Hetzner CX vs CAX: x86 and ARM measured on every test we run

Hetzner CX vs CAX comes down to x86 against ARM at almost the same size. On our servers in Helsinki on 5 November 2026, the ARM CAX plans won sysbench by a factor of two to 3.5, but the x86 CX plans won more of the app tests that matter for self-hosting. The 4 vCPU CX33 beat the CAX21 on Llama 3.2 3B (13.5 against 12.2 tokens/s) and on n8n webhooks (31.93 against 25.77 executions/s), and it costs 2 euro less.

Why this matters now: after the 2026 price changes, ARM is no longer the cheap option at Hetzner. In the Hetzner Cloud API every CAX size now costs more than the CX of the same vCPU and RAM, so the old advice "take ARM, it is cheaper and faster" needs a recheck.

The short version of what we found: synthetic CPU scores flatter ARM. Pick by the app you will run, not by the benchmark chart.

Tested in the Lab Hetzner CX23, CX33 (x86) · CAX11, CAX21 (ARM) · hel1 · Nov 2026

This comparison is part of our Hetzner hub. For the full lineup including the CX43, see the Hetzner review.

What CX and CAX are

Both are Hetzner's "Cost-Optimized" shared vCPU plans. Hetzner labels CX as "Intel/AMD" and CAX as "Ampere", says the line uses "older but reliable processor generations", and recommends it for "workloads that can handle variable CPU performance". A vCPU is one virtual CPU thread; "shared" means other customers' servers run on the same host.

The x86 side is a lottery. Our CX23 reported an Intel Xeon (Skylake) and our CX33 an AMD EPYC-Rome. Both CAX servers reported Neoverse-N1, an ARM core design, which Hetzner sells under the Ampere label. Hetzner's product page lists these plans for its EU locations only.

Prices: ARM now costs more at every size

Prices from the Hetzner Cloud API for hel1 on 5 November 2026, excluding VAT and the 0.50 euro net primary IPv4. Gross prices for our account are in the full list below. According to Bitdoze's price history, the CX23 and CAX11 launched at 2.99 and 3.29 euro net, so the gap was small before 2026 and has grown since. See Hetzner price increase 2026 for the timeline.

Round 1: CX23 vs CAX11 (2 vCPU, 4 GB)

The CAX11 scored 3,248.26 sysbench events/s on one thread against 913.87 on our Skylake CX23, about 3.5 times as much. In Ollama the CAX11 was ahead too, but by far less: 15 against 10.3 tokens/s on Llama 3.2 1B and 6.5 against 5.8 on Llama 3.2 3B. Tokens/s is how many word pieces the model writes per second.

In n8n the order flipped. Under ApacheBench with a webhook, Code node and response workflow, the CX23 reached 15.04 executions/s and the CAX11 13.9, with 0 failed requests on both. The x86 server also had faster random disk access (53,050 against 48,554 read IOPS, where IOPS means input/output operations per second). The ARM server pulled the Supabase images and started n8n faster.

Neither 4 GB server could run Llama 3.1 8B. The CAX11 failed with "insufficient memory (attempted to allocate 4402.48 MB)" and the CX23 was OOM-killed by the kernel after about 50 seconds.

Round 2: CX33 vs CAX21 (4 vCPU, 8 GB)

Here sysbench gave the CAX21 13,210.45 events/s on all threads, twice the CX33's 6,453.15. Yet the CX33 was faster on Llama 3.2 3B (13.5 against 12.2 tokens/s), on Llama 3.1 8B (6.3 against 5.7) and on n8n (31.93 against 25.77 executions/s; p95 at 50 concurrent requests 1,846 against 2,197 ms, where p95 is the time 95 % of requests beat). The CAX21 won the smallest model (26.8 against 23.2 tokens/s on Llama 3.2 1B), the disk IOPS tests and sequential reads.

A likely reason is that sysbench's CPU test is a tight prime-number loop that suits the Neoverse-N1 cores, while LLM inference leans heavily on memory bandwidth, where the EPYC-Rome CX33 measured 4,973.99 MiB/s against 4,263.84 on the CAX21. That is our reading of the numbers, not something we isolated in a separate test.

All five plans on one chart

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

The CX43 (8 vCPU, AMD EPYC-Rome) is shown for context. We did not test a CAX31, so we have no ARM result at that size.

App compatibility on ARM

ARM servers run arm64 software only. Most popular self-hosting images now ship both builds, but check before you commit. Here is what we verified for the three apps we test:

• n8n: the official n8nio/n8n image on Docker Hub lists linux/amd64 and linux/arm64 for the latest tag (checked 5 November 2026). Our compose with docker.n8n.io/n8nio/n8n:2.46.6 and postgres:16 ran on both CAX servers without changes. The n8n Docker docs do not mention ARM either way. More in n8n on ARM.
• Supabase: the self-hosting docs give 4 GB RAM and 2 cores as the minimum and do not mention ARM. In our test the official docker compose stack started all 11 containers with 0 unhealthy on both the CAX11 and CAX21. The arm64 images used 10.14GB of disk against 9.563GB on x86.
• Ollama: the Linux docs include an "ARM64 install" section with a dedicated package. The standard curl -fsSL https://ollama.com/install.sh | sh installer worked on both our CAX servers, in 14.9 s and 24.2 s. See Ollama on ARM for model choices.

Where ARM still bites: n8n community nodes or Code node packages with native x86 binaries, older single-architecture images, and anything you compiled yourself. Before you move, run docker manifest inspect on each image you use and look for an arm64 entry.

Full Hetzner price list

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.

Which one to pick

CX23 (x86, 2 vCPU, 4 GB)

Best for: n8n on a budget; 15.04 executions/s at best, the cheapest of the four. Main drawback: CPU lottery; ours got Skylake. Skip it if: you mainly run small local models.

CAX11 (ARM, 2 vCPU, 4 GB)

Best for: Llama 3.2 1B or 3B on 4 GB, and predictable CPU. Main drawback: 0.50 euro more and slower on n8n than our CX23. Skip it if: any image you need lacks an arm64 build.

CX33 (x86, 4 vCPU, 8 GB)

Best for: n8n plus Supabase or a 3B to 8B model; it won most app tests in its round. Main drawback: lower disk IOPS than the CAX21 (41,644 against 53,796 random reads). Skip it if: your workload is disk-bound and arm64-ready.

CAX21 (ARM, 4 vCPU, 8 GB)

Best for: database-heavy stacks that like fast random IO. Main drawback: 2 euro more than the CX33 and slower on n8n and larger models. Skip it if: local LLMs or n8n throughput are the point.

Next step: if Hetzner's CX or CAX plans do not fit your region or budget, compare the options in Hetzner alternatives.