Best RTX 4090 GPU Cloud (2026): Verified Hourly Prices

Only four clouds in our tracked set rent the RTX 4090 at all: RunPod, Vast.ai, TensorDock and Salad. The three with public price pages sit within two cents of each other, between $0.33 and $0.35 per hour. Picking between them is a question of interruption risk and billing model, not headline price.

Who actually rents an RTX 4090

Prices checked 28 July 2026 against each provider’s public pricing page.

Vast.ai: cheapest, with a caveat about which number you read

Vast.ai’s $0.33/h is the low end of a live marketplace, not a fixed rate. The platform’s own distribution for the 4090 puts the median at $0.38/h and the 90th percentile at $0.53/h, so a machine grabbed at a busy moment can cost 60% more than the headline. Interruptible instances drop to roughly $0.11/h, but the host can reclaim them mid-run. That is fine for checkpointed batch jobs and a poor fit for anything interactive.

RunPod: two tiers that differ by more than price

RunPod splits the 4090 into Community at $0.34/h and Secure at $0.69/h. The gap buys datacenter-operated hardware instead of third-party hosts, which matters if you are handling data you would rather not put on a stranger’s machine. For throwaway fine-tuning runs the Community tier is the obvious pick and lands within a cent of Vast.ai.

TensorDock: $0.35/h, thinner supply

TensorDock quotes from $0.35/h. The rate is competitive, but the fleet is smaller than RunPod’s or Vast.ai’s, so a specific region or a multi-GPU 4090 node is not always there when you want it. Check availability before you plan around it.

Why AWS, Azure, Google Cloud and Lambda are not on this list

None of them rent an RTX 4090. NVIDIA’s licence terms keep GeForce cards out of most datacenters, so the hyperscalers stock datacenter parts instead: AWS, Azure and Google Cloud all list A100 and H100, Lambda Labs runs A100, H100 and the Quadro RTX 6000. If you need a 4090 specifically, a marketplace is the only route. If you merely need 24GB of VRAM, an L4 or an A10 on a big cloud is the boring alternative.

What the 4090 is good at

At 165 TFLOPS FP16 and 24GB of VRAM, the 4090 handles QLoRA fine-tunes of 7B–13B models, Stable Diffusion and SDXL inference, and batch jobs you can checkpoint. The 24GB ceiling is the real constraint: full fine-tunes past about 13B parameters need an A100 80GB or an H100, and no amount of gradient checkpointing changes that.

FAQ

Is the RTX 4090 cheaper per unit of work than an A100?

Usually yes, for jobs that fit in 24GB. At $0.34/h and 165 TFLOPS FP16, the 4090 costs roughly $0.002 per TFLOP-hour; an A100 80GB at $1.09–$1.19/h and 312 TFLOPS costs about $0.0035. The A100 wins once your model no longer fits in 24GB and you would otherwise be splitting it across cards.

Why do the same 4090 listings show different prices hour to hour?

Vast.ai and RunPod Community are marketplaces where individual hosts set their own rates, so supply and demand move the price. Vast.ai’s published range for the 4090 spans $0.11 to $2.72/h. Filter and sort by price at the moment you rent rather than trusting a number you read in an article, including this one.

Can I rent an 8x RTX 4090 node?

Rarely, and not reliably. Consumer cards lack NVLink on the 4090, so multi-GPU scaling depends on PCIe bandwidth and falls off faster than it does on SXM A100 or H100 nodes. Multi-GPU 4090 boxes exist on Vast.ai and TensorDock, but for distributed training the interconnect, not the card, becomes the bottleneck.

Full provider-by-provider pricing across every GPU class is in our GPU cloud comparison.