The NVIDIA GeForce RTX 4090 is the card that made local AI mainstream: the full AD102 Ada Lovelace die with 16,384 CUDA cores, fourth-generation Tensor Cores with FP8, and 24 GB of GDDR6X at 1 TB/s. It is faster in raw FP32 than an A100 and, at packet.ai's price, roughly a quarter of the cost per hour.
packet.ai rents the RTX 4090 as a Dedicated single-tenant card for $0.39/hr: the whole GPU, nobody else on it, 99% SLA, billed hourly. It is the entry point of the rate card and, for 7–13B inference, image generation and embeddings, often the only card you need. Dynamic pricing for this SKU is launching soon.
Why rent the RTX 4090 on packet.ai
Dedicated at a Community Cloud price
$0.39/hr is what other clouds charge for a shared or marketplace 4090. On packet.ai it is a single-tenant card in a US data centre with an SLA. You are not renting from an anonymous host.
Ada Lovelace throughput
82.6 TFLOPS FP32 and 330 TFLOPS FP16 Tensor (660 with sparsity), with FP8 support. For 7B and 8B models in FP16 or 13B in 4-bit, the 4090 serves hundreds of tokens a second with room to spare.
24 GB GDDR6X at 1 TB/s
Enough for Llama 3.1 8B at full precision with a long context, Mistral 7B, Qwen 14B in 4-bit, SDXL with several LoRAs, or a large embedding model with a big batch. Bandwidth is the bottleneck for token generation and 1,008 GB/s is a lot at this price.
Cheapest hour on the rate card
Prototyping, notebooks, evaluation runs, small fine-tunes and CI jobs all cost cents. Leave it up all month and it is $285 at 730 hours, before monthly discounts.
Full CUDA ecosystem
PyTorch, TensorFlow, JAX, vLLM, llama.cpp, ComfyUI, Automatic1111, TensorRT: everything targets Ada. Driver and library support is the most mature of any consumer card.
RTX 4090 pricing on packet.ai
Three ways to buy the same card. Mix them on one account: Dynamic for development and bursty work, Dedicated for production, Clusters when a single node is not enough.
Dynamic
Shared card, scheduler-enforced isolation. Same peak performance and VRAM as Dedicated.
- Hourly billing, no commitment
- SSH-ready in under 5 minutes
- Never pre-empted
Dedicated
Whole card, exclusively yours. Zero scheduler interference, predictable p99.
- 99% SLA
- Single-tenant hardware
- 5–10 minute provisioning
Clusters
Multi-node, InfiniBand fabric, custom storage, named TAM.
- Bare-metal nodes and 8× configurations
- Up to 1,024+ GPUs
- 2–6 week provisioning
How it compares
Published on-demand starting rates for the same GPU at other providers, USD per GPU-hour. packet.ai's reference rate is $0.39/hr Dedicated.
| Provider | RTX 4090 rate |
|---|---|
| packet.ai | $0.39/hr Dedicated |
| Vast.ai (marketplace) | $0.35/hrlower by 10% |
| ShadeFarm | $0.60/hrpacket.ai −35% |
| RunPod Secure Cloud | $0.74/hrpacket.ai −47% |
| RunPod Serverless (active) | $1.10/hrpacket.ai −65% |
| Lambda / CoreWeave / AWS | Not offered |
Specifications
| Feature | Value |
|---|---|
| GPU architecture | NVIDIA Ada Lovelace (AD102) |
| CUDA cores | 16,384 |
| Tensor Cores | 512 (4th generation, FP8 / FP16 / BF16 / TF32) |
| RT Cores | 128 (3rd generation) |
| Boost clock | 2,520 MHz (reference) |
| Memory | 24 GB GDDR6X |
| Memory interface | 384-bit |
| Memory bandwidth | 1,008 GB/s |
| FP32 compute | 82.6 TFLOPS |
| FP16 Tensor compute | 330 TFLOPS dense / 660 TFLOPS sparse |
| FP8 Tensor compute | 660 TFLOPS dense / 1,321 TFLOPS sparse |
| Interconnect | PCIe 4.0 ×16 (no NVLink) |
| Media engines | 2 × NVENC (8th gen, AV1), 1 × NVDEC |
| Board power | 450 W |
What you can run on 24 GB
Where the RTX 4090 is the right card, where it is comfortable, and where you should step up.
| Workload | Fit | Notes |
|---|---|---|
| 7–8B LLM inference at FP16 | Single card, long context | Llama 3.1 8B, Mistral 7B, Gemma 2 9B |
| 13–14B LLM inference (4-bit) | Single card | Qwen 2.5 14B, Phi-4 at Q4 |
| SDXL / SD 1.5 image generation | Full precision with LoRAs and ControlNet | FLUX.1 dev in FP8 fits; schnell is comfortable |
| Embedding and reranking services | Large batches | BGE, E5, ColBERT at high QPS |
| Whisper / TTS / speech | Real-time and batch | Whisper large-v3 with plenty of headroom |
| LoRA fine-tuning to 8B | Single card with QLoRA | Hourly billing suits short runs |
| 30B+ models or 70B inference | Not enough VRAM | Step up to RTX 6000 Pro (96 GB) or A100 (80 GB) |
Deploy in minutes
- Create an accountSign up at dash.packet.ai. No credit card required to start; billing is usage-based.
- Pick the RTX 4090 and a tierChoose Dedicated (Dynamic launching soon), hourly or monthly billing, and the region. Multi-GPU and bare metal are on request.
- DeployFrom the dashboard, the API, or the packet CLI. Dedicated instances are typically SSH-ready in five to ten minutes.
- SSH in and runUbuntu, current NVIDIA drivers and CUDA are in place, with local NVMe attached. Bring your containers or install directly.
# after the instance shows Running in the dashboard ssh ubuntu@<instance-ip> nvidia-smi # confirm the RTX 4090 and driver # example: serve a model with vLLM pip install vllm python -m vllm.entrypoints.openai.api_server --model <hf-model-id> --port 8000
CLI and API reference, including how to launch and stop instances programmatically, is in the packet.ai docs.
FAQ
How much does the RTX 4090 cost on packet.ai?
$0.39 per GPU-hour on Dedicated capacity, billed hourly, no minimum term. That is about $285 a month at 730 hours; monthly commits take up to 20% off. Dynamic pricing for this card is launching soon.
Is it really dedicated at that price?
Yes. The RTX 4090 is sold on the Dedicated tier: a single-tenant card committed to your account with a 99% SLA. The rate is lower than most providers' shared or marketplace 4090 tiers because packet.ai's scheduler keeps its fleet busy across the whole rate card, not because this card is shared.
How does the RTX 4090 compare to the L40S and A100?
The 4090 has the highest FP32 throughput of the three and the lowest price, but the least memory (24 GB vs 48 GB and 80 GB) and no ECC. Choose it for models that fit in 24 GB and for image generation; choose the L40S for 48 GB and ECC in a data-centre part; choose the A100 for 80 GB and 2 TB/s of HBM bandwidth.
Can I run multiple 4090s together?
Multi-GPU configurations are available on request. The card has no NVLink, so inter-GPU traffic goes over PCIe 4.0; for models that need tensor parallelism across cards, an A100, RTX 6000 Pro or B200 is a better base.
Which region?
US Central (Texas) today. Check the pricing page region filter for the live list.
What comes pre-installed?
Ubuntu with current NVIDIA drivers and CUDA, SSH access and local NVMe. Install vLLM, llama.cpp, ComfyUI or any framework directly, or run your own containers.
How fast is provisioning?
Dedicated single-GPU instances are typically SSH-ready in five to ten minutes. No quota approval and no credit card to start.
Any hidden fees?
No platform fees, no ingress charges. Egress is $0.04/GB. Local NVMe up to 2 TB per node is included.
Same silicon. Smarter economics.
RTX 4090 from $0.39/hr, billed hourly, live in EU West. No quota approval, no credit card to start.