Private LLM hosting vs Hetzner GPU servers
I run GPUwerk, which rents DGX Sparks, so weigh everything below with that in mind. Hetzner, the German hosting company, rents dedicated GPU servers (its GEX line) in Germany and Finland, with no minimum contract. Its largest card is an NVIDIA RTX PRO 6000 Blackwell Max-Q with 96 GB of GDDR7, and the GEX430 puts four of them in one server. A DGX Spark is one machine with 128 GB of unified memory, rented by the minute. The choice is about memory layout, billing and how much server you want to run yourself, not price alone.
Side by side
| Hetzner GPU servers | Dedicated DGX Spark (GPUwerk) | |
|---|---|---|
| What it is | Dedicated root servers: GEX45 (one RTX PRO 4000 Blackwell SFF, 24 GB) at €214 a month, GEX63 and GEX131 (one RTX PRO 6000 Blackwell Max-Q, 96 GB) at €999 and from €1,499, and GEX430 (four RTX PRO 6000 Blackwell Max-Q, 384 GB in total) at €3,599, each plus a one-time setup fee, excluding VAT. One product line in a much larger server and cloud catalogue. | One dedicated DGX Spark, rented by the minute, configured by GPUwerk so you can start from the browser terminal. (advantage) |
| Region | GEX131 in Falkenstein, Germany, and Helsinki, Finland; GEX45, GEX63 and GEX430 in Helsinki, per Hetzner's GPU server pages. | EU-Central (Hradec Králové) on GPUwerk's own tier. One location. |
| Memory architecture | A discrete card with its own VRAM (24 GB or 96 GB GDDR7) plus separate system RAM (64 GB up to 768 GB, depending on the model). | 128 GB unified memory shared by CPU and GPU, enough for large quantized models in one pool. (advantage) |
| Who can see your data | Governed by Hetzner's terms and data processing agreement. It is a dedicated server, so no other customer shares it. | No other customer, and GPUwerk only under the DPA. GPUwerk's DPA states GPUwerk "hosts the machine but does not access, read, copy, index, or analyse the content of the controller's instance." |
| Model choice | Whatever you deploy. A 96 GB card holds models up to roughly the size of gpt-oss-120b; x86 hosts run any container image. | Whatever you deploy, within 128 GB unified memory: gpt-oss-120b, Llama 3.3 70B, Qwen3-Coder 30B, DeepSeek and similar. A linked pair doubles that. ARM64, so images need an arm64 build. |
| Speed and concurrency | Depends on the card. Hetzner publishes specs, not tokens per second; benchmark your model. | gpt-oss-120b: 33.5 tok/s single-stream and 862.8 tok/s aggregate at 256 concurrent requests (Dendro Logic's benchmark on its own Spark). Methodology on GPUwerk's benchmarks page. |
| Pricing model | A monthly price with an hourly rate and a one-off setup fee, per server model, on Hetzner's own pages. No minimum contract term. | Per hour, billed per minute, no setup fee: $0.79/hour on-demand, $0.59/hour while stopped and held for you, per pricing. (advantage) |
| Worked cost example | See the arithmetic below the table. | See the arithmetic below the table. |
| Contracts and DPA | Hetzner's customer agreement and data processing terms. Both companies are EU entities; check the specifics against your requirements. | A standard GDPR Article 28 DPA published free at /legal/dpa. The sub-processor list names none for instance workloads on GPUwerk's own tiers. |
| What you operate yourself | Everything on the server: OS, drivers, model server, monitoring, backups. Hetzner runs the hardware and network. | Less: GPUwerk maintains the host and drivers. You get root access, from the browser terminal or over SSH with your own key, to your own container: the model server, any RAG or agent layer, monitoring and backups. GPUwerk keeps no copy of /workspace, not even for a failed drive. (advantage) |
The worked cost example
Take a workload that generates 500 million output tokens a month, about what a busy internal support bot or a document pipeline running all day might produce.
Hetzner. Take the monthly price of the GEX server you would use from Hetzner's site, add the setup fee spread over the months you expect to keep it, and measure your model's tokens per second on that card. A fair comparison needs both numbers for your own model, so this page does not invent one.
Dedicated Spark. gpt-oss-120b reaches 862.8 tok/s aggregate at 256 concurrent requests, per Dendro Logic's concurrency benchmark on a single Spark (cited on GPUwerk's benchmarks page). That is 862.8 × 3,600 = 3,106,080 tokens per hour if the node stays saturated at that concurrency the whole time. 500,000,000 ÷ 3,106,080 ≈ 161 hours. At $0.79/hour, 161 × $0.79 = $127.19, before tax.
That figure holds only if the Spark serves 256 concurrent requests for all 161 hours. If you need the model available all month, compare Hetzner's monthly price with a Spark running 730 hours ($576.70). A Hetzner server kept busy around the clock can come out cheaper, and its discrete card can be faster per request. Benchmark your own model on both before you decide.
Migration path
Both are machines you deploy your own stack on. Run vLLM on either and you get the same OpenAI-compatible API, because the endpoint comes from vLLM, not from the hosting provider. Put LiteLLM in front of either for request logging and key management. What changes is the hardware underneath: x86 and a discrete card on Hetzner, ARM64 and unified memory on a Spark. Re-test throughput, and check your images have an arm64 build, when you move in either direction.
When Hetzner is the right choice
- You want a specific discrete GPU, x86, or a server you keep for months on one bill.
- You already run servers at Hetzner and want the GPU on the same account and tooling.
- You are comfortable maintaining the OS and drivers yourself, and a monthly price suits you better than paying by the minute.
When a dedicated Spark is the right choice
- Your workload is LLM inference and 128 GB of unified memory fits how you would quantize and serve the model.
- You want to pay by the minute for an evaluation or a part-time workload, with no setup fee.
- You would rather not maintain a general-purpose server for one job.
FAQ
Is Hetzner a good alternative to GPUwerk for LLM hosting?
It can be. Hetzner rents dedicated GPU servers in Germany and Finland with no minimum contract, including a 96 GB RTX PRO 6000 Blackwell model. You run the whole server yourself. A GPUwerk DGX Spark has 128 GB of unified memory, is billed by the minute and comes configured for you. If your model fits in 96 GB and you want a server for months, compare Hetzner's current prices directly.
What's the difference between a DGX Spark and a typical Hetzner GPU server?
Memory layout and CPU architecture. A DGX Spark has 128 GB of unified memory shared by an ARM64 CPU and the GPU. Hetzner's GEX servers pair an x86 CPU and system RAM with a discrete card that has its own 24 GB or 96 GB of GDDR7.
Is a DGX Spark cheaper than a Hetzner GPU server?
It depends on how many hours you need. In the worked example on this page, 500 million generated tokens a month cost about $127 in Spark compute at $0.79/hour, assuming the node sustains 256 concurrent requests the whole time. Running a Spark all month costs $576.70. Compare that with the monthly price and setup fee of the Hetzner server you would use.
Why choose a DGX Spark over a Hetzner GPU server?
For more memory in one pool than Hetzner's largest card, billing by the minute, and a machine GPUwerk maintains for you. If you need x86, a specific discrete GPU, or a server on the same account as your other Hetzner machines, Hetzner may serve you as well or better.
Hetzner claims checked on 9 November 2026 against Hetzner's GPU server overview, the GEX131 and GEX45 pages and the GPU server docs (models, GPUs, memory, locations, no minimum contract). Monthly prices in euros excluding VAT, from Hetzner's GPU server pages on 9 November 2026; Hetzner also lists an hourly rate for each server. Spark throughput figures come from Dendro Logic's concurrency benchmark, detailed on GPUwerk's benchmarks page.