Where to rent an RTX 5090 at the cheapest price

The cheapest hours on 15 Oct 2026 were on capacity that can be taken back: SaladCloud's lowest priority at $0.25 an hour on consumer PCs, then GPUBank's interruptible tier at $0.29 with five minutes of notice and the disk kept. Marketplace listings come and go: Vast.ai advertised $0.37 for the GPU alone on 15 October but showed no current offers on 14 and 16 October, and Clore.ai's own snapshot put its on-demand median at $0.542. HyperAI lists $0.35, with 16 vCPU and 40 GB of RAM according to GetDeploying, but publishes no disk or traffic terms and warns of queues at peak hours. Among managed clouds that state the whole machine, GPUBank was lowest at $0.48 with 9 vCPU, 35 GB of RAM, a 500 GB NVMe disk, a public IPv4 address and all traffic included, and a top-up bonus takes the effective rate to between $0.40 and $0.30.

Expect $0.25 to $1.19 an hour for on-demand or priority capacity in October 2026, with a median on-demand price of $0.72 across the 16 providers GetDeploying could price on 16 October. Managed offers we read at source sit at $0.48 (GPUBank), $0.65 (CloudRift), $0.72 (Beam) and $1.19 (Runpod Secure Cloud), with Runpod's Community Cloud at $0.69. Reservations run from GPUBank's $0.38 to CloudRift's $0.55 to $0.62, and serverless GPU time costs $1.09 an hour on Beam before CPU and memory and $1.58 on Runpod. A full 730-hour month at $0.48 comes to $350.40.

Because NVIDIA's GeForce driver licence says, in clause 2.8, that the software "is not licensed for datacenter deployment", a restriction NVIDIA first added in December 2017. The large clouds therefore build on datacenter parts such as the L4, the L40S and the H100, and the RTX 5090 is rented out by specialised GPU clouds and peer-to-peer marketplaces instead. None of AWS, Google Cloud, Azure, Oracle, Lambda or CoreWeave listed an RTX 5090 when we checked.

One card serves models up to about 32 billion parameters at 4-bit, around 14 billion at 8-bit and 8 billion at 16-bit, with room left for the context. Hardware Corner measured Qwen3 32B at 4-bit in llama.cpp at 61 tokens per second with a 4K context, and the card reached roughly 47,000 tokens of context before its memory was full. A 70B model needs two RTX 5090s or a 96 GB card such as the RTX PRO 6000; LoRA and QLoRA on 7B to 14B models fit, a full 16-bit fine-tune of a 7B model does not.

It is faster and carries 8 GB more memory, but it is not always cheaper per result. In ComfyUI's community benchmark FLUX.1 dev at FP8 takes 8.78 seconds an image on a 5090 against 11.28 on a 4090, so at GPUBank's rates a thousand images cost $1.17 on the 5090 and $0.72 on the 4090. For batched LLM serving CloudRift measured 4,570 against 2,259 tokens per second with vLLM, which makes the cost per token about the same. Rent the 4090 when 24 GB is enough, and the 5090 when you need the memory, lower latency or FP4.

Rent the RTX 5090 for models that fit in 32 GB and a handful of concurrent users, and the H100 for larger models, long contexts or training that spans several GPUs. The H100 SXM has 80 GB of HBM3 at 3.35 TB/s and NVLink between cards, the RTX 5090 32 GB at 1.8 TB/s and no NVLink. On GPUBank the H100 SXM costs $1.85 an hour against $0.48, and for a 70B model at 4-bit the 96 GB RTX PRO 6000 at $1.15 is often the cheaper middle ground.

Yes: it is the fastest consumer card for image generation, and its 32 GB holds FLUX.1 dev with room for LoRAs and ControlNets. In ComfyUI's community benchmark of the default FLUX dev FP8 workflow it took 8.78 seconds an image, and one user reached 5.46 seconds with optimisations, against 11.28 on an RTX 4090 and 26 on an RTX 3090; NVIDIA claims just over five seconds with an FP4 build. At GPUBank's $0.48 a thousand FP8 images cost about $1.17, or $0.73 at the tuned speed, against $0.72 on an RTX 4090, so the 4090 stays cheaper per image when 24 GB is enough and the 5090 wins on wait time and headroom. GPUBank's ComfyUI templates, including one for Wan 2.2 video, launch on it directly.

Sometimes for inference, and rarely for training. Two cards give 64 GB, enough for a 70B model at 4-bit, and one published run measured Llama 3.3 70B at 26.9 tokens per second across two RTX 5090s with Ollama, but every exchange between the cards crosses PCIe because the 5090 has no NVLink, so tensor-parallel serving and multi-card training scale poorly. On GPUBank two RTX 5090s cost $0.96 an hour against $1.15 for one 96 GB RTX PRO 6000, which holds the same 70B model on a single card without splitting it and is usually the cheaper answer per token. Data-parallel jobs, where each card holds its own copy of a model that fits in 32 GB, scale well on two or four cards.

Yes. GPUBank reserves the RTX 5090 at $0.38 an hour for a week or a month, prepaid on cards set aside for you, cancellable before the start date and not refundable once started, with the same machine as on demand; a 730-hour month comes to $277.40. CloudRift reserves it at $0.62 for a week, $0.59 for a month and $0.55 for three months, and LeaderGPU rents a single card at €449 a month or an eight-card server at €2,690 a month, both excluding VAT. Clore.ai's bare-metal blocks start at eight cards on 14 or 30-day terms at $0.64 to $1.06 per GPU-hour. A reservation only pays if the card will be busy most of the term: at GPUBank's prices it beats on-demand use past about 578 hours a month.

Renting is cheaper unless the card does real work for most of every day over two years. With the card at $4,999, a $1,500 host, $2,000 of resale value after two years and US power at 18.19 cents per kWh, owning costs about $187.46 a month plus $0.13 per hour of load, which only beats renting at $0.48 past about 538 hours a month, and never beats the reserved or interruptible rates. At the $1,999 launch price the break-even falls to about 179 hours.

They can move a monthly bill more than the gap between two headline rates. Per-second billing charges a 61-minute job for 61 minutes, an hourly meter charges it for 120. Vast.ai keeps billing storage while an instance is stopped and bills bandwidth per byte at host rates, Runpod charges $0.20 per GB-month for an idle volume, and GPUBank charges only a persistent disk at $0.03 per GB-month with traffic included. For 100 hours a month with 200 GB of models those terms decide whether the bill is about $54 or closer to $60 to $70.

It depends on the host, which is what a marketplace is. Vast.ai shows a reliability score and a maximum rental duration per listing, Clore.ai shows ratings and the hardware around the card, and both let you filter on them; some RTX 5090 hosts run older CPUs on PCIe 3.0, which slows data loading. On marketplaces the renter usually pays for time lost to a failing host, whereas GPUBank does not bill a failed launch and credits the whole session when a host fails during a job.

Use a persistent disk or volume and check what it costs while nothing runs. On GPUBank a persistent disk at $0.03 per GB-month is the only charge on a stopped machine, so 200 GB costs $6.00 a month. Runpod's network storage costs $0.07 per GB-month under 1 TB, Vast.ai bills an instance's storage for as long as the instance exists, and SaladCloud containers have no persistent disk, so the weights travel in the image or in object storage.

Yes for jobs that checkpoint often or can simply be retried, no for interactive sessions. GPUBank's interruptible RTX 5090 costs $0.29 instead of $0.48, gives five minutes of notice through the metadata endpoint, a webhook and the console, runs your checkpoint hook and stops the machine with its disk intact. SaladCloud's $0.25 tier reschedules the container on another PC, and Clore.ai's own snapshot shows the median spot price close to the on-demand median, so check the discount before counting on it.

Within a few minutes on a managed cloud with ready templates. GPUBank reports a median of 48 seconds from launch to a machine that passed its health check, and about two and a half minutes for vLLM 0.11 or four minutes for ComfyUI with its models synced, none of it billed before the health check. On a marketplace the start time depends on the host's disk, its uplink and whether your image is already cached there.