NVIDIA RTX 4090 cloud prices

The cheapest RTX 4090 hours are on capacity that can be taken back, and which one is cheapest depends on how much risk you accept. Clore.ai's snapshot of 21 September 2026 found a free spot server at $0.10 per GPU-hour, but that was one server among 243 and a higher bid can take it 8. GPUBank's interruptible tier is $0.14 with five minutes of notice and the disk kept, and SaladCloud's lowest priority is $0.16 on consumer PCs with no persistent disk 1 9. For capacity that cannot be interrupted, GPUBank's $0.23 is the lowest all-in price we found, ahead of Novita AI at $0.33, Runpod Community at $0.34 and Vast.ai's cheapest listing at $0.36 before storage and bandwidth 16 11 7.

An RTX 4090 costs between $0.10 and $0.89 an hour to rent in October 2026, and most offers that include a usable machine fall between $0.23 and $0.45. The low end is spot and priority capacity: Clore.ai at $0.10 in its September snapshot, GPUBank interruptible at $0.14 and SaladCloud from $0.16 8 1 9. On demand, GPUBank charges $0.23, Novita AI $0.33, Runpod Community $0.34, Vast.ai from $0.36, CloudRift and packet.ai $0.39 and Beam $0.44 7 13 15. Datacenter-style fixed offers cost more, with Runpod Secure Cloud at $0.89 11. GetDeploying's median on-demand rate across twelve providers was $0.47 on 16 October 21. Compare what each price includes before you compare the numbers.

The large clouds do not rent the RTX 4090 because NVIDIA's driver licence says that GeForce software is not licensed for datacenter deployment. The clause is 2.8 in the current licence, dated 25 February 2025, and it also limits GeForce software to GeForce hardware you own 5. NVIDIA first added the restriction at the end of 2017 and said at the time that it wanted to keep GeForce cards out of demanding, large-scale enterprise environments 6. AWS, Google Cloud, Azure and Oracle therefore offer the datacenter L4 and L40S instead. The 4090 comes only from specialised GPU clouds, peer-to-peer marketplaces and networks of consumer PCs, so read each provider's terms, reliability record and data policy more carefully than you would a hyperscaler's.

Vast.ai's cheapest RTX 4090 listing is close to Runpod Community Cloud's price, but its typical listing costs more. On 16 October Vast's public offer search showed the cheapest single card at $0.36 on demand, a median of $0.53 across its listings and interruptible bids from $0.27 7, while Runpod listed $0.34 on Community Cloud, out of stock on GetDeploying that day, and $0.89 on Secure Cloud 11 21. Vast's rate covers the GPU and the host's CPU and RAM share, and storage and bandwidth come on top at each host's rate, with storage billed even while the instance is stopped. Runpod charges no egress but bills volumes at $0.10 per GB-month running and $0.20 idle. With a large model library kept all month, the storage line usually decides the comparison rather than the GPU rate.

SaladCloud's RTX 4090 is cheap because it runs on consumer PCs that their owners use, so a container can be moved off a machine at any time and no tier behaves like dedicated on-demand capacity. The prices are $0.330 at high priority, $0.273 at medium, $0.217 at low and $0.160 at the lowest batch tier, billed per second with vCPU and RAM included 9, unchanged since 12 September 2026 10. A higher-priority workload displaces a lower one, and there is no persistent disk, so model weights are pulled again on every new node. For batch image generation or transcription that retries failed items, it is one of the cheapest published hours. For a production endpoint, plan on at least two replicas, which doubles the bill to about $482 a month at high priority.

A marketplace RTX 4090 at about $0.10 an hour is a spot slot on one host, and what comes with it is whatever that host offers. Clore.ai's snapshot of 21 September 2026 showed the cheapest free spot server at $0.100 against a spot median of $0.365, with a 1.25 percent renter fee on spot and per-minute billing 8. Spot rentals are won by bid, so a higher bid can take the machine during a job, and the cheapest servers tend to give the card few CPU cores and little RAM. On Vast.ai the GPU rate excludes storage and bandwidth, which each host prices separately and bills while the instance exists 7. Benchmark the host for a few minutes and price the job, not the hour, before you commit a long run to it.

An on-demand RTX 4090 stays yours until you stop it, while a spot or interruptible one can be taken back by the provider or by a higher bid. The discounts are large: GPUBank's interruptible tier costs $0.14 against $0.23 on demand, about 40 percent less, Novita lists spot at $0.17 against $0.33, and SaladCloud's lowest tier is about half its high tier 1 21 9. The terms of the interruption differ far more than the discounts. GPUBank gives five minutes of notice, runs your checkpoint hook, stops the machine rather than deleting it and credits the last 30 minutes if the notice is missed 1. On Vast.ai and Clore.ai a higher bid takes the machine, and SaladCloud treats every tier, including its high priority, as interruptible 7 8 9. Use interruptible capacity for batch jobs that checkpoint often, and on-demand or reserved capacity for anything a user is waiting on.

A single RTX 4090 cannot run a 70B model properly, because the weights alone need about 40 GB at 4-bit precision and the card has 24 GB. Offloading most of the model to system RAM produces output, but the GPU then spends its time waiting on the PCIe bus. Two cards work: CloudRift served Llama 3.3 70B at 4-bit AWQ across two RTX 4090s with vLLM at 467 tokens per second under load, against 1,230 on two RTX 5090s and 1,031 on one 96 GB RTX PRO 6000 14. The 4090 has no NVLink, so the split pays a PCIe cost on every request. For a 70B model, a single card with enough memory is usually simpler and cheaper per token than a pair of 4090s. One 4090 is comfortable up to about 14B at 8-bit or 32B at 4-bit with a short context.

A rented RTX 4090 handles SDXL easily and runs FLUX.1 dev well at FP8, which is the precision most ComfyUI workflows use on 24 GB. In ComfyUI's community benchmark of the standard FLUX dev FP8 template, a 4090 took 11.28 seconds an image, against 8.78 seconds on an RTX 5090 and 26 seconds on an RTX 3090 22. At GPUBank's $0.23 an hour, 1,000 FLUX images cost $0.72 on the 4090 and $1.17 on the 5090 at $0.48, so the older card is the cheaper one per image and the newer one the faster 1. The 4090's limit is memory: FLUX.1 dev in BF16, several ControlNets at once or large upscales can exceed 24 GB, and that is when a 32 GB or 48 GB card earns its higher price.

Rent the RTX 4090 when the model fits in 24 GB and cost per result matters, and the RTX 5090 when you need its extra 8 GB, lower latency or FP4. On GPUBank the 4090 costs $0.23 and the 5090 $0.48, so the 5090 must be 2.1 times faster to match the 4090 per job 1. It is not on FLUX.1 dev at FP8, where it takes 8.78 seconds against 11.28, so 1,000 images cost $1.17 against $0.72 22. In llama.cpp single-user generation it reaches 300.40 tokens per second against 188.96, so a million tokens cost $0.44 against $0.34 23. In batched vLLM serving it doubles the throughput, 4,570 against 2,259 tokens per second, and the cost per token ends up about equal 14.

A full 730-hour month of one RTX 4090 costs about $105 to $650 depending on the provider and the terms. On GPUBank it costs $167.90 at the on-demand rate, $131.40 reserved for the month, or about $104.94 of real money when paid from a $1,000 top-up credited as $1,600, with a persistent disk adding $0.03 per GB-month 1. CloudRift's one-month reservation comes to $255.50 with local storage included, Runpod Community at $0.34 to $248.20 plus volume storage, and Vast.ai's cheapest $0.36 listing to $262.80 before host storage and bandwidth 13 11 7. SaladCloud's high tier is $240.90 a node, but an always-on service needs a second replica because its nodes can be reclaimed 9. Fixed datacenter-style offers cost far more: Oblivus at $0.64 comes to $467.20 and Runpod Secure Cloud at $0.89 to $649.70 21 11.

Buying a used RTX 4090 is cheaper than renting only if the card works almost around the clock. Used cards were valued at $2,214 on 6 October 2026 by GetPCParts, sold for a median of €2,185 on eBay according to PCPrice.watch, and were listed at a median asking price of $2,999 on DealHunter 26 27 28. With $2,200 for the card, $1,200 for the machine, $1,400 of resale value after two years, a 0.6 kW draw and US power at 18.19 cents per kWh 29, owning costs $83.33 a month plus $0.11 per hour of use. That beats GPUBank's $0.23 list rate only past about 690 hours a month, and never beats the $0.18 reservation or the $0.17 effective rate from a $100 top-up 1.

On most providers a stopped RTX 4090 instance keeps billing its storage. On GPUBank a stopped machine costs only its persistent disk at $0.03 per GB-month; compute, the IPv4 address and the local NVMe cost nothing until you start it again 1. Runpod bills an idle volume disk at $0.20 per GB-month and a network volume at $0.07 under 1 TB 11. On Vast.ai the host's storage rate runs for as long as the instance exists, whether it is running or not, at about $0.20 per GB-month on the cheap listings we saw 7. SaladCloud has no persistent disk, so nothing is billed, but nothing is kept either. Read the stop and delete terms before you leave an instance parked for a week.

Yes, LoRA and QLoRA fine-tuning of an 8B model fits on one RTX 4090, and QLoRA reaches models of about 14B parameters within 24 GB. GPUBank positions its RTX 4090 for LoRA training on 8B models and provides Unsloth, Axolotl and Kohya templates as well as PyTorch 2.8 with CUDA 12.8 1. A full 16-bit fine-tune of an 8B model does not fit, because the weights, gradients and optimiser state together need several times the card's memory. Training across several 4090s works for data-parallel runs where each card holds the whole model, but it is slowed by the lack of NVLink once the model has to be split. For long runs, the interruptible tier at $0.14 is attractive if the trainer saves checkpoints regularly, since a reclaim then costs minutes.

Community-cloud RTX 4090s run on machines owned by people you do not know, so treat them as unsuitable for regulated or confidential data. On Vast.ai, Clore.ai, Runpod Community Cloud and SaladCloud, independent hosts operate the hardware, and container isolation protects you from other tenants but not from the owner of the machine 7 8 9. Runpod sells its Secure Cloud, at $0.89 for this card, as the option in vetted datacenters 11. For public models, open datasets and creative work, community hosts are usually fine, and their low price is the point. For customer data, unreleased weights or work under contract, a managed offer whose operator runs its own regions is the cheaper choice once the risk is counted, even at a few cents more an hour.