RunPod vs Hetzner for Ollama: Hetzner Sells GPUs Too, and They're Cheaper

Search this comparison and you’ll find the same framing everywhere: RunPod is the GPU option, Hetzner is the cheap CPU option, and the article is really about whether you can tolerate slow inference.

That framing is out of date. Hetzner sells dedicated GPU servers — the GEX line — and its own product page tells you what to run on them: “Get AI models and tools such as DeepSeek or Ollama running on our dedicated GPU servers.”

Once both vendors are selling GPUs, the comparison gets more interesting, because on the card they both offer, Hetzner is cheaper per hour and cheaper per month.

Two questions, not one

Untangle them before looking at any price:

1. Does your model fit, and do you need it fast? This is a RAM and vRAM question, and it’s what decides CPU versus GPU.
2. How much of the month is the GPU actually working? This is the duty-cycle question, and it decides hourly versus monthly.

Most people get the first one wrong in one direction and the second wrong in the other — renting a GPU for a workload a CPU would handle, or paying hourly rates around the clock for something a monthly server does cheaper.

The RAM floor

Ollama holds model weights in memory permanently. That’s the hard constraint, and it doesn’t care which vendor you use:

CPU inference is bounded by memory bandwidth, not core count. A quantised 7–8B model on a decent server produces output at roughly reading speed or slower. That is genuinely fine for summarising documents overnight, enriching records in a queue, or any job where nobody is watching the cursor. It is miserable for a chat interface.

So the honest first answer: if you’re running a 7–8B model for background work, you don’t need either vendor’s GPU. A 16GB CPU instance does it, and our best VPS for Ollama guide covers that side in detail.

The rest of this article is for the case where you do need the GPU.

What each vendor actually sells

Hetzner GEX — dedicated GPU servers, billed hourly with a monthly cap:

Hetzner removed the GEX131’s setup fee in December 2025 and says the server “can be used flexibly for a few hours or days without having to pay a setup fee” — so despite being dedicated hardware, it’s rentable by the hour.

RunPod pods, Secure Cloud tier:

RunPod also sells a cheaper Community Cloud tier on third-party hardware, and Serverless, which bills only while a request is processing.

The comparison nobody makes

Both vendors sell the RTX PRO 6000 with 96GB of vRAM. Line them up:

Hetzner is 24% cheaper per hour and 35% cheaper per month, with more RAM, more CPU and included NVMe. For a 96GB card — enough vRAM to hold a 70B model comfortably — that is not a close contest on price.

The monthly figure is a cap rather than a commitment: Hetzner bills the GEX131 hourly and stops at the monthly rate, which is why $989 is less than $1.5850 × 730 hours ($1,157). You get the monthly price without signing up for a month.

So why use RunPod at all?

Three genuine reasons, none of which is the hourly rate.

Scale to zero. RunPod Serverless bills only while a request is processing. Hetzner’s server bills whether it’s working or idle. For an endpoint that handles a handful of requests a day, that difference dwarfs the hourly rate — RunPod’s serverless RTX 6000 Pro at $3.49/hr costs about $283 a month at 39% duty, and far less below that. The crossover against Hetzner’s $989/month lands at roughly 283 hours of actual processing per month. Under that, serverless wins despite costing more than twice as much per hour.

Availability and geography. RunPod advertises thousands of GPUs across 30+ regions. Hetzner’s GPU servers are dedicated machines in European datacentres, subject to stock, and its own FAQ addresses whether particular cards are available at all. If you need a GPU in the next ten minutes, or one near users outside Europe, that’s RunPod’s argument.

Card selection. Hetzner offers two GPU configurations. RunPod offers a dozen, including H100, H200, B200 and A100 variants. If your workload is tuned for a specific card, Hetzner may simply not have it.

Putting it together

The GEX44 deserves a note of its own. At €184/month for a 20GB card with 64GB of system RAM, it’s the cheapest always-on GPU in this comparison by a wide margin — DigitalOcean’s comparable RTX 4000 Ada runs $0.76/hour, which is $555/month if you leave it on. The break-even against GEX44 arrives at about 37% duty cycle. Watch the €79 setup fee, which the GEX131 doesn’t carry.

The verdict

Use Hetzner if the GPU is going to be busy. Cheaper per hour, cheaper per month, more RAM and CPU around the card, and the monthly cap means you’re not committing to anything. For an always-on Ollama endpoint serving a team, this is the cheaper machine and it isn’t especially close.

Check Hetzner pricing → · Read our Hetzner review

Use RunPod if the GPU is going to be idle, or if you need a card Hetzner doesn’t stock, or a region outside Europe. Serverless is the genuinely differentiated product here — nothing in Hetzner’s range bills only while processing, and below roughly 280 GPU-hours a month that advantage is worth more than the hourly premium.

Check RunPod pricing → · Read our RunPod review

Use neither if you’re running a 7–8B model for background work. The most common mistake in this category is renting a GPU to do something a $15 CPU server does adequately overnight. Our GPU VPS guide covers the rent-versus-buy break-even, and RunPod vs DigitalOcean GPU Droplets works through the same duty-cycle arithmetic against a third vendor.

How we picked

The GEX131’s pricing and specifications come from Hetzner’s own pressroom announcement: €889.00 / $989.00 a month, €1.4247 / $1.5850 per hour, RTX PRO 6000 Blackwell Max-Q with 96GB GDDR7 ECC, Xeon Gold 5412U, 256GB DDR5 ECC and 2 × 960GB Gen4 NVMe, with the setup fee removed on 16 December 2025. The GEX44 and GEX131 specifications in the table above are visible in the screenshot from Hetzner’s own comparison matrix.

One figure we could not verify at source: the GEX44’s €184/month price. Hetzner does not render prices on its GPU product page — the numbers shown there (306.8 and 3511) are Tensor TFLOPS, not currency — and its cloud price calculator is a JavaScript widget that exposes nothing to read. The €184 and €79 setup figures come from a third-party plan listing, so treat them as indicative and confirm in Hetzner’s console before ordering. The same caveat applies to any Hetzner cloud CPU pricing you see quoted anywhere, including in our other guides.

RunPod’s rates are from its published pricing page and appear in the screenshot above.

What we did not do: we have not benchmarked tokens per second on either platform. The brief for this article called for measured throughput on a Hetzner CPU box against a RunPod pod, and we’d rather publish the cost analysis honestly than invent numbers to fill a table. Throughput depends on the model, the quantisation, the context length and the inference settings, and a figure measured on our hardware would not transfer to yours.

Measure it yourself in about ten minutes — ollama run <model> --verbose reports evaluation rate in tokens per second after each response, so run the same prompt on a CPU instance and a GPU pod and compare directly. That comparison, on your model, is worth more than any published benchmark.

Both vendor links above are affiliate links. Hetzner wins most of the price comparisons in this article on our own arithmetic; RunPod wins the ones about idle time and availability.

FAQ

Does Hetzner offer GPU servers?

Yes — the GEX line of dedicated GPU servers. GEX44 pairs an NVIDIA RTX 4000 SFF Ada (20GB vRAM) with 64GB of RAM; GEX131 pairs an RTX PRO 6000 Blackwell Max-Q (96GB vRAM) with a 24-core Xeon and 256GB of DDR5. Hetzner’s own page names Ollama as a supported use case.

Is Hetzner cheaper than RunPod for GPUs?

On the RTX PRO 6000 that both offer, yes: $1.5850/hour against RunPod’s $2.09, and $989/month against $1,526 always-on — with more RAM and CPU included. RunPod becomes cheaper when the GPU sits idle, because Serverless bills only while processing.

Do I need a GPU to run Ollama?

Not for background work. A quantised 7–8B model runs on an ordinary CPU server with 8GB of RAM at a few tokens per second, which is fine for summarisation, batch enrichment and scheduled jobs. You need a GPU for interactive use, for models above about 13B, or whenever a person is waiting on the output.

How much RAM does Ollama need?

Roughly 8GB for a quantised 7–8B model, 16GB for a 13B, 32GB for a 30B, and 40GB or more for a quantised 70B. The weights stay resident in memory, so this is a floor rather than a target.

At what point should I stop paying hourly?

Around 280 GPU-hours a month against Hetzner’s GEX131, which is the point where RunPod Serverless at $3.49/hour stops being cheaper than $989 flat. Below that, pay per request; above it, rent the machine.

Can I run a 70B model on Hetzner?

On the GEX131, yes — 96GB of vRAM holds a quantised 70B comfortably. The GEX44’s 20GB does not; that card suits 7–13B models.

Is RunPod’s Community Cloud worth using for Ollama?

For experimentation, yes — it’s meaningfully cheaper than Secure Cloud because it runs on third-party hardware. For anything holding customer data or needing predictable availability, use Secure Cloud or a dedicated server instead.