RTX 5090 cloud pricing: what 9 providers charge per hour, checked 7 November 2026
On 7 November 2026, renting one RTX 5090 cost between $0.25 and $1.19 per GPU-hour on the providers' own pages: $0.25 is Salad's interruptible Lowest tier and $1.19 is RunPod's Secure Cloud. Most self-serve rates that will not preempt you mid-job sat between $0.60 and $0.90, and Vast.ai's 108 single-GPU marketplace offers had a median of $0.85. Billing is per second at most of them; storage and bandwidth are where the bills differ.
- Lowest listed: Salad $0.25 (Lowest priority, preemptible); Vast.ai single-GPU offers from $0.44.
- Typical non-interruptible rate: CloudRift $0.65, RunPod Community $0.69, Hivenet EUR 0.75, Spheron $0.84, GPUwerk $0.89.
- Highest: RunPod Secure Cloud $1.19.
- Monthly: LeaderGPU from EUR 499 excluding VAT; CloudRift $0.55/h on a 3-month reservation.
- Checked but no RTX 5090 rate published: Hyperstack, TensorDock.
How much does it cost to rent an RTX 5090 per hour?
Expect $0.60 to $0.90 an hour for a 32 GB card you keep until you stop it, and less if you accept preemption. I read every row below from the provider's own pricing page or live listing on 7 November 2026. GPUwerk, which publishes this article, is in the table too and is marked.
| Provider | Price per GPU-hour | Billing | Storage and egress | Location stated | Source | Checked |
|---|---|---|---|---|---|---|
| Salad | $0.25 Lowest, $0.367 Low, $0.483 Medium, $0.60 High priority | Per second while running | vCPU and RAM included; storage and egress not stated | Not stated (distributed nodes) | salad.com/pricing | 2026-11-07 |
| Vast.ai (marketplace) | Single-GPU offers from $0.44; median $0.85; 10th to 90th percentile $0.49 to $1.33 | Per second | Storage billed while stopped; bandwidth per TB, set by each host | Per host; 91 of 288 offers in EU countries | Vast search API | 2026-11-07 |
| Lium (marketplace) | $0.50 to $0.94 live asks; one GPU from $0.66 | Per second, no minimum | No fees listed | Each pod names its city and country | lium.io | 2026-11-07 |
| CloudRift | $0.65 on demand; $0.62 (1 week), $0.59 (1 month), $0.55 (3 months) reserved | Per second while running; paused VMs keep billing | 800 GB local storage included; no egress fees | Not stated | cloudrift.ai/pricing | 2026-11-07 |
| RunPod Community Cloud | $0.69 | Per second | Container disk $0.10/GB-month; volume $0.10 running, $0.20 stopped; no egress fees | 17 regions, not named | runpod.io | 2026-11-07 |
| LeaderGPU | From EUR 499 a month, excl. VAT (about EUR 0.68/h over 730 h) | Daily, weekly or monthly; EUR 10 minimum order | Not stated | Netherlands | leadergpu.com | 2026-11-07 |
| Hivenet | From EUR 0.75 | Per second | Not stated; VAT not stated | France, UAE and USA deployment paths | hivenet.com | 2026-11-07 |
| Spheron | $0.84 dedicated | Per minute, 20-minute minimum | 200 GB NVMe in the spec; fees not stated | US, Europe, Canada | spheron.network | 2026-11-07 (page updated 2026-10-21) |
| GPUwerk (publisher) | $0.89, USD excl. tax | Per minute from prepaid credit | No egress fees | Hradec Králové, Czech Republic (EU); one machine | gpuwerk.com | 2026-11-07 |
| RunPod Secure Cloud | $1.19 | Per second | As Community Cloud | 14 regions, not named | runpod.io/pricing | 2026-11-07 |
Prices in USD unless the provider shows EUR. Vast.ai figures are my count of on-demand offers returned by its public search API on 7 November 2026, using each offer's base GPU price without storage. LeaderGPU's hourly equivalent is my division of EUR 499 by 730 hours. GPUwerk is the publisher of this page; its row comes from its own pricing page. Hyperstack's pricing page lists no RTX 5090 (its cheapest H100 is $2.50). TensorDock's public site lists no RTX 5090 rate and its deploy dashboard needs a login, so neither appears in the table.
Why does the same card cost $0.25 at one provider and $1.19 at another?
Mostly because you are buying different terms of service, not a different GPU. Salad's page spells out the trade: Lowest priority is the most interruptible, Medium and Low can be preempted by higher-priority groups after roughly 30 minutes, and High is never preempted by other workloads but can still lose a node that disconnects. That preemption is what the $0.25 floor buys.
RunPod sells the same 32 GB card in two tiers drawn from different host pools: Community Cloud at $0.69 across 17 regions and Secure Cloud at $1.19 across 14, and the main table on its pricing page shows only the Secure figure. The 72% premium buys a different class of host, not more VRAM. The RunPod comparison covers what each tier means for data handling. RunPod's docs add that pods are billed per second for compute and storage, with no fees for data ingress or egress, and that you are not charged while the host machine is unavailable.
Marketplaces sit in between and spread widely. Lium showed $0.50 to $0.94 across 46 pods. CloudRift lists one on-demand price, $0.65, with 800 GB of local storage bundled into it.
What does the Vast.ai marketplace actually charge?
A median of $0.85 for one RTX 5090, with half the offers between $0.59 and $1.01 once you include multi-GPU machines. Vast's static RTX 5090 pricing page showed "No current offers" when I loaded it, so I queried the same public search endpoint the console uses. It returned 288 rentable on-demand offers covering 897 GPUs. By my count:
- All offers, per GPU: minimum $0.40, median $0.72, a $26.67 outlier at the top.
- Single-GPU offers (108): minimum $0.44, median $0.85, 90th percentile $1.33.
- Verified hosts (175 offers): minimum $0.47, median $0.87.
- Hosts in EU countries (91 offers): minimum $0.44, median $0.87. The four cheapest were in Slovakia; Czechia (25 offers) and Hungary (17) had the most.
The base rate is not the whole bill. Vast's billing docs say storage is charged for every second an instance exists, including while stopped, and bandwidth is charged for every byte sent or received in any state, at a per-TB price that varies by machine. Across the RTX 5090 offers the median upload price was about $4 per TB, and fewer than 4% of hosts charged nothing for bandwidth. For trade-offs beyond price, the Vast.ai comparison covers who runs the hardware.
Which RTX 5090 clouds are in the EU?
Four providers in the table name an EU location, and Vast lets you filter by host country. Hivenet lists a France deployment path at EUR 0.75 an hour. LeaderGPU says its servers sit in EU data rooms in the Netherlands, billed by the day, week or month from EUR 499 a month excluding VAT. Spheron's RTX 5090 page names US, Europe and Canada regions for its $0.84 rate. GPUwerk's machine is in the Czech Republic.
A country on a marketplace listing tells you where the box is, not who operates it. On Vast and Lium the host is an independent party, which matters for a data processing agreement. If the location question is about GDPR, the GDPR and AI Act page explains what a processor needs to state. RunPod's RTX 5090 page lists "Varies by region" without naming them.
What do storage, minimums and egress add to a real job?
For short jobs, a few dollars at most; for a model you park between sessions, storage can outlast the compute. A worked example, my arithmetic from the published rates: a 10-hour fine-tuning job on one RTX 5090 with a 200 GB volume, then the volume kept for 7 days while stopped, then 500 GB downloaded.
| Provider | 10 h compute | 200 GB volume, 10 h running | 200 GB kept 7 days stopped | 500 GB egress | Total |
|---|---|---|---|---|---|
| RunPod Community | $6.90 | $0.27 | $9.21 | $0 | $16.38 |
| Vast.ai median single GPU | $8.50 | Host-set | Host-set | About $2 | $10.50 plus storage |
| CloudRift | $6.50 | Included (800 GB) | Pausing keeps billing the GPU | $0 | $6.50, if terminated |
| GPUwerk (publisher) | $8.90 | Host disk | Stop holds the machine at 75% of the rate | $0 | $8.90, if terminated |
RunPod: $0.69 x 10; volume $0.10/GB-month running and $0.20 stopped, prorated over 730 hours (200 x 0.20 x 168/730 = $9.21). Vast: median single-GPU base price and the median $4/TB upload price; storage is set per host. CloudRift and GPUwerk bill a paused or stopped GPU, so copy your data off and terminate. My calculation, 7 November 2026 rates.
The lesson from the RunPod row: keeping a 200 GB stopped volume for a week cost more than the ten hours of GPU time. Spheron's 20-minute minimum only matters if you start and stop many short sessions.
What does a full month cost, and when does a reservation win?
Running 730 hours costs $475 to $650 at the common rates, and reservations cut that by about 10 to 15%. At $0.65 on demand CloudRift is $474.50 a month; its 3-month rate of $0.55 brings that to $401.50. RunPod Community is $503.70, Spheron $613.20, GPUwerk $649.70 and RunPod Secure $868.70. LeaderGPU's EUR 499 is a monthly price outright.
A month of continuous rental is also where buying starts to look reasonable: it is a tenth to a seventh of the $4,400 to $4,900 street price recorded on the DGX Spark vs RTX 5090 page for August 2026. The RTX 5090 price article works through the buy-versus-rent break-even with electricity included, so I will not repeat it here. If your load is a few hours a day, per-second or per-minute billing means you pay only for those hours.
Is a rented RTX 5090 the right GPU for the job?
It is the cheapest memory bandwidth you can rent per dollar, as long as the job fits in 32 GB. At $0.69 to $0.89 an hour you get 1,792 GB/s of memory bandwidth (the figure in Tom's Hardware's RTX 5090 review); an H100 SXM starts at $2.69 an hour on RunPod's Community Cloud for 3.35 TB/s. By that arithmetic the 5090 is the best tokens-per-dollar choice for 8B to 32B models at 4-bit, and a good fit for image and video diffusion and small LoRA runs. The H100, H200 and B200 price comparison has the datacenter side.
It is the wrong card once weights plus KV cache pass about 30 GB. A 70B model at 4-bit is about 40 GB and does not load. Your options then are an RTX PRO 6000 with 96 GB (RunPod lists $1.69 Community and $2.49 Secure) or a 128 GB DGX Spark, which is slower per token but holds models a 5090 cannot; the Spark comparison linked above lays out where each wins, and the best LLMs for an RTX 5090 lists what fits.
My pick for one person testing models: RunPod Community or CloudRift if the region does not matter to you, Hivenet or a verified EU host on Vast if the data must stay in the EU. I would avoid the interruptible tiers for anything longer than a few minutes unless the job checkpoints itself. GPUwerk's own card is the right choice only if you want a whole EU machine, billed per minute, and it happens to be free; it is one machine, and the pricing page shows whether it is.
FAQ
How much does it cost to rent an RTX 5090 per hour?
On 7 November 2026 the providers' own pages listed $0.25 (Salad's lowest, interruptible priority) to $1.19 (RunPod Secure Cloud) per GPU-hour. Most non-interruptible self-serve rates sat between $0.60 and $0.90: CloudRift $0.65, RunPod Community $0.69, Hivenet EUR 0.75, Spheron $0.84 and GPUwerk $0.89. Vast.ai's single-GPU marketplace offers had a median of $0.85.
What is the cheapest RTX 5090 cloud?
Salad's Lowest priority tier at $0.25 per GPU-hour, but those nodes can be preempted. For capacity that is not taken away mid-job, the lowest listed rates on 7 November 2026 were Vast.ai marketplace offers from $0.44 for one GPU, Lium from $0.50, Salad High priority at $0.60 and CloudRift at $0.65 on demand.
Where can I rent an RTX 5090 in the EU?
Hivenet lists a France deployment path at EUR 0.75 per hour, LeaderGPU rents RTX 5090 servers in the Netherlands from EUR 499 a month excluding VAT, GPUwerk has one machine in the Czech Republic at $0.89 per hour, and 91 of Vast.ai's 288 RTX 5090 offers on 7 November 2026 were hosted in EU countries. RunPod does not name regions on its RTX 5090 page.
How is RTX 5090 cloud time billed?
Per second at RunPod, Vast.ai, Salad, CloudRift, Lium and Hivenet, per minute at Spheron (with a 20-minute minimum) and GPUwerk, and by the day, week or month at LeaderGPU. Storage is usually extra: RunPod charges $0.20 per GB a month for a stopped volume, and Vast.ai keeps billing storage and bandwidth while an instance is stopped.
Is it cheaper to rent an RTX 5090 or buy one?
At $0.65 to $0.89 an hour, a 730-hour month of continuous rental costs about $475 to $650, which is why heavy daily use pushes people toward buying. Light or bursty use favors renting, because an idle rented GPU costs nothing once you terminate it, while a bought card costs the same whether it runs or not.
Sources
Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.
- Pricing (Pods, Serverless, Storage), RunPod
- RTX 5090 GPU Rental, RunPod
- Pod pricing, RunPod docs
- RTX 5090 GPU pricing, Vast.ai
- Offer search API (public bundles endpoint), Vast.ai
- Billing, Vast.ai docs
- SaladCloud pricing, Salad
- GPU rental prices per hour, CloudRift
- NVIDIA RTX 5090 GPU compute, CloudRift
- NVIDIA RTX 5090 rental, Spheron, 2026-10-21
- Compute: GPU and CPU rental, Hivenet
- GPU servers for rent, LeaderGPU
- RTX 5090 pricing, Lium
- GPU pricing, Hyperstack
- TensorDock GPU cloud, TensorDock
- Nvidia GeForce RTX 5090 Founders Edition review, Tom's Hardware