Hetzner review 2026: five Hetzner Cloud plans measured in our Lab

This Hetzner review is based on five Hetzner Cloud servers we rented and benchmarked in Helsinki on 5 November 2026. Short answer: Hetzner is still very good value for self-hosted apps in Europe. None of the five showed any CPU steal, disk speed was strong on every plan, and the 4 vCPU CX33 ran Llama 3.2 3B at 13.5 tokens/s for 8.49 euro net per month. The catch is the cheapest x86 plan: our CX23 got an older Intel CPU that scored 913.87 sysbench events/s on one thread, against 3,248.26 on the ARM CAX11 that costs 0.50 euro more.

Why now: Hetzner raised cloud prices twice in 2026, the cost-optimized line has had long stretches of "currently unavailable", and many people are moving n8n, Supabase and small local LLMs onto a VPS. Old reviews with 2023 prices and no app tests no longer help you choose.

The finding that surprised us most: the same plan family can land on very different hardware. Inside one evening we saw an Intel Xeon Skylake, an AMD EPYC-Rome and an Ampere Neoverse-N1, and the speed gap between them was larger than the price gap.

Tested in the Lab Hetzner CX23, CX33, CX43, CAX11, CAX21 · hel1 · Nov 2026

This page sits in our Hetzner hub. If you only care about one plan, read the CX23 benchmark or the CX vs CAX comparison.

How we tested

We created the five servers in Hetzner's Helsinki location (hel1) with Ubuntu 24.04.4 LTS and ran the same script on each, started within 21 seconds of each other. The script runs sysbench for CPU and memory, fio for disk, curl downloads for network, vmstat for CPU steal, then installs Docker, the official Supabase docker compose stack and Ollama. Our separate production n8n server is also a CX23, running Ubuntu 26.04. The full method is on how we test.

Two limits to keep in mind. Each number is one run on one server, so it tells you what that server did, not what every server of that plan does. And sysbench's CPU test counts prime-number calculations per second ("events/s"), which is a narrow synthetic workload. We show app results (Ollama and Supabase) next to it for that reason.

Benchmarks of five Hetzner Cloud plans

All five plans are "shared vCPU": you get virtual CPU threads on a host that other customers also use. CPU steal is the share of time your virtual machine wanted the CPU but the host gave it to someone else. We measured it over 30 seconds on each server and got 0.0 % on all five.

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

The "Download" column is the cross-datacenter test from Falkenstein. The in-datacenter numbers are in the network section below.

The CPU lottery: Intel Skylake, AMD EPYC-Rome or Ampere

Hetzner says its cost-optimized plans use "older but reliable processor generations from Ampere, Intel, and AMD" and labels the CX plans "Intel/AMD". It does not tell you which one you will get. Here is what each of our servers reported:

• CX23: Intel Xeon Processor (Skylake, IBRS, no TSX)
• CX33 and CX43: AMD EPYC-Rome Processor
• CAX11 and CAX21: Neoverse-N1 (Ampere, ARM)

On one thread, the CX23 did 913.87 events/s. The CX33 did 1,620.94, and the ARM CAX11 did 3,248.26, about 3.5 times the CX23. That ratio exaggerates the real-world gap, because ARM cores do very well on sysbench's prime test. In Ollama the CAX11 generated Llama 3.2 3B at 6.5 tokens/s against 5.8 on the CX23, and Llama 3.2 1B at 15 against 10.3. The ARM server still won, just by much less.

We are not the only ones who drew the Intel CPU. A public CX23 run in Falkenstein from April 2026 on VPSBenchmarks reports the same "Intel Xeon Processor (Skylake, IBRS, no TSX)" string. Treat the CX23 as a plan that may be Intel or AMD, and check with lscpu right after you create it. If you get Skylake and your workload is CPU-bound, delete the server and create a new one, or pick a CAX11.

The sysbench gap also did not carry over to n8n. In our webhook load test at 25 concurrent requests, the Skylake CX23 handled 15.04 executions/s and the CAX11 12.92. A CPU benchmark that ranks one plan 3.5 times ahead can still lose the app test.

One more oddity: the CX33 and CX43 both reported AMD EPYC-Rome, yet the CX43 did 3,470.38 single-thread events/s against 1,620.94 on the CX33. Same CPU name, more than twice the per-thread result. We only have one run per server, so we report it and do not guess the cause.

Disk

IOPS means input/output operations per second; 4k random reads at queue depth 32 are a good proxy for database work. Every plan landed between 41,644 (CX33) and 53,796 (CAX21) random-read IOPS, and between 35,013 and 47,996 random-write IOPS. Sequential 1 MB reads ran from 1,936 MB/s (CX33) to 2,965 MB/s (CX23). Sequential writes were the most uneven: 1,283 MB/s on the CX23 against 2,139 MB/s on the CAX11. For Postgres behind n8n or Supabase on a small server, any of these is plenty.

Network: in-datacenter fast, cross-datacenter much slower

We timed a 40-second download of Hetzner's own test file from three places. From the Helsinki test server (same location) the five servers pulled 6,332 to 8,013 Mbit/s. From the CacheFly CDN they got 3,759 to 5,189 Mbit/s. From Hetzner's Falkenstein test server in Germany they got only 285 (CX23) to 835 Mbit/s (CX43).

So the CX23 was about 22 times slower pulling from Falkenstein than from Helsinki. This is one sample per server over a long route, not a ceiling on Hetzner's network. It matters if you plan to back up a Helsinki server to a Storage Box or second server in Germany: put the backup target in the same region, or expect transfers measured in hundreds of Mbit/s, not gigabits. Latency to public DNS was low everywhere: 1.351 to 2.416 ms to 1.1.1.1.

What real apps did on each plan

Ollama (local LLMs on CPU)

Tokens/s is how many word pieces the model writes per second; around 10 feels like a fast typist. We ran a 256-token answer at temperature 0.

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

The CX43 led with 21.2 tokens/s on Llama 3.2 3B. The CX33 (13.5) beat the more expensive CAX21 (12.2). Llama 3.1 8B ran at 9.7 tokens/s on the CX43, 6.3 on the CX33 and 5.7 on the CAX21. On the 4 GB CAX11 the 8B model failed with "insufficient memory (attempted to allocate 4402.48 MB)". On the 4 GB CX23 it was OOM-killed after about 50 seconds: Ollama's fit check expected 1,698 MiB to stay free, then the kernel's out-of-memory killer stopped llama-server. Neither 4 GB plan has swap by default. More on model sizing in Ollama on a VPS.

Supabase (official docker compose)

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

All five servers started the 11 containers with 0 unhealthy after 90 seconds idle, including both ARM plans. Whole-system RAM use ranged from 1,564 MiB (CAX11) to 1,870 MiB (CX43). The image pull is where the small plans hurt: 110.7 s on the CX23 against 29.2 s on the CX43.

n8n

On our production CX23 (Ubuntu 26.04), n8n 2.46.6 with Postgres 16 went from nothing to a healthy instance in 127 seconds, and the whole server used 1,036 MiB at idle. On the test servers we also ran a webhook, Code node and response workflow under ApacheBench at 1, 10, 25 and 50 concurrent requests. An execution here is one full run of that workflow.

Measured by AI Hosting Lab on 2026-11-05 · 5 servers · method · raw data

No server failed a single request. Best throughput per plan: CX43 36.14 executions/s, CX33 31.93, CAX21 25.77, CX23 15.04 and CAX11 13.9. At 50 concurrent requests the median response took 1,539 ms on the CX43 and 3,786 ms on the CAX11, so small plans queue rather than fail. The highest n8n RAM we saw per plan was between 859.3 MiB (CAX21) and 1,049.6 MiB (CX33). Setup steps are in install n8n on Hetzner.

Hetzner prices today, from the API

These are the prices the Hetzner Cloud API returned for Helsinki on 5 November 2026. "Gross" includes the VAT applied to our account; your total depends on where you are billed.

From the Hetzner Cloud API, location hel1, checked 2026-11-05. "Incl. VAT" is what our Czech (EU consumer) account pays; businesses and buyers outside the EU usually pay the price excl. VAT. Primary IPv4 is billed separately.

Two things the price list does not tell you. First, the API price is the server alone. A primary IPv4 address costs 0.50 euro net per month on top, per Hetzner's Primary IPs docs. That is why Hetzner's website shows the CX23 from 5.99 euro net while the API says 5.49. Second, the newest CPX types are not always the cheapest per resource: the API lists the 1 vCPU, 1 GB CPX02 at 5.99 euro net, more than the 2 vCPU, 4 GB CX23.

The 2026 price changes in brief

Hetzner announced a general price rise for 1 April 2026; heise reported that existing products were included and cited hardware procurement costs, especially RAM and NVMe. A second cloud change followed on 15 June 2026. Hetzner's own changelog says the updated pricing applies "to new orders and rescales", so servers you already run keep their price unless you resize them. Per Bitdoze's price history, the CX23 went from 2.99 euro net at launch to 3.99 in April and 5.49 in June, which matches today's API price. The x86 CPX and CCX lines rose much more. Details and sources are in our Hetzner price increase 2026 page.

Availability is the other 2026 story. On 5 November 2026 Hetzner's cost-optimized page showed every CX and CAX model as "currently unavailable", yet we created all five test servers in hel1 the same day. Hetzner's status notice from 26 June 2026 says creation is restricted for new customers and for some existing customers chosen at random. Plan for the possibility that you cannot order the plan you want.

Which plan for which job

CX23

Best for: one n8n instance with Postgres; our production n8n idles at 1,036 MiB of RAM on it, and the test CX23 handled up to 15.04 webhook executions/s with no failures. Main drawback: ours got the Skylake CPU, 913.87 single-thread events/s, the slowest we measured. Skip it if: you want to run Supabase and n8n together or any local model above 1B.

CAX11

Best for: small local models on a budget; Llama 3.2 1B ran at 15 tokens/s against 10.3 on the CX23. Main drawback: ARM, so every image must have an arm64 build, and 4 GB was not enough for Llama 3.1 8B. Skip it if: n8n is the main job (13.9 executions/s at best, below the CX23) or you depend on an x86-only binary.

CX33

Best for: the all-rounder. 13.5 tokens/s on Llama 3.2 3B and 6.3 on Llama 3.1 8B for 8.49 euro net. Main drawback: its random-read IOPS (41,644) were the lowest of the five, still far above what a small database needs. Skip it if: you need 8B answers faster than about 6 tokens/s.

CAX21

Best for: ARM-native stacks and Supabase; it had the best random-read IOPS (53,796). Main drawback: 2 euro more than the CX33 and slower on our 3B and 8B Ollama tests. Skip it if: local LLMs are your main workload.

CX43

Best for: Llama 3.1 8B (9.7 tokens/s) or many containers at once. Main drawback: nearly double the CX33 price. Skip it if: a 3B model is good enough; the CX33 delivers it for 8.49 euro.

Who should use Hetzner, and who should not

Use Hetzner if your users or your data should stay in the EU, you are comfortable running Ubuntu and Docker yourself, and you want the most RAM and CPU per euro for n8n, Supabase or a small Ollama model. Every plan includes 20 TB of EU traffic, per Hetzner's product page.

Look elsewhere if most of your users are in the Americas or Asia (the cost-optimized CX and CAX plans are listed only for EU locations, and heise reported the steepest June rises in the US), if you need a managed server or phone support, or if you need to order today and the plans show as unavailable. We compare options in Hetzner alternatives.

A setup note for Ubuntu 24.04 and newer: sshd is started by ssh.socket. After you harden /etc/ssh/sshd_config.d/, systemctl reload ssh fails with "ssh.service is not active", even though new connections already use the new config. Our hardening script handles it like this:

sshd -t
# Ubuntu 24.04+ starts sshd through ssh.socket: "systemctl reload ssh" fails with
# "ssh.service is not active". New connections already read the new config.
systemctl reload ssh 2>/dev/null || echo "ssh.socket activation: config applies to new connections"

Next step: if the CX23 is on your shortlist, read the full Hetzner CX23 benchmark before you order.