GPU Server Rental: Dedicated, GPU VPS or Hourly Cloud

Pick between a monthly dedicated GPU server, a GPU VPS and an hourly cloud GPU for an always-on job, with dated list prices and a break-even formula.

By Faiz Ahmed
14 min read

For a workload that runs all day, every day, price a monthly dedicated GPU server first if the card you need is a workstation or inference part and you will keep it for several months. An hourly cloud GPU is the better start if your real utilisation is low, if you need a training part such as a single H100 or H200, or if the job might end next month. A GPU VPS is the right answer only when a slice of a GPU is enough. The rest of this page gives you the prices and the formula to check that for your own case.

Decision table

Your workloadRent thisWhy
Inference or rendering that is busy most of the day, for three months or more, on one or two workstation-class cardsMonthly dedicated GPU serverFlat price, traffic included, whole machine is yours
Always on, but the GPU sits idle most of the time (a low-traffic API, a dev box, a bot)GPU VPS with a fractional GPU, or an hourly cloud GPU that you stopYou pay for a slice, or only for the busy hours
Busy in bursts: a few hours a day, or a few days a monthHourly cloud GPUStopping it stops the GPU charge
One H100, H200 or B200 around the clockHourly cloud GPU, a monthly-priced cloud VM, or a reserved dealFew dedicated hosts sell these one at a time
Eight GPUs in one box, for monthsBare metal 8-GPU server or a reserved clusterSee our GPU cluster pricing page
You do not yet know how long you will need itHourly cloud GPUNo setup fee to lose

The three products

Monthly dedicated GPU server

This is a whole physical machine, rented from a hosting company, with a GPU inside it. Nobody else runs on the hardware. You get root access, the full CPU, memory and disks, and a network port with a traffic allowance. The price is a flat monthly figure, and some hosts add a one-off setup fee.

The cards on offer are mostly workstation and inference parts. On 21 September 2026 Hetzner's GPU line-up was two servers, one with an RTX PRO 4000 Blackwell and one with an RTX PRO 6000 Blackwell. DataPacket listed L4 and RTX PRO 6000 Blackwell servers at the low end, and H200 and B300 parts only as eight-GPU machines. Latitude.sh was the exception in our research: it listed a bare metal server with a single H100.

"Bare metal GPU" means the same thing sold the cloud way. Latitude.sh describes "fully automated bare metal servers" billed hourly, monthly or yearly, and Vultr's plan API lists bare metal GPU machines next to its virtual ones.

GPU VPS

A GPU VPS is a virtual machine with a GPU attached. Often it is only part of a GPU. On 21 September 2026 Vultr's plan API listed slices of an NVIDIA A16 with 2 GB of GPU memory, 2 vCPUs and 8 GB of RAM, A40 slices that also start at 2 GB, and whole-GPU virtual machines such as an L40S with 16 vCPUs. OVHcloud's GPU products in our research were also virtual machines on its Public Cloud, not dedicated servers, even though they carry a monthly price.

Ask how the GPU is shared. NVIDIA's MIG documentation says MIG gives "fault isolation for different clients such as VMs, containers or processes". Time-sliced sharing does not give you that hardware isolation. Our MIG and fractional GPU explainer covers the difference.

Hourly cloud GPU

This is what the live tables on this site track: a virtual machine or a container with one or more whole GPUs, billed by the hour or finer, that you can stop. On Vast.ai an instance is a Docker container with "exclusive GPU access", and the CPU, RAM and storage are a proportional share of the host. RunPod bills pods "by the second for compute and storage". The price is per GPU, and the rest of the machine comes in whatever ratio the provider chose.

Dedicated server list prices on 21 September 2026

These are list prices before tax, read from each host's own price page or price feed on 21 September 2026. Re-check before you order. Hetzner prices are in euros and exclude VAT.

HostServerTypeMonthly list priceSetup feeTraffic and term
HetznerGEX45, 1x RTX PRO 4000 Blackwell SFF, Core i5-13500, 64 GB RAMDedicated€214.00 including IPv4€209Unlimited traffic on a 1 Gbit/s port, no minimum term
HetznerGEX131, 1x RTX PRO 6000 Blackwell Max-Q 96 GB, Xeon Gold 5412U, 256 GB RAMDedicated€1,199.00 including IPv4€599Unlimited traffic on a 1 Gbit/s port, no minimum term
DataPacket1x L4 24 GB, EPYC 4464PDedicatedfrom $516Not recordedUnmetered bandwidth plans offered
DataPacket1x RTX PRO 6000 Blackwell 96 GB, EPYC 9355Dedicatedfrom $1,785Not recordedUnmetered bandwidth plans offered
DataPacket2x RTX PRO 6000 BlackwellDedicatedfrom $2,687Not recordedUnmetered bandwidth plans offered
DataPacket8x H200 NVLDedicatedfrom $16,176Not recordedUnmetered bandwidth plans offered
OVHcloudl4-90, 1x L4Cloud VM, monthly rate$720None shownTraffic not metered
OVHclouda100-180, 1x A100 80 GBCloud VM, monthly rate$1,229None shownTraffic not metered
OVHcloudl40s-90, 1x L40SCloud VM, monthly rate$1,296None shownTraffic not metered
OVHcloudh100-380, 1x H100 80 GBCloud VM, monthly rate$2,070None shownTraffic not metered

Hetzner's price has moved a lot. Hetzner's press release for the GEX131 gave a launch price "from € 889.00" a month including an IPv4 address, with no setup fee. An edit to the release notes a €159 setup fee from 16 December 2025. On 21 September 2026 the price feed showed €1,199 a month and a €599 setup fee. That is about 35% above the launch price, and older blog posts still quote the launch figure. Hetzner's terms are generous in another way: "No minimum contract term" and "Cancellation period: immediately", with hourly billing up to the monthly maximum. The setup fee is the commitment.

OVHcloud's monthly rate is sometimes a real discount and sometimes not. By our arithmetic the a100-180 monthly price is about 45% below 730 hours at OVHcloud's own hourly rate. The h100-380 monthly price is only about 5% below. Do the sum for the SKU you want.

Some things could not be captured. OVHcloud also sells bare metal GPU servers and Leaseweb sells GPU dedicated servers, but both price pages render only in a browser, so we have no verified figure for either. Setup fees for DataPacket are not in our research.

Two hosts in our research are also in our live tables, so their prices are shown live instead of typed here. Latitude.sh bare metal plans all included "20 TB free out / ∞ in" on 21 September 2026. Vultr's plan API showed 15,360 GB of transfer with each bare metal GPU server and 10,240 GB with an L40S virtual machine. Neither price list showed a setup fee.

ProviderH100 $/GPU-hrL40S $/GPU-hrA100 $/GPU-hrB200 $/GPU-hr
Latitude.shnone in stocknone in stocknone in stocknone in stock
Vultrnone in stocknone in stock$2.80none in stock
Cheapest in-stock on-demand price per GPU-hour at each provider, from providers with live stock tracking. Latest stock observation: .

What an always-on cloud GPU costs per month today

This table takes the cheapest current hourly price for each GPU and multiplies it by 730 hours. "Current" means an in-stock, on-demand offer, seen in the last 15 minutes, from a provider whose stock we track live. Stock is rechecked about every minute.

GPU$/GPU-hrPer day (×24)Per month (×730)Provider
RTX 4090$0.74$17.76$540RunPod
L40S$0.97$23.28$708Massed Compute
A100$0.68$16.32$496LeaderGPU
H100$2.89$69.36$2,110Massed Compute
Cheapest in-stock on-demand price per GPU-hour, for one GPU running the whole time. Per day = hourly price × 24 hours. Per month = hourly price × 730 hours (365 days × 24 ÷ 12), rounded to the nearest dollar. Storage, data transfer and tax are not included. Latest stock observation: .

And the cheapest current offer for the GPUs that overlap with the dedicated hosts above:

GPUCheapest $/GPU-hrProviderProviders in stock
RTX 4090$0.74RunPod3
RTX PRO 6000 Blackwell$0.59RunPod6
L40S$0.97Massed Compute6
A100$0.68LeaderGPU9
H100$2.89Massed Compute6
Cheapest in-stock on-demand price per GPU-hour, from providers with live stock tracking. Latest stock observation: .

The monthly column assumes the GPU never stops, so it is the most the cloud option can cost you. Stop the instance at night and you pay less. A dedicated server gives you no such option.

The comparison method

Dedicated server, effective monthly cost = monthly price + (setup fee ÷ months you will keep it)

Hourly cloud GPU, monthly cost = hourly price × 730 × utilisation

Utilisation is the share of the month the instance is running and billed, from 0 to 1. It is not how hard the GPU works while it is on.

Set the two equal and you get the break-even:

Break-even utilisation = effective monthly cost ÷ (hourly price × 730)

Here is the dedicated side worked through with the Hetzner GEX131 figures from 21 September 2026. Keep it one month and you pay €1,199 + €599 = €1,798, which is €2.46 for each of the 730 hours. Keep it twelve months and the setup fee adds €599 ÷ 12 = €49.92 a month, so the effective cost is €1,248.92 a month, or €1.71 an hour. With no setup fee at all it would be €1.64 an hour. One month costs about 44% more per hour than twelve.

Now take the hourly price P for the same GPU from the live table above, convert it to the same currency, and divide. If P is double your effective hourly figure, the break-even is 50%: run the cloud GPU more than half the month and the dedicated server is cheaper. If P is equal to or below your effective hourly figure, the dedicated server never wins on price, and you would choose it only for the reasons below.

For context, Uptime Institute estimated in February 2025 that owning hardware outright beats a hyperscaler at 22% average utilisation or above, and beats a neocloud only at 66% or above. That is about buying, not monthly rental, but it shows how much the answer turns on which cloud you compare against.

What the formula does not capture

  • Bandwidth. The dedicated prices above include traffic: unlimited at Hetzner on a 1 Gbit/s port, not metered on the OVHcloud SKUs. Hourly clouds vary. RunPod charges "no fees for data ingress or egress", while on Vast.ai "Data transfer costs vary by host and include both upload and download traffic". If you serve a lot of data, read our data egress reference and add the cost to the cloud side.
  • The rest of the machine. A dedicated server price covers a named CPU, a fixed amount of RAM and local disks. A per-GPU cloud price covers whatever share the provider attaches. Compare the whole shape, not the GPU alone.
  • Single-tenant hardware. On a dedicated or bare metal server nobody else is on the box. That can matter for compliance.
  • No stop button. A dedicated server bills whether or not you use it. Some clouds lack one too: Lambda's documentation says instances "can only be launched, restarted, or terminated". And a stopped cloud instance still costs something, because storage keeps billing. On Vast.ai, "Storage charges continue even when instances are stopped".
  • The class of GPU. The RTX 4090 in the live table is a consumer card. None of the dedicated hosts in our research listed it. They listed workstation, inference and data center parts. A consumer card in the cloud and a workstation card on a dedicated server are different products, so start from the memory you need. The 4090 has 24 GB and the RTX PRO 6000 Blackwell has 96 GB.

Commitments on the cloud side

Also price a committed cloud deal, which is the cloud's answer to the dedicated server. Terms recorded on 21 September 2026:

  • RunPod savings plans: "Commit to a 3-month or 6-month term upfront". Its docs say plans are non-refundable.
  • Vast.ai reserved instances: "up to 50% discount with commitment", pre-paid, with credits locked to the specific instance.
  • Latitude.sh advertises reserved pricing paid upfront as "Monthly 50% cheaper than hourly" and "Paid yearly 65% cheaper than hourly". That is a site-wide claim. GPU-specific reserved prices were not shown, and the "monthly" figures on its price list are the hourly rate times 730.
  • OVHcloud Savings Plans run 12 or 36 months. AWS Savings Plans and Azure reserved instances run one or three years.
  • At the large end, CoreWeave's annual report describes committed contracts of one to six years that require "payment regardless of the level of utilization".

Our guide to on-demand, spot and reserved pricing goes through these models in detail.

Questions to ask before you sign

  1. What is the setup fee, and do you pay it again if you change server?
  2. What is the minimum term and the notice period? Get the cancellation terms in writing.
  3. Is the monthly price fixed for the term? Hetzner's GEX131 list price rose about 35% between launch and 21 September 2026.
  4. What traffic is included, on what port speed, and what happens above the allowance?
  5. Is the GPU the full card or a slice? If it is a slice, is it MIG or time-sliced?
  6. Which exact GPU variant is it? Hetzner's GEX131 lists the Max-Q variant of the RTX PRO 6000 Blackwell, and its GEX45 lists the SFF variant of the RTX PRO 4000 Blackwell.
  7. Who replaces failed hardware, and how fast?
  8. Is the price before tax? Hetzner's list prices exclude VAT.

The rule

Work out your effective monthly cost with the setup fee spread over the months you are sure of, not the months you hope for. Divide by 730 times the live hourly price. If the break-even utilisation is below what you will really run, and the host sells the card you need one at a time, rent the dedicated server. In every other case, start on an hourly cloud GPU such as the cheapest current RTX PRO 6000 Blackwell or H100 offer, measure your utilisation for a month, and then run the numbers again. Every tracked provider is on the providers page.

Sources

Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.

Frequently asked questions

Is a dedicated GPU server cheaper than an hourly cloud GPU?

Only when the GPU is busy most of the month and you keep the server long enough to spread the setup fee. Divide the monthly price plus the amortised setup fee by 730 times the live hourly price to get the utilisation at which the two cost the same.

What is a GPU VPS?

It is a virtual machine with a GPU attached, often only a slice of one. On 21 September 2026 Vultr's plan list included slices of an A16 or A40 card with 2 GB of GPU memory, next to whole-GPU virtual machines.

Do dedicated GPU servers have setup fees?

Some do. Hetzner charged a one-off setup fee on both of its GPU servers on 21 September 2026, while the Latitude.sh, Vultr and OVHcloud price lists showed none. Always add the fee to the first month before you compare.

Can I rent a bare metal server with a single H100?

Yes, but the choice is narrow. Latitude.sh listed a bare metal server with one H100 on 21 September 2026, while DataPacket listed H200 and B300 parts only as eight-GPU servers.

What is included in a dedicated GPU server price that an hourly GPU price leaves out?

Usually a traffic allowance, the whole machine's CPU, memory and disks, and single-tenant hardware. What it does not include is a stop button, so an idle dedicated server costs the same as a busy one.

Related Posts