Servers & GPU Hosting

How much does a GPU dedicated server cost in 2026?

Published 2026-10-28 · Researched and price-checked on official sources

A GPU dedicated server starts at €214 per month for 24 GB of VRAM at Hetzner, or €1,499 per month for 96 GB, both verified on 28 October 2026. Hourly cloud GPUs look much cheaper: RunPod lists an RTX 4090 pod at $0.34 per hour on Community Cloud. The break-even depends on utilisation, not on the sticker price.

Key facts

What matters

Hetzner's 24 GB GEX45 costs €214.00 per month plus a one-off €209.00 setup fee, which works out to about $239 per month at the ECB rate of 0.89686 EUR per USD.
The 96 GB GEX131 is €1,499.00 per month with a €1,499.00 setup fee, so the first year costs €19,487 excluding VAT, not €17,988.
RunPod charges $0.34 per hour for a 24 GB RTX 4090 on Community Cloud and $0.89 per hour for the same card on Secure Cloud, a 2.6x gap that changes every comparison you read.
A Hetzner GEX45 only beats a $0.34/hour 4090 pod after 702 billable hours in a month; against Community Cloud's $0.22/hour RTX 3090 it never catches up.
At fp16 a model needs roughly 2 GB of VRAM per billion parameters, so 24 GB holds about 9B parameters after overhead and 96 GB holds about 38B.

What is a GPU dedicated server?

A GPU dedicated server is a physical machine with one datacentre GPU that you rent whole: the card, the CPU, the RAM and the local NVMe storage are yours for the billing period, and no other tenant schedules work on that card. You get root access, install your own drivers and CUDA build, and pay a fixed monthly price.

That is a different product from a cloud GPU pod. Pods bill by the minute or second, run on shared hosts, and can be preempted or simply unavailable when the region is busy. The trade is flexibility against isolation and predictable cost.

Hetzner is the clearest example of the monthly model. Its GEX45 pairs an NVIDIA RTX PRO 4000 Blackwell SFF Edition (24 GB GDDR7) with an Intel Core i5-13500, 64 GB DDR4 and 2x512 GB NVMe on a 1 Gbit/s port with unlimited traffic. The GEX131 steps up to an RTX PRO 6000 Blackwell Max-Q with 96 GB GDDR7, a Xeon Gold 5412U, 256 GB DDR5 and 2x960 GB NVMe.

Verified GPU dedicated server prices, October 2026

Hetzner's GEX45 product page lists €214.00 per month or €0.3429 per hour, both excluding VAT, plus a one-off €209.00 setup fee. At the ECB reference rate of 1 USD = 0.89686 EUR (27 October 2026) that is about $239 per month before tax.

The GEX131 page lists €1,499.00 per month, €2.4022 per hour and a €1,499.00 one-time setup fee. The setup fee is easy to miss and it does real damage to the first-year number: twelve months of GEX131 plus setup is €19,487 excluding VAT.

Both prices are for the whole server, not just the GPU. That distinction matters when you compare against hourly rentals, because CPU cores, system RAM and local NVMe are included here and usually billed separately, or capped, in cloud pods.

GEX45: 24 GB GDDR7, 64 GB RAM, 2x512 GB NVMe, unlimited traffic, from €214.00/month.
GEX131: 96 GB GDDR7, Xeon Gold 5412U, 256 GB DDR5, 2x960 GB NVMe, €1,499.00/month.
Contract terms: Hetzner's GEX131 page states no minimum contract term and immediate cancellation, so the monthly price is not locked behind a year.

What the same class of GPU costs by the hour

RunPod's pricing page opens with the Secure Cloud tier selected, which is exactly how most price comparisons get the numbers wrong. Checked on 28 October 2026, the same page lists an RTX 4090 with 24 GB at $0.34 per hour on Community Cloud and $0.89 per hour on Secure Cloud. For the 96 GB RTX Pro 6000 the pair is $1.69 and $2.49 per hour.

The rest of the ladder, Community Cloud, per hour: RTX A6000 48 GB $0.33, A40 48 GB $0.35, L40S 48 GB $0.79, RTX 6000 Ada 48 GB $0.74, A100 80 GB PCIe $1.19, H100 PCIe $1.99, H100 SXM $2.69, H200 141 GB $3.59. Paperspace, now part of DigitalOcean, lists an H100 at $3.09 per hour.

Two things follow. First, the Community versus Secure split is worth 2.6x on a 4090, more than the gap between most vendors. Second, an older card is startlingly cheap: the RTX 3090 with the same 24 GB is $0.22 per hour on Community Cloud, a third of what the newer 4090 costs.

The break-even: when monthly rental beats hourly GPU cloud

Convert the monthly price to an hourly one and the comparison becomes honest. €214 per month is $238.61, which is $0.327 per hour if the server runs all 730 hours in a month.

Against RunPod's $0.34/hour 4090 the GEX45 breaks even at 702 hours per month, roughly 96% utilisation. Against the Secure Cloud 4090 at $0.89 it breaks even at 268 hours. Against the $0.22/hour RTX 3090 it needs 1,085 hours, which a month does not contain: 730 hours on that 3090 costs $161, less than the Hetzner box before its setup fee.

The 96 GB comparison is worse for the dedicated option. GEX131 is $1,671 per month, or $2.29 per hour across 730 hours. Community Cloud's 96 GB RTX Pro 6000 at $1.69 per hour costs $1,234 for the same month, so GEX131 needs 989 hours to break even, again impossible. It only wins against the $2.49 Secure Cloud rate, at 671 hours.

The conclusion is uncomfortable for the old advice. On pure GPU hours, monthly rental is not the cheap option any more. It wins when you are buying the rest of the machine: 256 GB of RAM, local NVMe, unlimited traffic, a single tenant and a bill that does not move when an experiment runs long.

VRAM arithmetic before you pick a card

Model size decides the card, not the other way round. At fp16 each parameter takes about 2 bytes, so a 70B model needs roughly 140 GB of weights before anything else. Add about 20% for activations and fragmentation and you are at 168 GB.

24 GB card: about 19 GB usable after overhead, so roughly 9B parameters at fp16, or a 70B model at 4-bit if the KV cache stays small.
48 GB card: about 38 GB usable, so roughly 19B at fp16 or 70B at 4-bit with headroom.
96 GB card: about 77 GB usable, so roughly 38B at fp16. A 70B model at fp16 still does not fit on one card.

4-bit quantization roughly halves the weight footprint again, at some quality cost. Long context is the other consumer: the KV cache grows with sequence length and batch size, so a model that fits at 2K tokens can fall over at 32K. Size the card for the context you actually serve, not for the benchmark prompt.

Who should rent a GPU dedicated server

The monthly model fits a narrow set of workloads, and it fits them well.

You serve your own fine-tuned weights continuously, so the card is busy most of the month.
Your job is bound by local disk or system RAM, for example fine-tuning with large dataloaders or rendering with big asset libraries.
You need root access, a specific CUDA or driver version, and no noisy neighbour on the GPU.
You want a fixed monthly line item instead of a metered bill that tracks every experiment.

It fits badly when your usage is bursty. An evaluation run of 100 GPU hours a month on a €214 box costs $2.39 per GPU hour, roughly seven times the effective rate of a server that stays busy. If your workload is a few experiments a week, hourly pods at Community rates are cheaper and you never pay for idle time.

The on-prem argument has strengthened: IBM's CEO said in October 2026 that open-weight models are now good and cheap enough to run on your own hardware (The Register, 22 October 2026). That case assumes the hardware is utilised. A GPU that idles is the most expensive way to buy inference.

Hidden costs the price pages do not lead with

Four costs sit outside the monthly headline. The setup fee is the biggest: €209 on the GEX45 and €1,499 on the GEX131, charged once and not refunded if you cancel in month two. VAT is excluded from every Hetzner figure quoted here, so add your local rate before comparing with a US vendor's pre-tax dollars. Currency drift matters on a 12-month commitment when your revenue is in dollars and the invoice is in euros; the ECB rate used here is 0.89686 and it will not stay there.

The fourth is traffic policy. Hetzner's unlimited traffic is genuine on the standard 1 Gbit/s port, but the GEX131 page notes that the policy does not apply to servers with the 10 Gbit/s uplink addon, where outgoing traffic over 20 TB is billed per TB. If your inference service streams large responses or model downloads to users, price that addon before you commit.

One more trap: the cheapest hourly card is usually the older one. An RTX 3090 at $0.22 per hour is a third of a 4090 and has the same 24 GB, but it is slower per token and the host is shared. Compare throughput per dollar, not VRAM per dollar.

At a glance

Option	GPU and VRAM	List price	Cost at 730 h/month	Fits
Hetzner GEX45	RTX PRO 4000 Blackwell SFF, 24 GB	€214.00/mo + €209 setup	$0.33/h equivalent	Always-on single-tenant inference
Hetzner GEX131	RTX PRO 6000 Max-Q, 96 GB	€1,499.00/mo + €1,499 setup	$2.29/h equivalent	Large models plus 256 GB RAM
RunPod Community	RTX 4090, 24 GB	$0.34/h	$248	Bursty experiments
RunPod Secure	RTX 4090, 24 GB	$0.89/h	$650	Managed hosts, less variability
RunPod Community	RTX Pro 6000, 96 GB	$1.69/h	$1,234	Same GPU class, cheaper hours
RunPod Community	RTX 3090, 24 GB	$0.22/h	$161	Cheapest 24 GB per hour

GPU dedicated server vs hourly cloud GPU, verified 28 October 2026

FAQ

Is a GPU dedicated server cheaper than renting GPUs by the hour?

Usually not, at current prices. Hetzner's 24 GB GEX45 works out to $0.327 per hour across a 730-hour month, which only beats RunPod's $0.34/hour RTX 4090 after 702 hours of use. Against Community Cloud's $0.22/hour RTX 3090 it never breaks even. The dedicated box wins on CPU, RAM, local NVMe, traffic and isolation, not on GPU dollars.

How much VRAM do I need to run a 70B model?

About 140 GB of weights at fp16, or roughly 168 GB once you allow for activations and fragmentation, so a single 96 GB card is not enough. At 4-bit quantization the weights drop to roughly 35 GB and fit on a 48 GB card if the context and batch size stay modest.

Do GPU dedicated servers require a long contract?

Not at Hetzner. Its GEX131 product page states there is no minimum contract term and cancellation is immediate, so you are not locked into twelve months. The one-off setup fee is charged at order time regardless, which is the real cost of leaving early.