in BF16, 1 in FP8 and about 0.5 in 4-bit. The extra 8 GB matters most for models that sit just above 24 GB.
| Job | Memory needed | RTX 4090 (24 GB) | RTX 5090 (32 GB) |
|---|---|---|---|
| Flux.1 dev in BF16, text encoder in FP8 | About 24 GB plus 5 GB | No | Yes |
| Wan 2.1 14B video model in FP8 | About 14 GB plus encoders and activations | Tight | Yes |
| A 32B LLM in 4-bit with an 8k context | About 18 to 20 GB plus cache | Short contexts only | Yes |
| A 14B LLM in BF16 | About 28 GB | No | Yes, little cache |
| SDXL LoRA training | About 12 to 16 GB | Yes | Yes, larger batches |
| A 70B LLM in 4-bit | About 35 to 40 GB | No | No: use an L40S or larger |
Rules of thumb; test your workflow on a 60-hour Burst block first. The RTX 4090 page covers the jobs that already fit in 24 GB.
Bandwidth and FP4 on a GeForce card
The RTX 5090's 1,792 GB/s is 78% more than the 4090's 1,008 GB/s and roughly twice the L40S's 864 GB/s. Token generation in an LLM is limited mostly by how fast weights stream from memory, so for models that fit, the 5090 serves noticeably faster than both. The LLM inference guide explains how to estimate it.
Its Tensor Cores also run FP4, the format Blackwell introduced. FP4 halves the memory of FP8 weights and speeds up inference where your framework and model support it; accuracy depends on how the model was quantized. For image and video work, FP8 checkpoints are more common today, and the card handles those too.
Software for the RTX 5090
Blackwell cards need a recent stack. CUDA 12.8 was the first release to support them, and PyTorch added Blackwell builds with its CUDA 12.8 wheels from version 2.7. If you bring your own Docker image, base it on CUDA 12.8 or newer, and check with nvidia-smi and torch.version.cuda after you connect. For ComfyUI on a rented card, install a PyTorch build for CUDA 12.8 or newer before the custom nodes. Renderers such as Blender, Redshift and Octane support the RTX 50 series in their recent versions; check the one you pin.
Power and multi-card machines
Each RTX 5090 draws up to 575 W, so a machine with several of them is a serious server, not a desktop. Cards talk over PCIe Gen 5, without NVLink, which suits jobs where each card works on its own batch: rendering frames, generating images in parallel or serving one model per card. For one model split across cards, the L40S has 48 GB per card, and SXM machines with NVLink are covered on the GPU server page.
RTX 5090 rental questions
How much does it cost to rent an RTX 5090 per hour?
$0.49 per GPU-hour on Petaflare, $0.47 on the 1,000-hour block and $0.44 on the 3,000-hour block. The smallest block is 60 GPU-hours for $29.40.
Is the RTX 5090 worth it over the RTX 4090?
When your model needs more than 24 GB or your job is limited by memory speed, yes: 32 GB and 78% more bandwidth for 44% more per hour. If the job fits comfortably in 24 GB, the RTX 4090 at $0.34 often costs less per result.
Can I run Flux on a rented RTX 5090?
Yes. Flux.1 dev fits in BF16 with an FP8 text encoder, which the 24 GB RTX 4090 cannot do without FP8 weights for the main model.
Is the RTX 5090 good for cloud gaming?
Petaflare rents compute instances over SSH, not game-streaming desktops. The card would cope, but our machines are set up for AI, rendering and other compute work.
Sources
- NVIDIA GeForce RTX 5090, specificationsread 19 November 2026
- RunPod pricingread 19 November 2026
- Salad pricingread 19 November 2026
- Vast.ai public GPU price feedread 19 November 2026