Which dedicated GPU server should you rent in 2026?
Five providers compared on the numbers that decide the bill: VRAM, monthly price, setup fee and how many hours your GPU really runs.
Short answer: Hetzner is the cheapest dedicated GPU server per GB of VRAM we found, with a 24 GB RTX PRO 4000 at €214 a month before tax, hosted in Helsinki, Finland. If your GPU sits idle half the day, skip the monthly server: RunPod rents an RTX 4090 at $0.69 an hour and an H100 PCIe at $2.89. Lambda is the cleaner pick for multi-day training runs on H100 or B200 hardware it owns. OVHcloud and Ikoula make sense when your data has to stay in France.
Pick by workload. Serving an 8B to 14B open-source model around the clock: 24 GB, Hetzner GEX45. Serving a 70B model: 96 GB, Hetzner GEX63. Fine-tuning for a few days: Lambda or RunPod by the hour. Not sure yet: rent hourly for a month, log your usage, then decide.
The ranking is editorial: a sponsored placement is always labeled on its listing, and no affiliate link changes a score. Some links on this page are affiliate links.
Who really rents you a dedicated GPU, and at what price?
Most "dedicated GPU server" guides stop at a definition. Here is the split that matters for your budget. A dedicated GPU server is a whole physical machine, GPU included, rented by the month: nobody else touches it and you get root access. Hetzner, OVHcloud and Ikoula sell this. An hourly GPU instance rents you a card for the time you use it, on hardware you hand back: that is RunPod and Lambda. Monthly wins when the card runs most of the day. Hourly wins when it doesn't.
The first spec to read is VRAM, the memory on the GPU. It decides whether your model fits at all. So we added a cost per GB of VRAM, which no vendor page shows. Scaleway and Contabo also rent GPUs, but their pricing pages did not show us their numbers, so they are not ranked here rather than quoted on a guess.
| What you rent: GPU, VRAM, location | Entry price, setup fee, cost per GB of VRAM | |
|---|---|---|
| HetznerDedicated, monthly, Finland (GEX45, GEX63) or Germany | RTX PRO 4000 Blackwell, 24 GB (GEX45); RTX PRO 6000 Blackwell, 96 GB (GEX63, GEX131); four RTX PRO 6000 for 384 GB (GEX430). Location: GEX45 and GEX63 in Helsinki only, GEX131 also in Germany. Billing: hourly, capped at the monthly price. | GEX45 at €214/mo + €209 setup; GEX63 at €999/mo + €999. Prices shown in euros, before tax. Per GB of VRAM: about €8.90/mo on the GEX45. |
| RunPodHourly, per-second billing | From the 24 GB RTX A5000 to the 180 GB B200, with RTX 4090, A100 and 80 GB H100 in between. | RTX A5000 at $0.27/hr, RTX 4090 at $0.69/hr, H100 PCIe at $2.89/hr, B200 at $5.89/hr. |
| LambdaHourly, its own data center hardware | Single-GPU instances from the 24 GB Quadro RTX 6000 to the 180 GB B200, including A100, H100 and GH200. | A10 (24 GB) at $1.29/hr, GH200 (96 GB) at $2.29/hr, H100 PCIe at $3.29/hr, B200 at $6.99/hr, plus sales tax. |
| OVHcloudDedicated, monthly, listed as "Soon available" | Scale-GPU-1: NVIDIA L4 (24 GB per NVIDIA's spec sheet), 16-core AMD EPYC Genoa 9124, 192 GB to 1.125 TB of RAM, 5 to 25 Gbps public bandwidth. | From $1,145/mo + $1,145 setup. Per GB of VRAM: about $48/mo, the highest here. |
| IkoulaDedicated, monthly, hosted in France | RTX 4070 (12 GB), RTX 5070 Ti (16 GB) or RTX Pro 4000 Blackwell (24 GB), single or dual. 2 x 2 TB NVMe on every plan. 1 Gbps shared, 100 Mbps guaranteed. | From €195/mo (RTX 4070); €260/mo for 24 GB. Prices shown in euros, before tax. Setup fee not listed on the pricing page. |
Prices seen on vendor pricing pages in 2026, before tax. Hetzner and Ikoula list euros, which we don't convert; OVHcloud, RunPod and Lambda list US dollars. Cost per GB of VRAM is our math: monthly price divided by GPU memory.
Hetzner
Hetzner is the German host that keeps undercutting the dedicated server market, and its GPU line follows the same playbook. Four machines in 2026. The GEX45 pairs a 24 GB RTX PRO 4000 Blackwell (GDDR7 with ECC) with a Core i5-13500 and 64 GB of RAM for €214 a month plus €209 setup. The GEX63 jumps to a 96 GB RTX PRO 6000 Blackwell for €999. The GEX131 keeps that card with a Xeon Gold and at least 256 GB of RAM, from €1,499. The GEX430 runs four cards, 384 GB of VRAM, for €3,599.
Here is what most comparisons miss: Hetzner bills by the hour and caps the bill at the monthly price. The GEX45 costs €0.3429 an hour and stops counting at roughly 624 hours. So a ten-day project costs ten days plus setup. The catch: a server is billed as long as it exists, powered on or not. Cancel it, don't just shut it down. And setup fees come back with every new order.
The trade-off is service. You get bare metal: drivers, CUDA, security and backups are your job. There are no H100 or H200 cards in the lineup, the older GEX44 and GEX130 are gone, and the GEX45 and GEX63 run in Helsinki, Finland, only; the GEX131 is also available in Germany. Hosting forums also pointed out with the previous generation that compact SFF cards are low-power parts: great for inference, weaker for heavy training.
- 24 GB of VRAM for €214/mo
- 96 GB of VRAM for €999/mo
- Hourly billing capped at the monthly price
- No minimum term
- ✓Lowest cost per GB of VRAM in this comparison
- ✓Short rentals work thanks to the hourly cap
- ✓Current Blackwell cards with ECC memory
- ✗Setup fee equals a full month on bigger machines
- ✗Bare metal: drivers and security are on you
- ✗No H100 or H200 option
Our top pick if your GPU runs most of the day. A GEX45 serves an 8B to 14B model for about €2,777 in year one, setup included. If you need an H100, look at RunPod or Lambda instead.
RunPod
RunPod isn't a dedicated server. It's a GPU marketplace billed by the second, in "pods" you start and stop at will. It belongs on this page because it's the option every team weighs against a monthly box. The catalog is wide: RTX A5000 (24 GB) at $0.27 an hour, L4 at $0.39, RTX 4090 at $0.69, A100 SXM (80 GB) at $1.49, H100 PCIe at $2.89, up to the 180 GB B200 at $5.89.
Run the math before you commit. An RTX 4090 running all 730 hours of a month costs about $504, more than twice a Hetzner GEX45 with the same VRAM. At 150 hours a month, it's about $104. The break-even sits near 300 hours a month, about ten hours a day. Above that, dedicated wins. Below it, hourly wins. And RunPod gives you H100s and B200s that none of the dedicated servers here offer.
Traps users keep reporting: pod storage keeps billing after you stop the GPU, so delete volumes you don't need. Community Cloud is cheaper but runs on third-party hosts with uneven reliability; Secure Cloud costs more and runs on dedicated infrastructure. Support is the weak spot, and it's the lowest score in our review.
- Per-second billing
- RTX A5000 through B200
- H100 PCIe at $2.89/hr
- No subscription
- ✓Zero GPU cost when you're not using it
- ✓Access to data center cards (H100, B200)
- ✓Switch GPUs in minutes
- ✗Pricier than dedicated past ~300 hours a month
- ✗Storage billed while the GPU is stopped
- ✗Support rated weak by users
The smart first move: rent hourly long enough to measure your real usage, then move to dedicated if the card runs more than ten hours a day. Full price breakdown in our RunPod pricing guide.
Lambda
Lambda also rents GPUs by the hour, with a catalog of single-GPU instances. Single-GPU instances run from the 24 GB Quadro RTX 6000 at $0.69 an hour to the 180 GB B200 at $6.99, with the A10 (24 GB) at $1.29, the A100 (40 GB) at $1.99, the H100 PCIe at $3.29 and the 96 GB GH200 at $2.29. Sales tax comes on top.
Card for card, Lambda costs more than RunPod: $3.29 vs $2.89 for an H100 PCIe, $6.99 vs $5.89 for a B200. What you pay for is a single vendor rather than a marketplace of third-party hosts, which matters when a multi-day training run can't stop halfway. The GH200 at $2.29 an hour for 96 GB is the sleeper deal on the list if you need memory more than raw speed.
The limit: no self-serve monthly rental. Reserved capacity at the lowest prices goes through sales, and 1-Click Clusters run from two weeks to a year. For a model that serves traffic 24/7, a European dedicated box is far cheaper.
- No subscription, hourly billing
- GH200 (96 GB) at $2.29/hr
- H100 PCIe at $3.29/hr
- B200 at $6.99/hr
- ✓One vendor for long training runs
- ✓Wide choice of data center GPUs
- ✓Clusters from two weeks to one year
- ✗Pricier than RunPod card for card
- ✗Sales tax on top of listed prices
- ✗No self-serve monthly server
Pick it for a few days of training or fine-tuning on an H100 or B200. For always-on inference, go back to Hetzner.
OVHcloud
OVHcloud's Scale-GPU-1 is a full server built around a GPU, not a cheap AI box. You get an NVIDIA L4, a 16-core AMD EPYC Genoa 9124, 192 GB to 1.125 TB of RAM, NVMe drives from 1.92 to 7.68 TB, a 50 Gbps private network and 5 to 25 Gbps of public bandwidth.
The price reflects that: from $1,145 a month and the same again in setup, close to $14,900 in year one. For an L4, a 24 GB card designed for inference and video, that's about $48 per GB of VRAM per month, several times Hetzner. You're paying for the server around the GPU and the network it sits on.
Two caveats. The product page still reads "Soon available", so confirm you can order before you plan around it. And user reviews of OVHcloud often mention slow delivery and canned support replies. An independent GPU cloud ranking published in late 2025 also placed OVHcloud among the laggards, with European data sovereignty as its main strength.
- NVIDIA L4 with AMD EPYC Genoa
- Up to 1.125 TB of RAM
- 50 Gbps private network
- European provider
- ✓Complete, expandable server
- ✓Much faster network than the other dedicated boxes
- ✓Fits an existing OVHcloud setup
- ✗Highest cost per GB of VRAM here
- ✗Setup fee equals a full month
- ✗Listed as "Soon available"
Only worth it if your infrastructure already lives at OVHcloud. For raw GPU per dollar, Hetzner or RunPod win easily.
Ikoula
Ikoula is a French host, and it ranks fifth on this page only because most readers of this guide don't need French hosting. If you do, it's the cheapest dedicated option we found. Six plans, all with 2 x 2 TB NVMe: the GPU 4070 (RTX 4070, 12 GB, Core i9-13900, 64 GB RAM) at €195 a month, the GPU 5070 Ti (16 GB GDDR7) at €253, and the GPU PRO 4000 with a 24 GB RTX Pro 4000 Blackwell at €260. Dual-GPU versions go up to €360 for 2 x 24 GB.
Against Hetzner, the 24 GB card costs €46 more a month. In return you get hosting in France, Linux or Windows Server, and optional Plesk. GeForce cards with 12 or 16 GB handle video and 3D well but run out of memory fast with language models. For AI, go straight to the PRO 4000.
Check the network before you order: 1 Gbps shared, with only 100 Mbps guaranteed. And the pricing page doesn't list setup fees or contract length, so get both in writing.
- Hosted in France
- 24 GB RTX Pro 4000 at €260/mo
- Dual-GPU plans up to €360/mo
- Windows Server or Linux
- ✓Cheapest French-hosted dedicated GPU we found
- ✓4 TB of NVMe storage
- ✓Windows Server for rendering work
- ✗Only 100 Mbps guaranteed bandwidth
- ✗Setup fee and term not listed
- ✗Nothing above 24 GB per card
The pick when "hosted in France" is a hard requirement and 24 GB is enough. Otherwise, Hetzner gives you more for less.
What to work out before you rent a GPU server
Vendor pages talk CUDA cores and teraflops. Projects go wrong on simpler things: memory, hours and fees you didn't plan for. Six checks. If you'd rather have someone build and host your model for you, see our LLM agency.
How much VRAM do you need?
A large language model (LLM) has to fit entirely in GPU memory to respond quickly. Rule of thumb: 24 GB runs an 8B to 14B parameter model, or a 32B model quantized to 4 bits (quantization lowers number precision to save memory). A 70B model at 4 bits needs about 96 GB. 12 and 16 GB cards are mostly for rendering, video and small image models.
Monthly or hourly?
Count your real hours. A 24 GB card costs €214 a month at Hetzner and $0.69 an hour for an RTX 4090 on RunPod. The break-even is around 300 hours a month:
- Under 300 hours: hourly wins, as long as you actually stop and delete instances.
- Over 300 hours: dedicated wins, and the gap grows every hour.
- A project of a few weeks: Hetzner's hourly billing with a monthly cap covers both.
Your real year-one cost
Add setup to the monthly price. Year one: Hetzner GEX45, €214 x 12 + €209 = €2,777; GEX63, €12,987; OVHcloud Scale-GPU-1, $14,885; Ikoula GPU PRO 4000, €3,120 before a setup fee Ikoula doesn't list. At Hetzner and OVHcloud, setup equals a month on the larger machines, so switching servers often gets expensive.
Where does your data live?
If you handle customer data, location matters. Hetzner hosts the GEX45 and GEX63 in Finland and the GEX131 also in Germany, OVHcloud and Ikoula in France: useful if you serve EU clients under GDPR. RunPod and Lambda are US companies: ask where the card actually runs and read their data processing agreement.
Is the network good enough?
A public model serving hundreds of users, or a studio pushing multi-GB renders, depends on bandwidth. Ikoula guarantees 100 Mbps on a shared 1 Gbps link; OVHcloud lists 5 to 25 Gbps public. Fine for internal use. For a public service, ask before you sign.
Who sets up and watches the box?
A dedicated server shows up bare or close to it. NVIDIA drivers, updates, firewall and backups are on you. Make sure someone on the team knows Linux. RunPod and Lambda ship ready-made images, which saves time on day one.
- Your model fits in the card's VRAM, with headroom
- Your monthly GPU hours are estimated against the 300-hour mark
- Setup fees are in your year-one math
- Hosting location works for your customers
- Guaranteed bandwidth is known
- Someone on the team can install drivers and secure the server
Is your tool missing from this page? Email us: hello@hackceleration.com
Your dedicated GPU server questions
How much does a dedicated GPU server cost per month?
From about €195 to over €3,500 a month before tax in Europe in 2026. Entry level is Ikoula (12 GB RTX 4070 at €195) and Hetzner (24 GB at €214). A 96 GB card is €999 at Hetzner. OVHcloud's L4 server starts at $1,145. Add setup fees, up to a full month at Hetzner and OVHcloud.
Is a dedicated GPU server better than cloud GPU?
Only if the GPU runs a lot. Above roughly 300 hours a month, a dedicated server is cheaper. Below that, hourly cloud GPUs like RunPod or Lambda cost less and let you switch cards. Tip: rent hourly for a month first and log your usage.
What GPU do I need to run an LLM?
A 24 GB card runs an 8B to 14B model, or a 32B model quantized to 4 bits. A 70B model needs about 96 GB. Data center cards like the H100 (80 GB) are only worth it for training or high-traffic serving.
Can I rent a dedicated GPU server for just one month?
Yes. Hetzner has no minimum term and bills hourly up to the monthly cap. But setup fees apply to each order: €999 on a GEX63. For a few days, RunPod or Lambda by the hour is cheaper.
Is a consumer RTX card good enough, or do I need a pro card?
For 3D, video and small models, an RTX 4070 or 5070 Ti is fine. For AI, memory decides: RTX PRO cards come with 24 or 96 GB and ECC, which matters on a server running 24/7. Compact SFF versions draw little power: good for inference, weaker for heavy training.
Am I still billed when the server is off?
Yes. A dedicated server is billed as long as it exists; you have to cancel it. On RunPod, a stopped GPU costs nothing, but pod storage keeps billing until you delete it.
Can I get an H100 as a dedicated monthly server?
Not self-serve from the providers here. Hetzner tops out at the 96 GB RTX PRO 6000 and OVHcloud's Scale-GPU uses the L4. For H100 access, RunPod lists it at $2.89 an hour and Lambda at $3.29; long-term reserved H100 capacity at Lambda goes through sales.
Get the next comparison in your inbox
One tool broken down every month, prices checked, limits included.
