Hetzner GPU Servers Compared: GEX45, GEX63 and GEX131 (October 2026)

Want to run a language model on a dedicated GPU without handling drivers, vLLM and monitoring yourself? The Managed GPU Server is located in a German data centre, with 24 GB of VRAM for €699 or 96 GB of VRAM for €1,799 net per month plus setup, including model, vLLM, Open WebUI and operations. Book an initial consultation

Hetzner rebuilt its GPU range in autumn 2026: the GEX45 arrived on 1 September, the GEX63 on 6 October, and the GEX44 can no longer be ordered. Many comparisons online therefore list models and prices that no longer apply.

This article puts every orderable Hetzner GPU server into one table: GPU, VRAM, CPU, RAM, storage, location, net and gross price, setup fee, term and cost per GB of VRAM. It then covers which LLM sizes fit on 24 and 96 GB, what self-managed operation involves and when an API is the better choice. All figures as of 13 October 2026.

Table of contents

• The Hetzner GPU servers at a glance
• Prices and cost per GB of VRAM
• Hourly price and billing
• Which LLMs fit on 24 and 96 GB
• GEX63 or GEX131
• Location: Germany or Helsinki
• What a self-managed GPU server needs
• Self-managed or Managed GPU Server
• When an API or GPU cloud is the better fit
• Common misconceptions
• Status and methodology
• Our approach at WZ-IT
• Further guides

The Hetzner GPU servers at a glance

Hetzner offers GPUs as dedicated root servers in its GEX line. The Hetzner documentation lists five current server types across three models:

GPU data comes from NVIDIA (RTX PRO 4000 Blackwell SFF, RTX PRO 6000 Blackwell Max-Q), server data from the product pages and the configurator.

All models share these points:

• One GPU per server. According to the documentation, additional graphics cards cannot be installed.
• Network: 1 Gbit/s port with guaranteed bandwidth and unlimited traffic. Only the GEX131 can be equipped with a 10 Gbit uplink for €43 net per month; traffic above 20 TB is then charged.
• Access: full root access, one IPv4 address, an IPv6 subnet, rescue system and images for Debian and Ubuntu, among others.
• Term: no minimum contract term, immediate cancellation.

No longer in the range are the GEX44 with an RTX 4000 SFF Ada (20 GB GDDR6), whose product page redirects to the GEX45 and whose order page lists it as currently not available, and the GEX130 with an RTX 6000 Ada (48 GB). Hetzner does not list data centre GPUs such as the H100 or H200.

Prices and cost per GB of VRAM

All prices include one IPv4 address (€1.70 net per month), as of 13 October 2026. Hetzner's English press releases state net prices ("All prices without VAT"), the German ones gross prices with 19% VAT; the website adds VAT depending on the selected country.

The GEX45 and GEX63 prices match the press releases (GEX45: €214.00, GEX63: €999.00), the GEX131 prices match the price data on the product page.

Three points stand out:

• The GEX45 has the cheapest GB of VRAM, but 24 GB limits model size.
• The GEX63 delivers the same 96 GB GPU as the GEX131-1 for €500 less per month, a third below the GEX131-1.
• GEX131-2 and -3 cost more without offering more VRAM. The premium pays for RAM and storage.

Prices are rising. The GEX131 launched in December 2025 at €889 net without a setup fee (since 16 December 2025: €159). In the price list for the 15 June 2026 adjustment, the GEX131-1 was listed at €1,197.30 excluding IPv4 with a €599 setup fee. On 13 October 2026 it is €1,499 and €1,499. The GEX44 cost €184 net at its launch in March 2024.

Hourly price and billing

Besides the monthly price, Hetzner lists an hourly price. According to the billing documentation, usage is billed by the hour, partial hours are rounded up, and the total never exceeds the monthly price.

For planning this means:

• From about 26 days of runtime, the full monthly price applies.
• Switching off saves nothing. A server is billed as long as it exists in the account. To stop paying, you have to cancel it.
• The setup fee applies per order. For short projects it makes up most of the cost, about 79% of a week on the GEX63.

Which LLMs fit on 24 and 96 GB

The calculation follows the article Sizing GPU and VRAM for LLMs: weights need about 2 GB per billion parameters in FP16, about 1 GB in FP8 and about 0.5 GB with 4-bit quantisation. vLLM uses 92% of GPU memory by default (--gpu-memory-utilization), so about 22 GB on 24 GB and about 88 GB on 96 GB.

The table shows how much of this budget remains after the weights. A few GB of that go to activations and CUDA graphs; the rest is available for the KV cache, i.e. context and concurrent requests.

How to read the figures:

• 24 GB (GEX45) handles 8B and 14B models with headroom for context. A 32B model only fits in 4-bit and leaves about 6 GB, so little room for parallel sessions.
• 96 GB (GEX63, GEX131) handles 32B models in FP8 with plenty of context and 70B models in 4-bit. According to its model card, gpt-oss-120b runs on a single 80 GB GPU in its MXFP4 quantisation.
• 4-bit files are larger in practice than the rule of thumb: a 70B model takes about 40 to 45 GB instead of 35 GB.
• Architecture matters for the cache. Qwen3-32B needs 256 KiB per token (8 GiB for one session with 32,768 tokens), the hybrid Qwen3.8-27B only 64 KiB. On 24 GB, Qwen3.8-27B in 4-bit leaves room for about 80,000 tokens of cache, a model with 256 KiB per token in the same space for about 20,000.

Beyond 96 GB, Hetzner has no single-server option. A 120B model in FP8 takes about 117 to 125 GB (for example Qwen3.5-122B-A10B-FP8) and needs two GPUs, as shown in the field report Qwen3.5-122B on two RTX PRO 6000. Spreading a model across two Hetzner servers is only possible over the network; the options are covered in Multi-GPU inference without NVLink.

Which configuration fits a specific use case depends on context length and concurrent requests. Formulas and worked examples are in the article on VRAM sizing, formats in the article on LLM quantization.

GEX63 or GEX131

Both servers have the same GPU. They differ in platform, location and price:

The provisioning times come from the order processing overview, last updated on 6 October 2026.

For inference with a single model, the GEX63 is the cheaper choice, because the GPU determines the speed. The GEX131 pays off if one of these applies: German location, ECC system memory, more than 128 GB of RAM (for example for data preparation or many containers next to the model), more local storage or 10 Gbit/s.

Location: Germany or Helsinki

Hetzner operates its own ISO 27001-certified data centre parks in Nuremberg and Falkenstein (Germany) and Helsinki (Finland). For GPU servers, the location is tied to the model:

Under data protection law, Helsinki is not a third country. Finland is an EU member state, and under Art. 1(3) GDPR the free movement of personal data within the Union may be neither restricted nor prohibited for data protection reasons. A server in HEL1 is therefore not at a disadvantage compared with one in Falkenstein.

A German location is still often a requirement, for example in customer contracts, tenders or internal policies. In that case Hetzner offers only the GEX131 for 96 GB and currently no model for 24 GB. WZ-IT's Managed GPU Server is located in a German data centre in both sizes.

The location alone does not define the data path. Model downloads, external APIs, telemetry and backup targets are separate data paths that also need to be defined.

What a self-managed GPU server needs

Hetzner provides hardware, network and images. The GPU server documentation states that the customer is responsible for all normal system administration tasks, explicitly software, monitoring, RAID configuration, backups and security. For a production LLM server, these are the building blocks:

The one-time setup is the smaller part. What remains is ongoing: updates, assessing vulnerabilities, model changes and on-call duty for incidents.

Self-managed or Managed GPU Server

A self-managed GEX is the cheapest way to get a dedicated GPU. WZ-IT's Managed GPU Server uses the same GPU models and includes operations. The comparison in figures, all prices net:

Self-managed fits if your team has Linux, NVIDIA and vLLM skills and someone can cover on-call duty. A GEX is then considerably cheaper.

Managed fits if nobody on your team should respond to alerts at night, if you need 24 GB in a German data centre, or if operations and changes must be documented. Additional models, RAG, SSO and high availability are added according to scope.

Already running a GEX? We also take over the operation of existing AI servers at the customer's hosting provider after a technical review, see AI server maintenance and operations. The price depends on configuration and scope.

When an API or GPU cloud is the better fit

A dedicated server pays off with continuous use. Other load profiles have better options:

The Hetzner Inference API is an experimental service with an OpenAI-compatible interface. As of October 2026 it offers Qwen3.6-35B-A3B-FP8 and Qwen3.8-27B, each with a context of 262,144 tokens. It is free of charge while it remains experimental and limited to 10 requests and 4 million input and 100,000 output tokens per minute and API key. According to the documentation, Hetzner does not store the content of requests and responses, but gives no guarantee for performance and availability and advises against production use. It is a good starting point for a first test with open-weight models, not for a production service.

A decision guide between model API and own inference is in the article Cloud AI vs. self-hosted.

Common misconceptions

"More VRAM makes the model faster." VRAM size determines whether model and KV cache fit on the card and how many requests run in parallel. According to NVIDIA, token generation is mostly limited by memory bandwidth. The RTX PRO 6000 is faster than the RTX PRO 4000 because its 1,792 GB/s is more than four times the bandwidth, not because of the 96 GB.

"Max-Q means throttled and unsuitable for LLMs." The Max-Q edition is designed for 300 W instead of 600 W. According to NVIDIA, its theoretical peak AI performance is 3,511 instead of 4,000 AI TOPS, about 12% below the Workstation Edition. Memory and bandwidth are identical: 96 GB GDDR7 with ECC and 1,792 GB/s.

"The GEX131 is faster than the GEX63 because it costs more." Both have the same GPU. The premium pays for CPU, ECC memory, storage, network option and the Falkenstein location.

"A mixture-of-experts model only needs memory for its active parameters." With MoE models, all parameters must be loaded. Qwen3.6-35B-A3B computes with about 3 billion active parameters but occupies memory for 36 billion.

"Hetzner servers are located in Germany." The GEX45 and GEX63 are only available in Helsinki. Only the GEX131 can be ordered in Falkenstein.

"Switching off saves money." Hetzner bills a server as long as it exists in the account. Costs only end with cancellation, and every new order triggers the setup fee again.

"With --api-key, vLLM is secured." According to the vLLM documentation, some endpoints remain reachable without authentication even with an API key. A reverse proxy that only exposes the required endpoints is recommended.

Status and methodology

• Data collected: 13 October 2026.
• Hetzner data: product pages and configurator for the GEX45, GEX63 and GEX131, the price data used by the website, press releases and the documentation on GPU servers, provisioning times, billing and price adjustment.
• GPU data: NVIDIA product pages for the RTX PRO 4000 Blackwell SFF, RTX PRO 6000 Blackwell Max-Q and RTX PRO 6000 Blackwell Workstation Edition.
• Model data: parameter counts from the Hugging Face model cards.
• Prices: net including one IPv4 address (€1.70). Gross = net × 1.19 (German VAT). Hourly price = server plus IPv4.
• € per GB VRAM per month = net monthly price ÷ VRAM in GB.
• € per GB VRAM over 12 months = (12 × monthly price + setup) ÷ 12 ÷ VRAM.
• Monthly threshold = monthly price ÷ hourly price.
• LLM table: weights = parameters × 2 bytes (FP16), × 1 byte (FP8), × 0.5 bytes (4-bit). Remaining = vLLM budget (0.92 × VRAM, i.e. 22 GB or 88 GB) minus weights, rounded to whole GB.
• WZ-IT prices: net excluding VAT according to the Managed GPU Server offer page.

Prices and availability change; the provider's order page is authoritative. Tables and figures may be reused with attribution and a link to this article.

Our approach at WZ-IT

We provide dedicated GPU servers with operations and look after existing AI servers.

• Assess the workload: check model, quantisation, context, concurrent use and other compute loads against 24 or 96 GB of GPU memory.
• Managed GPU Server: Managed GPU Server 24 with an RTX PRO 4000 Blackwell SFF for €699 net per month plus €499 setup, Managed GPU Server 96 with an RTX PRO 6000 Blackwell Max-Q for €1,799 plus €999, both cancellable monthly and located in a German data centre.
• Included: one agreed model with vLLM, Open WebUI, proactive 24/7 monitoring, maintenance of operating system, drivers, vLLM and Open WebUI with CVE assessment, incident handling according to the service level, documentation and a personal point of contact.
• More than 96 GB: servers with two, three or four RTX PRO 6000 Blackwell Max-Q (192, 288 or 384 GB), price on request.
• Existing server: operation of AI servers at the customer's hosting provider after a technical review via AI server maintenance and operations.

If a self-managed GEX or a model API is the better fit, we will tell you.

Further guides

• Sizing GPU and VRAM for LLMs - formula for weights, KV cache and overhead
• Multi-GPU inference without NVLink - tensor parallelism, pipeline parallelism or replicas
• vLLM with Qwen3.5-122B on two RTX PRO 6000 - field report for models above 96 GB
• Managed GPU Server 96 with RTX PRO 6000 Blackwell - the 96 GB configuration in detail
• vLLM, Ollama or llama.cpp - which inference software fits which use case
• European LLM APIs compared - when an API is enough instead of your own GPU
• Hetzner Cloud vs dedicated server - every type with price per vCPU and GB of RAM
• GPUs on Proxmox: passthrough and vGPU - virtualising a GPU server
• WZ-IT Hetzner expertise - operating servers in your Hetzner account

Planning to run your own language model and looking for the right GPU? We review model, context and number of users and propose the right path: self-managed server, Managed GPU Server or API. Book an initial consultation

Sources

• Hetzner: GPU server overview
• Hetzner: GEX45
• Hetzner: GEX63
• Hetzner: GEX131
• Hetzner Docs: GPU Server, possible configurations and add-ons
• Hetzner press release: GEX45 (1 September 2026)
• Hetzner press release: GEX63 (6 October 2026)
• Hetzner press release: GEX131 (11 December 2025)
• Hetzner press release: GEX44 (4 March 2024)
• Hetzner press release: GEX130 (24 September 2024)
• Hetzner Robot: GEX44 order page
• Hetzner Docs: Price adjustment 15 June 2026
• Hetzner Docs: Billing system
• Hetzner Docs: Order processing
• Hetzner Docs: Inference API
• Hetzner Docs: Robot firewall
• NVIDIA: RTX PRO 4000 Blackwell SFF Edition
• NVIDIA: RTX PRO 6000 Blackwell Max-Q Workstation Edition
• NVIDIA: RTX PRO 6000 Blackwell Workstation Edition
• NVIDIA Technical Blog: Mastering LLM Techniques, Inference Optimization
• NVIDIA Technical Blog: Open-Source GPU Kernel Modules
• NVIDIA: CUDA 12.8 Release Notes
• vLLM: Engine arguments
• vLLM: Security
• vLLM: Metrics
• Ollama: FAQ
• NVIDIA DCGM Exporter
• Hugging Face: Mixture of Experts Explained
• Hugging Face: gpt-oss-120b model card
• Regulation (EU) 2016/679 (GDPR)