Buying an RTX 5090 in November 2026 means paying $5,600 to $11,900 for a card NVIDIA launched at $1,999 — and that is if you can find one in stock at all. Renting one is a different story: you can rent an RTX 5090 GPU server for as little as $0.25 an hour, spin it up in minutes, run your model, and kill it when you are done. No $6,000 purchase, no 575-watt card heating your room, no wondering whether the card will still be worth anything in two years.
But the headline price is not the whole story. The same card costs $0.25 an hour at one provider and $1.19 at another, and the gap is not a scam — it is different terms of service, different host pools, and very different storage and egress bills hiding underneath. One provider’s $0.69-an-hour pod cost a real user $16.38 for a 10-hour job, and more than half of that was not GPU time at all.
This guide gives you the real November 2026 numbers: what 9 providers actually charge, what the hidden costs add, which AI models fit in 32 GB of VRAM, and the break-even math that tells you whether to rent or buy. Every price here comes from the providers’ own pages, checked in November 2026.
- How Much Does It Cost to Rent an RTX 5090?
- Why the Same Card Costs $0.25 at One Provider and $1.19 at Another
- The Real Bill: Storage, Egress, and Stopped Instances
- What a Full Month Costs — and When a Reservation Wins
- Rent vs Buy: The Break-Even Math
- What Fits in 32GB of VRAM?
- When an RTX 5090 Is the Wrong GPU
- How to Rent an RTX 5090 in 10 Minutes
- Marketplace vs Fixed-Price Cloud: Which Should You Pick?
- The Demand Squeeze: Why GPU Rental Prices Are Moving
- EU Options and the GDPR Question
- FAQ
- The Bottom Line
How Much Does It Cost to Rent an RTX 5090?
Short answer: $0.60 to $0.90 an hour for a card nobody can take away from you mid-job, and $0.25 to $0.44 if you accept preemption. Here is the full table, from prices read off the providers’ own pages on 7 November 2026, cross-checked against a second independent price review on 13 November 2026.
| Provider | Price per GPU-hour | Billing | What you should know |
|---|---|---|---|
| Salad | $0.25 Lowest / $0.367 Low / $0.483 Medium / $0.60 High | Per second | Interruptible tiers; High is never preempted by other workloads |
| Vast.ai (marketplace) | From $0.44; median $0.85 (single-GPU offers) | Per second | 288 offers; storage billed even while stopped; bandwidth per TB |
| Lium (marketplace) | $0.50–$0.94; one GPU from $0.66 | Per second, no minimum | 46 pods; each names its city and country |
| CloudRift | $0.65 on demand; $0.62 (1 week), $0.59 (1 month), $0.55 (3 months) reserved | Per second | 800 GB local storage included; no egress fees |
| RunPod Community Cloud | $0.69 | Per second | 17 regions; container disk $0.10/GB-month; no egress fees |
| Hivenet | From €0.75 | Per second | France, UAE and USA deployment paths |
| Spheron | $0.84 dedicated | Per minute, 20-minute minimum | 200 GB NVMe included; US, Europe, Canada |
| GPUwerk | $0.89 | Per minute from prepaid credit | One machine in the Czech Republic; no egress fees |
| RunPod Secure Cloud | $1.19 | Per second | 14 regions; different host pool from Community Cloud |
| LeaderGPU | From €499/month excl. VAT (≈ €0.68/h) | Daily, weekly or monthly | Netherlands; €10 minimum order |
Two things jump out. First, the spread between the cheapest non-interruptible rate (CloudRift at $0.65) and the most expensive (RunPod Secure at $1.19) is 83% — for the exact same GPU. Second, a second price review a week later (13 November 2026) found nearly identical numbers — Vast.ai median $0.79, RunPod $1.39 on its headline page — which means these rates are stable, not a one-day glitch. If you want the raw provider-by-provider breakdown with sources, the GPUwerk RTX 5090 cloud pricing comparison lists every figure with its source page.
For context, the card itself: 32 GB of GDDR7 memory, 1,792 GB/s of memory bandwidth, and a 575 W power draw. That bandwidth-per-dollar is the whole reason the 5090 is the card people rent — it is the cheapest memory bandwidth you can buy by the hour, as long as your job fits in 32 GB.
Why the Same Card Costs $0.25 at One Provider and $1.19 at Another
You are not buying a different GPU. You are buying different terms. Once you understand the three things that move the price, the table above stops looking random.
Preemption. Salad spells the trade out plainly: its Lowest tier at $0.25 can be interrupted at any time, Low and Medium can be preempted by higher-priority groups after roughly 30 minutes, and only High at $0.60 is never preempted by other workloads (though a node can still disconnect). That $0.25 floor is the price of a job that might die halfway through. It is fine for a render you can restart or a training run that checkpoints every few minutes. It is a disaster for a 6-hour fine-tune with no checkpoints.
Host pool. 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. The 72% premium buys a different class of host, not more VRAM — and RunPod’s own pricing page shows only the Secure figure, so a lot of people quote $1.19 as “the RunPod price” without ever seeing the $0.69 tier. Vast.ai is the same story in marketplace form: its 288 offers ranged from $0.40 to $1.33 per GPU-hour, with verified hosts at a $0.87 median and the four cheapest offers sitting in Slovakia.
What is bundled. CloudRift’s $0.65 includes 800 GB of local storage and no egress fees. RunPod Community’s $0.69 looks cheaper than it is once you add a volume: container disk at $0.10 per GB-month, and a stopped volume at $0.20 per GB-month. Spheron bundles 200 GB of NVMe into its $0.84. The headline rate is only comparable when you know what is underneath it — which brings us to the part that actually decides your bill.
The Real Bill: Storage, Egress, and Stopped Instances

For a short job, the extras are a few dollars. For a model you park between sessions, storage can outlast the compute. Here is a worked example, using the published November 2026 rates: a 10-hour fine-tuning job on one RTX 5090 with a 200 GB volume, 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 + storage |
| CloudRift | $6.50 | Included (800 GB) | Paused VMs keep billing the GPU | $0 | $6.50, if terminated |
| GPUwerk | $8.90 | Host disk | Stop holds the machine at 75% of the rate | $0 | $8.90, if terminated |
Look at the RunPod row: keeping a 200 GB stopped volume for a week cost $9.21 — more than the ten hours of GPU time that did the actual work. That is the single most common billing surprise in GPU rental, and it is entirely avoidable: copy your weights and checkpoints off, then terminate the instance instead of stopping it. CloudRift and GPUwerk bill a paused or stopped GPU, so the same rule applies there — terminate, do not pause.
Vast.ai has its own version of the trap: its 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 its RTX 5090 offers the median upload price was about $4 per TB, and fewer than 4% of hosts charged nothing for bandwidth. On a marketplace, the cheapest hourly rate is not the cheapest job — check the host’s storage and bandwidth lines before you deploy.
Spheron’s 20-minute minimum only matters if you start and stop many short sessions. If you are the kind of person who spins a pod up, realizes the model name was wrong, and kills it 3 minutes later, those minimums add up across a day of fiddling.
What a Full Month Costs — and When a Reservation Wins
Running one RTX 5090 for all 730 hours of a month costs $475 to $650 at the common rates, and reservations cut that by about 10 to 15%:
| Provider / plan | Hourly rate | 730-hour month |
|---|---|---|
| CloudRift on demand | $0.65 | $474.50 |
| CloudRift 3-month reservation | $0.55 | $401.50 |
| RunPod Community | $0.69 | $503.70 |
| Spheron | $0.84 | $613.20 |
| GPUwerk | $0.89 | $649.70 |
| RunPod Secure | $1.19 | $868.70 |
| LeaderGPU (monthly) | ≈ €0.68 | €499 excl. VAT |
A month of continuous rental is also where buying starts to look reasonable: $475 to $650 a month is a tenth to a seventh of the $5,600 street price of the card. If your load is a few hours a day, per-second billing means you pay only for those hours, and renting wins by a mile. If the card would run 24/7 for a year, do the break-even math below before you commit to either side.
Rent vs Buy: The Break-Even Math
The question everyone actually asks is simple: at what point does renting cost more than just buying the card? The honest answer depends on your electricity price, because a bought card still costs money every hour it runs.
Work it through with the November 2026 numbers. A new RTX 5090 in the US cost $5,599.99 at the cheapest in-stock listing (an ASUS ROG Astral OC). Renting one at $0.89 an hour, the card draws about 0.65 kW at full load (575 W card plus roughly 75 W for the host). At German household electricity — €0.3618 per kWh in the first half of 2026 — that is about $0.26 of electricity per hour. So every hour you own instead of rent saves you $0.89 − $0.26 = $0.63, and the break-even is $5,599.99 ÷ $0.63 ≈ 8,930 hours.
| Usage pattern | Hours to break even | Calendar time |
|---|---|---|
| 24/7, never idle | ~8,930 hours | About 12 months |
| 40 hours a week | ~8,930 hours | About 4.3 years |
| 10 hours a week (evenings + weekends) | ~8,930 hours | About 17 years |
That table is the whole decision. A team running a model server around the clock reaches break-even in about a year, and owning wins — assuming the card does not die and you do not want a newer one by then. A hobbyist running 10 hours a week never reaches it; renting is cheaper for roughly 17 years, and by then the card is a museum piece. The middle case — 40 hours a week — takes 4.3 years, which is longer than most people keep a GPU, so renting still wins for most individuals.
There is also the buy side nobody quotes: the street price in November 2026 was $6,199 to $11,896 depending on the model and the country, and board partners reportedly stopped receiving new GB202 chips in October, with supply moving to RTX PRO cards. If you do buy, you are buying into a shrinking supply at 3 to 6 times MSRP. Renting sidesteps that entirely — and if you want a dedicated box rather than an hourly pod, our Hetzner GEX430 guide covers the other end of the market: four RTX PRO 6000 cards at €3,599 a month.
What Fits in 32GB of VRAM?
The 5090’s 32 GB is generous for a consumer card and tight for a datacenter one. What fits depends on quantization, and the rule of thumb is simple: a model at 4-bit precision needs roughly half a gigabyte per billion parameters, plus room for the KV cache while it generates.
| Model size | 4-bit weights | Fits on a 5090? | Good for |
|---|---|---|---|
| 7B–8B | ~4–5 GB | Yes, easily | Chatbots, agents, coding assistants with big context |
| 14B | ~8 GB | Yes | Stronger reasoning, longer conversations |
| 32B | ~19 GB | Yes, with room for KV cache | Near-frontier open-weight quality at 4-bit |
| 70B | ~40 GB | No | Needs a 96 GB RTX PRO 6000 or two cards |
| Image diffusion (SDXL, Flux) | ~12–24 GB | Yes | Fast image generation; Flux at full precision is tight |
| Video diffusion (short clips) | ~20–28 GB | Yes, short clips | 5–10 second generations; long clips need more VRAM |
In practice, 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 fine-tuning runs. 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 simply does not load. If that is your workload, read the next section before you rent anything.
One more thing the VRAM table does not show: the 5090’s 1,792 GB/s of bandwidth is what makes it fast for inference, not the raw compute. An H100 SXM starts at $2.69 an hour on RunPod Community for 3.35 TB/s — nearly twice the bandwidth, but roughly four times the hourly price. For a single-user chatbot or a small API, the 5090 gives you most of the speed at a quarter of the cost. If you are still deciding what to run on it, our guide to running AI models on a VPS covers the setup side.
When an RTX 5090 Is the Wrong GPU
Renting the wrong card is the most expensive mistake in this whole guide, because you discover it after you have already loaded your data. Check these three cases first:
Your model does not fit. A 70B model at 4-bit needs about 40 GB. Your options are an RTX PRO 6000 with 96 GB (RunPod lists $1.69 Community and $2.49 Secure per hour) or a 128 GB DGX Spark, which rents for about $0.79 an hour — slower per token than a 5090, but it holds models a 5090 cannot load at all.
You need datacenter reliability. A rented consumer card on a marketplace is a stranger’s PC. If your workload is a production API with an SLA, you want a datacenter card (H100, H200, B200) on a fixed-price cloud with support, not the cheapest Vast.ai host in the list. The hourly rate is higher; the 3 a.m. page is not yours to answer.
You need many GPUs at once. One 5090 is cheap. Eight of them on one machine is a different product — and in November 2026, an 8× RTX 5090 server in the Chengdu-Chongqing region was renting for about 17,000 yuan a month on a one-year contract, roughly double the equivalent 8× RTX 4090 rate. Multi-GPU rental is a negotiated, contracted market, not a self-serve one, and the per-GPU price is worse than single-card rental. If you need that scale, a dedicated multi-GPU box like the GEX430 is usually the better comparison.
How to Rent an RTX 5090 in 10 Minutes

The flow is nearly identical on RunPod, Vast.ai, and CloudRift, so here it is once, with the checks that catch the expensive mistakes.
Step 1 — Pick the provider by your job, not by the headline rate. Short test under an hour: any per-second provider. A fine-tune that must not die: CloudRift, RunPod Community, or Salad High — never an interruptible tier. Data that must stay in the EU: Hivenet, LeaderGPU, or a verified EU host on Vast.ai. A production API: RunPod Secure or a datacenter card instead.
Step 2 — Create the pod and check the GPU before you install anything. Once the pod is up, open its terminal and confirm the card is what you paid for:
nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv
# Expect: NVIDIA GeForce RTX 5090, 32768 MiB, driver 580.xx or newer
If it shows a different card, or 24 GB instead of 32 GB, kill the pod immediately and open a support ticket — marketplace listings do get mislabeled, and you do not want to find out after an hour of setup. Also check the CUDA version matches what your framework needs; a pod template with CUDA 12.8 will not run a container built for CUDA 13, and vice versa.
Step 3 — Pull your model and start the server. The fastest path for most people is vLLM in a container, which handles the model download, quantization, and the OpenAI-compatible API in one command:
docker run --gpus all -p 8000:8000 \
-e HF_TOKEN=your_huggingface_token \
vllm/vllm-openai:latest \
--model Qwen/Qwen3-32B-AWQ \
--max-model-len 32768 --gpu-memory-utilization 0.92
That serves a 32B model at 4-bit on one 5090 with a long context window. If you would rather run a local model with a chat interface, Ollama is the simpler route — ollama run qwen3:32b after pulling the image — and our AI models on a VPS guide walks through both in more detail.
Step 4 — Test it, then set a reminder to terminate. Hit the API once and check the speed:
curl http://localhost:8000/v1/models
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{"model": "Qwen/Qwen3-32B-AWQ", "prompt": "Hello, who are you?", "max_tokens": 50}'
Then decide: are you done, or is this a long-running server? If you are done, copy your outputs off and terminate the pod now — do not stop it, do not leave it running overnight “just in case”. A stopped volume on RunPod costs $0.20 per GB-month, and a paused GPU on CloudRift keeps billing. The cheapest GPU hour is the one you never buy.
Marketplace vs Fixed-Price Cloud: Which Should You Pick?
The market splits into two kinds of provider, and they suit different people.
Marketplaces (Vast.ai, Lium, Salad) are the cheapest per hour and the most variable. You are renting a stranger’s hardware, prices move with supply, and the host sets the storage and bandwidth terms. Vast.ai’s own numbers show the spread: a median of $0.85 for single-GPU offers, but a 10th-to-90th percentile range of $0.49 to $1.33. Pick a verified host, read the storage and bandwidth lines, and never put unencrypted sensitive data on a marketplace pod.
Fixed-price clouds (RunPod, CloudRift, Spheron, GPUwerk, Hivenet, LeaderGPU) charge one published rate, run their own support, and publish their storage and egress terms. You pay 10 to 40% more per hour than the cheapest marketplace offer, and in exchange you get a pod that behaves the same way every time. For anyone running a business workload, that premium is cheap insurance.
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.ai if the data must stay in the EU. Avoid the interruptible tiers for anything longer than a few minutes unless the job checkpoints itself. And if you are building something that needs to stay up — an agent, an API, a bot — remember that a rented GPU is a compute box, not a home: our OpenClaw on a VPS guide and best VPS for AI agents comparison cover the always-on side of the same problem.
The Demand Squeeze: Why GPU Rental Prices Are Moving
Two things are pushing on this market right now, and both showed up in the news in the last week.
First, the RAM shortage that has already raised VPS prices across the industry is hitting GPUs too. Memory and storage costs are several times what they were in 2025, because chip makers moved production toward AI servers — the same shortage behind Hetzner’s three price waves, OVHcloud’s unbundling, and netcup’s two rounds of increases this year. We covered the full picture in why VPS prices are rising in 2026; GPU rental is the same story, one layer up the stack.
Second, demand itself is being pushed around by policy. On 12 November 2026, Reuters reported that Chinese AI companies are trying to secure NVIDIA-powered compute at data centers outside China, because Washington is considering expanding export controls to cover remote access to advanced chips. The proposed Remote Access Security Act passed the US House in January 2026 by 369 votes to 22 and is still awaiting Senate action — so nothing is law yet, but companies are already reserving overseas capacity before any rule changes. More buyers chasing the same overseas GPUs is upward pressure on rental prices, and it is one more reason to lock a reservation rate if you plan to rent for months.
None of this means prices will spike tomorrow. It means the $0.65-an-hour rate you see today is not guaranteed next quarter, and a 3-month reservation at $0.55 is worth more than it looks on paper.
EU Options and the GDPR Question
If your data has to stay in the EU, four providers in the table name an EU location, and Vast.ai lets you filter by host country. Hivenet lists a France deployment path at €0.75 an hour. LeaderGPU rents RTX 5090 servers in the Netherlands from €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. On Vast.ai, 91 of the 288 RTX 5090 offers on 7 November 2026 were hosted in EU countries, with the most in Czechia (25) and Hungary (17).
One caution: a country on a marketplace listing tells you where the box is, not who operates it. On Vast.ai and Lium the host is an independent party, which matters for a data processing agreement. If the location question is about GDPR, you need a provider that will sign as a processor — a fixed-price EU cloud, not a marketplace host you cannot identify. RunPod’s RTX 5090 page lists its regions only as “varies by region”, so if EU residency is a hard requirement, confirm it with support before you deploy.
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 €0.75, Spheron $0.84 and GPUwerk $0.89. Vast.ai’s single-GPU marketplace offers had a median of $0.85. A second check on 13 November 2026 found nearly identical numbers.
What is the cheapest RTX 5090 cloud?
Salad’s Lowest priority tier at $0.25 per GPU-hour, but those nodes can be preempted at any time. 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 €0.75 an hour, LeaderGPU rents RTX 5090 servers in the Netherlands from €499 a month excluding VAT, GPUwerk has one machine in the Czech Republic at $0.89 an 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. The break-even is about 8,930 hours of use — roughly 12 months at 24/7, or 4.3 years at 40 hours a week.
Can an RTX 5090 run a 70B model?
No. A 70B model at 4-bit precision needs about 40 GB of VRAM, and the 5090 has 32 GB. You need an RTX PRO 6000 with 96 GB (about $1.69 an hour on RunPod Community) or a 128 GB DGX Spark (about $0.79 an hour rented), or two 5090s with tensor parallelism.
The Bottom Line
To rent an RTX 5090 GPU server in November 2026, budget $0.65 to $0.89 an hour for a card that will not be taken away from you — CloudRift at $0.65 or RunPod Community at $0.69 are the honest starting points, with Vast.ai from $0.44 if you are comfortable picking a host on a marketplace. The headline rate is not your bill: a stopped volume can cost more than the GPU time, a paused instance can keep billing, and bandwidth on a marketplace is charged per terabyte. Copy your outputs off and terminate the pod when you are done.
Buy only if the card would run 24/7 for a year or more; at $5,600 street price and $0.26 an hour of electricity, the break-even is about 8,930 hours. For everyone else — a weekend of fine-tuning, a week of testing a 32B model, a month of image generation — renting wins, and it wins by a lot. Pick the provider by the job, check the card with nvidia-smi before you install anything, and never rent an interruptible tier for a job that cannot checkpoint.
Prices checked 16 November 2026 from the providers’ own pages and the GPUwerk and Thundercompute price reviews. GPU rental rates move with supply — re-check the provider’s page before you commit to a reservation.


