Spheron GPU Catalog

Rent NVIDIA RTX 5090 GPUs on Demand from $0.86/hr

32GB GDDR7 Blackwell, deployed in under 2 minutes.

At a glance

You can rent an NVIDIA RTX 5090 on Spheron starting at $0.86/hr per GPU per hour on dedicated (99.99% SLA, non-interruptible), with spot instances cheaper still. Per-minute billing, no contracts, deployed in under 2 minutes across data center partners in multiple regions. The RTX 5090 packs 32GB of GDDR7 memory and 5th gen Tensor Cores, making it the best price-to-performance choice for LoRA/QLoRA fine-tuning of 7B-13B models, Stable Diffusion XL inference, local LLM serving with Ollama or vLLM, and general AI development work. Launch a container with your CUDA/PyTorch image, SSH in, and start training in minutes.

GPU ArchitectureNVIDIA Blackwell
VRAM32 GB GDDR7
Memory Bandwidth1.79 TB/s

Technical specifications

GPU Architecture
NVIDIA Blackwell
VRAM
32 GB GDDR7
Memory Bandwidth
1.79 TB/s
Tensor Cores
5th Generation
CUDA Cores
21,760
RT Cores
4th Generation
FP32 Performance
104.8 TFLOPS
FP16 Tensor (dense)
209.5 TFLOPS
FP8 Tensor (dense)
419 TFLOPS
INT8 Tensor (dense)
838 TOPS
FP4 Tensor (sparse)
3,352 TOPS
System RAM
24 GB DDR5
vCPUs
8 vCPUs
Storage
200 GB NVMe SSD
Network
PCIe Gen5
TDP
575W

Pricing comparison

ProviderPrice/hrSavings
SpheronYour price
$0.86/hr-
CloudRift
$0.65/hr-
NeevCloud
$0.69/hr-
RunPod (Community)
$0.69/hr-
RunPod (Secure)
$0.99/hr1.2x more expensive
Custom & Reserved

Need More RTX 5090 Than What's Listed?

Reserved Capacity

Commit to a duration, lock in availability and better rates

Custom Clusters

8 to 512+ GPUs, specific hardware, InfiniBand configs on request

Supplier Matchmaking

Spheron sources from its certified data center network, negotiates pricing, handles setup

Need more RTX 5090 capacity? Tell us your requirements and we'll source it from our certified data center network.

Typical turnaround: 24–48 hours

When to pick the RTX 5090

Scenario 01

Pick RTX 5090 if

Your workload is LoRA/QLoRA fine-tuning on 7B-13B models, Stable Diffusion XL or Flux inference, or local LLM serving where 32GB VRAM is plenty. You want the cheapest Blackwell-generation GPU with 5th gen Tensor Cores and aren't bottlenecked by multi-GPU interconnect.

Recommended fit
Scenario 02

Pick RTX 4090 instead if

You need the absolute lowest hourly rate and 24GB VRAM is enough for your model. Your workload doesn't benefit from Blackwell's 2x AI throughput or the bandwidth jump from GDDR6X to GDDR7.

Recommended fit
Scenario 03

Pick RTX PRO 6000 instead if

You need 48GB or 96GB VRAM on Blackwell silicon to serve 30B+ quantized models on a single GPU, or you want pro-tier drivers and ECC memory for production workloads.

Recommended fit
Scenario 04

Pick H100 instead if

You're training or fine-tuning 30B+ parameter models end-to-end, need HBM3 bandwidth and NVLink/InfiniBand for multi-GPU, or your workload requires the Hopper FP8 Transformer Engine.

Recommended fit

Ideal use cases

Use case / 01
🛠️

AI Prototyping & Development

Rapidly iterate on AI models at low cost, making the RTX 5090 ideal for development workflows and early-stage experimentation.

Model architecture experimentationRapid prototypingDevelopment and debuggingCI/CD ML pipelines
Use case / 02
🎯

Small Model Fine-Tuning

Perform LoRA and QLoRA fine-tuning of models up to 13B parameters with 32GB of fast GDDR7 memory.

Domain-specific fine-tuning (7B-13B models)Instruction tuningRLHF experimentsAdapter training
Use case / 03
💰

Cost-Effective Inference

Deploy smaller models at minimal cost for production inference workloads that demand high throughput at a budget-friendly price.

7B model inferenceChatbot deploymentImage classification APIsReal-time NLP services
Use case / 04
📚

AI Education & Research

Affordable GPU access for learning, research, and open-source contributions without the overhead of expensive data center GPUs.

ML courses and workshopsAcademic researchKaggle competitionsOpen-source model development

Performance benchmarks

Llama 3.1 8B Inference
~3,500 tokens/s
FP16, vLLM batched
Llama 3.1 8B (Q4_K_M)
~65 tokens/s
llama.cpp, single stream
Stable Diffusion XL
~16 img/min
1024x1024, base + refiner
Mistral 7B QLoRA
~720 tokens/s
INT4 fine-tuning
Memory Bandwidth
1,792 GB/s
GDDR7, 512-bit bus
vs RTX 4090
+28-50%
LLM tokens/s uplift

Serve Llama 3.1 8B on RTX 5090 with vLLM

Spin up an OpenAI-compatible inference endpoint on a single RTX 5090. The 32GB GDDR7 fits Llama 3.1 8B in FP16 with room for an 8K context window.

bash
Spheron
# SSH into your RTX 5090 instancessh root@<instance-ip> # Install vLLM (CUDA 12.x compatible)pip install vllm # Serve Llama 3.1 8B in FP16 on a single RTX 5090vllm serve meta-llama/Meta-Llama-3.1-8B-Instruct \  --dtype float16 \  --max-model-len 8192 \  --gpu-memory-utilization 0.9 \  --port 8000 # Test the OpenAI-compatible endpointcurl http://localhost:8000/v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "meta-llama/Meta-Llama-3.1-8B-Instruct",    "messages": [{"role": "user", "content": "Hello"}]  }'

RTX 5090 vs alternatives

Related resources

Frequently asked questions

How does the RTX 5090 compare to the RTX 4090?

The RTX 5090 features the next-generation Blackwell architecture compared to the RTX 4090's Ada Lovelace. Key improvements include 32GB GDDR7 memory (vs 24GB GDDR6X on the 4090), approximately 2x AI performance, 5th generation Tensor Cores (vs 4th gen), and significantly higher memory bandwidth. The RTX 5090 delivers a substantial leap in AI workload performance while maintaining consumer-grade affordability.

Is the RTX 5090 good for AI training?

The RTX 5090 is excellent for training small to medium models up to approximately 13B parameters. Its 32GB GDDR7 memory handles LoRA and QLoRA fine-tuning efficiently. For larger models requiring more VRAM or higher interconnect bandwidth, consider the H100 (80GB HBM3) or A100 (80GB HBM2e) for full-scale training workloads.

What AI models can I run on 32GB VRAM?

With 32GB of GDDR7, you can comfortably run Llama 3.1 8B (FP16, ~16GB), Mistral 7B (~14GB), Qwen 2.5 14B (FP16, marginal at ~28GB, needs context limits), Stable Diffusion XL, Flux.1 Dev, and Whisper Large V3. Quantized (Q4/INT4) versions of larger models such as Qwen 2.5 32B (~20GB) also fit. Llama 3.3 70B does not fit on a single RTX 5090 even at Q4; use an H100 or H200 for that class.

How does the RTX 5090 compare to the H100?

The H100 features 80GB HBM3 memory vs the RTX 5090's 32GB GDDR7, and is 2-3x faster for large-scale training workloads. However, the RTX 5090 is approximately 2x cheaper per hour and provides excellent performance for development, inference, and fine-tuning of smaller models. Choose the RTX 5090 for cost-effective development and the H100 for production-scale training.

Can I use the RTX 5090 for video and gaming workloads?

Yes! The RTX 5090 features 4th generation RT Cores, making it excellent for real-time ray tracing, video editing, game development, and 3D rendering workloads. It is a versatile GPU that handles both AI/ML and creative professional workloads with outstanding performance.

What deep learning frameworks work with the RTX 5090?

All major deep learning frameworks are fully supported: PyTorch, TensorFlow, JAX, and ONNX Runtime. The RTX 5090 has full CUDA 12.x support, ensuring compatibility with the latest framework versions, libraries, and tools in the AI/ML ecosystem.

What's the minimum rental period?

There's no minimum rental period! Spheron charges with per-minute billing granularity. Rent an RTX 5090 for as little as a few minutes to test your workload, or keep it running as long as you need. You only pay for what you use with no long-term contracts or commitments.

Is 32GB VRAM enough for fine-tuning?

Yes, 32GB is well-suited for LoRA and QLoRA fine-tuning of models up to 13B parameters. Full fine-tuning works for 7B-class models. For full fine-tuning of larger models (30B+), consider the H100 with 80GB HBM3. The RTX 5090's fast GDDR7 memory also helps accelerate data loading during the fine-tuning process.

What regions are RTX 5090 GPUs available in?

RTX 5090 GPUs are currently available in US, Europe, and Canada regions. We're continuously expanding capacity and availability. Check our app or contact sales for specific region requirements and current availability.

Do you offer support for production deployments?

Our platform is plug-and-play for standard deployments. For 100+ GPU clusters, you get dedicated support via Slack or Discord, plus sourcing assistance. Enterprise customers get dedicated support channels and SLA guarantees.

Book a call with our team

Can I run RTX 5090 on Spot instances? What are the risks?

Yes. Spot is the interruptible tier of on-demand, priced up to 70% off the dedicated rate. Dedicated instances carry a 99.99% SLA and are non-interruptible; spot instances can be terminated when capacity is reclaimed by a dedicated workload. Key risks: job interruption during training/inference, loss of unsaved state, restart from last checkpoint. Best practices: checkpoint every 15-30 minutes, use spot for fault-tolerant or development workloads, save model weights to persistent storage, and run production serving on dedicated instances. Given the RTX 5090's already-low base price, spot makes it an exceptionally budget-friendly option for experimentation.

Also consider