Basis of this comparison
What we could verify
We checked Hetzner's current GPU matrix, all three product pages and Robot documentation on 20 October 2026. We did not run an independent three-server benchmark, so this page does not invent a performance score or a universal winner.
This link passes through Filedesign for click measurement. Referral attribution was not demonstrated in the latest route audit.
What Hetzner currently sells
Hetzner currently lists three dedicated GPU server families. GEX45 is the 24 GB entry system. GEX63 and GEX131 both use a 96 GB NVIDIA RTX PRO 6000 Blackwell Max-Q GPU, but GEX63 pairs it with a desktop-class CPU and 128 GB RAM while GEX131 starts with a server-class Xeon and 256 GB ECC RAM.
These servers are managed through Robot. They are not Hetzner Cloud virtual machines, and turning one off does not delete or cancel it. That distinction matters when comparing them with platforms where a stopped instance stops compute billing.
Current GEX45, GEX63 and GEX131 configurations
| Server | GPU | Base system | Public price context | Best starting point for |
|---|---|---|---|---|
| GEX45 | RTX PRO 4000 Blackwell SFF, 24 GB GDDR7 ECC | Core i5-13500, 64 GB RAM, 2 x 512 GB NVMe | 214 EUR/month incl. IPv4, plus 209 EUR setup | Inference, fine-tuning, CAD, rendering and video work that fits within 24 GB VRAM |
| GEX63 | RTX PRO 6000 Blackwell Max-Q, 96 GB GDDR7 ECC | Core Ultra 7 265, 128 GB DDR5, 2 x 1.92 TB NVMe | 999 EUR/month incl. IPv4, plus 999 EUR setup | 96 GB GPU work that does not need the larger ECC memory configurations of GEX131 |
| GEX131 | RTX PRO 6000 Blackwell Max-Q, 96 GB GDDR7 ECC | Xeon Gold 5412U, 256 to 768 GB DDR5 ECC, 2 x 960 GB to 4 x 3.84 TB NVMe | From 1,499 EUR/month, plus 1,499 EUR setup | 96 GB GPU work that also needs more ECC system memory, storage or a server-class CPU |
The amounts are the current public EUR prices shown on Hetzner's GPU matrix and product pages on 20 October 2026. The GEX45 and GEX63 launch announcements explicitly include one IPv4 address. Confirm the complete GEX131 order price and any IPv4 item in the live configurator. All amounts exclude VAT.
Strengths and limits
- Dedicated hardware with full root access
- 96 GB VRAM on one GPU with GEX63 or GEX131
- European data-center options and an optional DPA
- Hourly Robot billing with a monthly price cap
- One GPU per server and no additional graphics card
- Setup fees matter for short projects
- No H100 or A100 in the current GEX catalogue
- Powering off does not cancel the server or stop billing
How Hetzner GPU billing works
Hetzner documents pro-rated billing for dedicated servers managed through Robot. The server is billed for the time it exists, by the hour, and the server charge does not exceed the published monthly price cap. A dedicated server can be cancelled immediately.
Powering the server off does not cancel it, so billing continues until cancellation. The one-time setup fee is separate from usage. Additional hardware and licences may follow different rules, so include them when comparing a short project with an hourly GPU platform.
GEX45, GEX63 or GEX131
Choose GEX45 when the exact model or application fits within 24 GB VRAM and the fixed 64 GB system memory and two 512 GB NVMe drives are sufficient. Hetzner positions it for inference, fine-tuning, CAD, rendering and video work.
Choose GEX63 when the workload needs 96 GB VRAM but 128 GB of system memory, two 1.92 TB NVMe drives and the Core Ultra CPU are sufficient. It is the lower-cost 96 GB option in the current public matrix.
Choose GEX131 when the workload needs the same 96 GB GPU alongside a Xeon server CPU, registered ECC memory or more RAM and local storage. The GEX131 configurations change RAM and storage, not GPU count.
Hetzner's current pages disagree about exact details. The GPU matrix lists GEX131 in Falkenstein and Nuremberg, while the product detail page lists Helsinki and Falkenstein. The matrix also labels GEX45 memory differently from the GEX45 detail page. Confirm the live configurator before ordering rather than treating either summary as a contract.
None of the three current families can be configured with multiple GPUs. Hetzner's GPU server documentation says customers cannot install additional graphics cards. If a workload needs H100, A100, multiple GPUs or a tightly coupled accelerator cluster, the current GEX catalogue is not a match.
Data location and customer responsibility
Hetzner states that its GPU servers operate in ISO 27001-certified European data centers and offers an optional Data Processing Agreement under Article 28 GDPR. This can simplify an EU data-location requirement, but the server alone does not make an application compliant.
The customer remains responsible for system administration, software, monitoring, backups, security, access control, data flows and any external services used by the application. Confirm the selected location and contract requirements for the actual workload.
A practical acceptance test
Before keeping any server, test the exact workload rather than relying on GPU memory alone:
- Record the server type, location, image, driver and CUDA version.
- Record the model, quantization, context length, batch size and concurrency.
- Measure cold start, time to first result, steady throughput and 95th percentile latency.
- Record peak GPU memory, system memory, storage and network use.
- Confirm that the application fits on the single installed GPU.
- Run long enough to detect thermal, storage or network constraints.
- Calculate server time, setup fee, IPv4, licences and required add-ons for the same test period.
Publish a winner or score only after the same test has been repeated on every candidate and the result data is available.
Primary sources and method
This is a source audit, not a hands-on benchmark. Product facts and billing statements link to Hetzner's own current pages:
- Current GPU server matrix
- GEX45 launch announcement from 1 September 2026
- GEX45 product details
- GEX63 launch announcement from 6 October 2026
- GEX63 product details
- GEX131 product details
- GEX configurations and expansion limits
- Robot cancellation and hourly billing
Related GPU comparisons
- Compare GPU clouds with European options
- Review RunPod for shorter or more variable workloads
- Compare all GPU cloud providers
- Read the GPUHosted editorial methodology
Verdict
GEX45 is the entry point when 24 GB VRAM is enough. GEX63 is the lower-cost current path to 96 GB VRAM. GEX131 is the relevant 96 GB option when the workload also needs a server-class CPU, registered ECC memory or larger RAM and storage configurations. Setup fees, the single-GPU limit and cancellation-based billing make all three a stronger fit for sustained, well-defined workloads than for brief experiments.
Hetzner GPU FAQ
Does Hetzner Cloud offer GPU instances?
The current GPU products are dedicated GEX servers managed through Robot. Do not treat them as stoppable Hetzner Cloud virtual machines.
Does Hetzner offer H100 or A100 servers?
Hetzner's current GEX matrix lists RTX PRO 4000 Blackwell and RTX PRO 6000 Blackwell Max-Q GPUs across GEX45, GEX63 and GEX131. It does not currently list H100 or A100. This describes the current catalogue, not every past or future Hetzner product.
Can I install a second GPU?
No. Hetzner's GPU server documentation says additional graphics cards cannot be installed, and the current GEX server families each use one GPU.
Does powering off the server stop billing?
No. Hetzner bills the dedicated server until it is cancelled. Robot dedicated servers are billed by the hour up to the monthly price cap, and cancellation can take effect immediately.
Is a Hetzner GPU server automatically GDPR compliant for my application?
No infrastructure provider can determine that from the server alone. Hetzner states that its GPU servers operate in European data centers and offers an optional DPA. The customer still needs to assess configuration, access, logs, backups, external APIs and the actual data flow.