Where to rent an RTX 4090 at the cheapest price

The cheapest RTX 4090 hours come from consumer machines because the people who own them have already paid for the card. A host on Vast.ai or Clore.ai, or a gamer whose PC runs SaladCloud's agent, bought the 4090 for play or as a side business, pays household electricity and no rent, and is happy with anything above the cost of power, which is how a card that resells for about $2,200 ends up rented for ten to thirty-six cents an hour²⁶²⁵. Because the licence keeps those cards out of the big clouds⁸, a large share of the supply sits in homes, small server rooms and colocated racks rather than in the facilities a renter would picture. The floors show it: Vast's cheapest on-demand single card was $0.36 and its median $0.53, Clore's spot floor was $0.10 against a $0.375 on-demand median, and Salad's lowest tier was $0.16⁵²³.

For the renter that has three consequences. The floor is real but thin: Clore's $0.10 was one free server among 243², Vast's listings change by the hour, and the cheapest slot is often the one with the fewest CPU cores, the slowest disk or the weakest uplink, so the price that matters is the cost per finished job after a short benchmark on the host you rent. Interruption is part of the deal, since the owner can take the machine back or a higher bid can take it from you, and on most of these networks the renter pays for the lost time. And the hardware is an end-of-life consumer card without ECC that has often run hot for years, which is fine for image generation and experiments and a poor fit for customer data, unreleased weights or anything a user waits on. A managed all-in rate costs a few cents more an hour and buys a known machine, a notice period and somebody who answers for the host.

The GeForce RTX 4090 is NVIDIA's Ada Lovelace flagship on the AD102 chip, with 16,384 CUDA cores, 512 fourth-generation Tensor Cores, 24 GB of GDDR6X on a 384-bit bus and a 450 W total graphics power that NVIDIA pairs with an 850 W minimum supply⁹. Its 1,008 GB/s of memory bandwidth is what sets the speed at which it writes tokens in a language model, Clore.ai gives its dense FP16 and BF16 tensor throughput as 165 TFLOPS, noting that the 330 TFLOPS sometimes quoted assumes 2:4 sparsity, and its Tensor Cores run FP8 natively², the precision at which FLUX.1 dev fits in its memory and runs well. It went on sale on 12 October 2022 at $1,599¹⁰, was replaced at the top of the range by the RTX 5090, on sale from 30 January 2025 at $1,999, and is no longer in production, so the supply of new cards has dried up and the used market sets the price of the hardware that hosts rent out. It has no NVLink, so two cards in one machine are two separate 24 GB devices joined over PCIe.

The large clouds do not rent it, and the reason is a licence clause rather than a missing price list. Clause 2.8 of the current NVIDIA driver licence says that GeForce software is licensed only for GeForce hardware you own and is not licensed for datacenter deployment⁸, a restriction NVIDIA first introduced at the end of 2017 to keep consumer cards out of large enterprise installations¹¹. The supply therefore comes from specialised GPU clouds, peer-to-peer marketplaces and networks of consumer PCs, where the 4090 is one of the most listed cards of all: Clore.ai alone counted 378 of them on 243 servers in its snapshot of 21 September².

People rent it for ComfyUI, Stable Diffusion WebUI Forge and InvokeAI, where SDXL fits easily and FLUX.1 dev runs at FP8, for LoRA and QLoRA fine-tuning of models up to about 8B to 14B parameters, and for single-card inference of language models up to about 14B at 8-bit or about 32B at 4-bit with a short context; Clore's rule of thumb of half a gigabyte per billion parameters at 4-bit puts a 32B model near 16 GB of weights and a 70B model near 40 GB, beyond one card². GPUBank positions its RTX 4090 for ComfyUI, SDXL and LoRA training on 8B models¹. The price range follows from where the cards sit: a card in a spare room costs its owner little more than electricity, a card in a rack carries the cost of the building, the network and the staff, and the rate reflects which of the two you are renting.

The fair way to compare RTX 4090 offers is to group them by what you are actually buying before looking at the number. A managed on-demand offer sells a defined machine at a fixed hourly rate that stays in force until you stop it, and the useful question is what that rate leaves out: GPUBank, CloudRift and Beam include the local disk and charge nothing for traffic¹¹²¹³, while Runpod bills volumes separately and packet.ai bills egress at $0.04 per GB⁴⁷. A marketplace listing sells a slot on a host whose CPU share, memory, disk speed, PCIe generation and reliability vary from one listing to the next, at a GPU rate the host sets, with storage and bandwidth as separate lines. Putting the two side by side without saying so is how most cheapest-GPU lists end up comparing things that cannot be compared.

The other three kinds of offer each trade something for the price. Spot and priority tiers can be taken back, either by a higher bid on Vast.ai and Clore.ai or by a higher-priority workload on SaladCloud, and they differ less in their discount than in what happens at the moment of interruption: some delete the container, some leave a disk you keep paying for, and only a few give notice. Reserved capacity lowers the rate in exchange for paying for a week, a month or longer whether you use it or not, and for this card it ranges from GPUBank's $0.18 for a single card to Clore.ai's five-GPU and eight-GPU bare-metal blocks at $0.40 to $0.94¹². Serverless workers bill only the seconds spent on requests, at an hourly equivalent of $0.69 on Beam for the GPU alone and $1.10 for a Runpod flex worker¹³⁴, which only pays when traffic is rare and bursty: once a $0.69 worker is busy for a third of the month, a dedicated $0.23 machine already costs the same.

The captures below were taken on 16 Oct 2026 from each provider's own RTX 4090 or pricing page, so every figure can be read next to the conditions printed beside it.

Several managed offers sit just above the providers shown in the cards. Novita AI's own RTX 4090 page lists $0.33 an hour on demand and warns that the price follows the computing power market, while GetDeploying showed its single-card configuration with 16 vCPU and 125 GB of RAM out of stock on 16 October, along with a $0.17 spot rate and a 12-month four-card reservation at $0.27 per GPU, all sold out⁶¹⁴. packet.ai rents a dedicated card at $0.39 with hourly billing, local NVMe and $0.04 per GB of egress, and its own page names both US Central and EU West as the region⁷. Beam's on-demand machine with 8 vCPU, 64 GB of RAM and 500 GB of NVMe starts at $0.44 with no egress fees¹³. Spheron lists 8 vCPU, 24 GB of RAM and 500 GB of storage at a last known $0.58, billed per minute after a 20-minute minimum, but its page said no RTX 4090 capacity was live when we read it¹⁵.

The marketplace tail is thinner than its reputation. Lium's own page shows a reference price of $0.30 with no pod available, while GetDeploying listed one Lium node in stock at $0.62¹⁶¹⁴, and TensorDock's page quotes a typical $0.35 that varies by host, with vCPU at $0.003 an hour, RAM at $0.002 per GB-hour and NVMe at $0.00005 per GB-hour billed on top, on a pricing page last updated in July 2024¹⁷. Further up, GetDeploying tracked Oblivus at $0.64 on demand and out of stock, a 36-month Oblivus reservation at $0.44, GPUhub at $0.44, SwissGPU at $0.48, Database Mart reservations at $0.51 to $0.76 and Runpod's Secure Cloud at $0.89 in stock, for a median on-demand rate of $0.47 across twelve providers on 16 October¹⁴. io.net's explorer, which our first survey quoted at $0.25, would not load for us, so it is left out of the comparison.

Storage is the line that most often changes the ranking for this card, because image and video work keeps large model libraries between sessions. On GPUBank the local NVMe is part of the rate and a persistent disk costs $0.03 per GB-month, so 100 GB of checkpoints and LoRAs costs $3 a month¹. Runpod charges $0.07 per GB-month for a network volume under 1 TB and $0.20 for an idle volume disk⁴. Vast.ai bills each host's storage rate for as long as the instance exists, stopped or not, and the cheap listings we queried charged about $0.20 per GB-month, so the same 100 GB costs about $20⁵. SaladCloud has no persistent disk at all, so weights are pulled again whenever the container lands on a new PC³.

Billing granularity and lost time are the other two hidden lines. A 61-minute ComfyUI session at $0.23 costs $0.23 when billed per second, as on GPUBank, Runpod, CloudRift, SaladCloud and Vast.ai, and $0.46 when the meter rounds up to the started hour, while Spheron's 20-minute minimum makes a five-minute test cost twenty minutes¹⁵. Lost time matters more on the cheapest offers, because on a marketplace the renter usually pays for it: an outbid spot instance is gone and billed to the minute, a host that runs the card on PCIe 3.0 or gives it two CPU cores makes the job longer at your expense, and the cheapest Vast listing reported a measured memory bandwidth of 876 GB/s for a card rated at 1,008⁵. Clore's own guide says renters are not charged for the downtime when an instance goes down unexpectedly, but a slow host is still paid in full¹⁹. GPUBank does not bill a failed launch and credits a session interrupted by a host failure¹, which is why the scenarios below add an explicit allowance for lost time to every marketplace and spot row.

Per job rather than per hour, the RTX 4090 is the cheapest card GPUBank rents for anything that fits in 24 GB. On FLUX.1 dev at FP8, 1,000 images cost $0.72 on the 4090 at 11.28 seconds each and $1.17 on the RTX 5090 at 8.78 seconds and $0.48 an hour²⁰, so the 5090 buys speed rather than savings. For a single user in llama.cpp, a million tokens cost $0.34 on the 4090, $0.44 on the 5090 and $1.31 on an 80 GB A100 at GPUBank's $0.95 A100 rate, because the A100 generates only 6 percent faster here²⁴. Under batched vLLM serving the 4090 and the 5090 converge at about $0.03 per million tokens each, since the 5090's 4,570 tokens per second are almost exactly twice the 4090's 2,259 at a little more than twice the price²⁵.

The other cards earn their price through memory rather than speed. The RTX 5090 at $0.48 adds 8 GB and FP4 support, which is enough for FLUX at higher precision, more LoRAs loaded at once or a 32B model with a usable context. The L40S at $0.55 brings 48 GB of ECC memory, which suits BF16 FLUX, stacked ControlNets and high-resolution upscaling, although we found no measurement on the same tests to price it per job. The A100 SXM at $0.95 has 80 GB of HBM2e at 2.0 TB/s and is the card for models and batches that need it, not a faster 4090: to match the 4090 per job it would have to finish 4.1 times sooner, and on the workloads measured here it is barely faster at all¹²⁴.