Arm64 Hosting: Cheaper Than CPX, Dearer Than CX

An earlier version of this article said that after Hetzner's 15 June 2026 price change, Arm cost roughly a third of x86 per vCPU on the same provider. The arithmetic was right. The comparison was not fair. It set Hetzner's Arm line, CAX, against its regular-performance x86 line, CPX, and skipped the cheaper x86 line that Hetzner sells at exactly the same sizes.

Correction: the CAX against CPX gap is real, but CPX is the wrong yardstick for most small workloads. The table below uses the like-for-like x86 line, CX.

That matters because the story we told, arm64 hosting quietly winning on price, is the sort of claim that gets repeated. If you are choosing a server this month, you should be choosing on the right pair.

Hetzner sells four sizes where the cost-optimised Arm servers and the cost-optimised x86 servers have identical vCPU, RAM and disk. Prices are per month excluding VAT for Germany and Finland, from Hetzner's price adjustment page. The specs come from its cost-optimised plans page.

On the cheaper x86 line, arm64 hosting is not a saving at any size. It is a premium that grows with the server. Both lines rose by about the same proportion in June, roughly 30 to 38 percent, so the June change did not create an Arm advantage here. It stayed a small disadvantage.

Set the same Arm servers against CPX and the original claim reappears. CPX, Hetzner's regular-performance line, went up by 144 to 175 percent in the same June update, for example from €7.99 to €19.49 for the two-vCPU size.

So the arm64 hosting gap widened because one x86 line was repriced steeply, not because Arm got better. Before June, CAX11 at €4.49 was already cheaper than CPX22 at €7.99, but by about 44 percent, not 70. CPX also carries double the disk at each size, which the table shows and a per-vCPU figure hides. Hetzner has not published a reason for the uneven sizes. Its February statement, ahead of the 1 April rise, says only that the costs to operate its infrastructure and to buy new hardware have both increased dramatically.

At the large clouds, the case for arm64 hosting is stronger and the numbers are smaller. AWS states its own figure on the Graviton page.

Note the two hedges, up to and comparable. This is a vendor claim. So are the case studies in Arm's January 2026 blog post: one startup moved from $2,000 to $1,300 a month on Graviton3, which is a 35 percent cut; another reported about 25 percent lower monthly costs, and a third about 20 percent lower compute cost. Those are chosen examples from the company that sells the architecture. The same post says close to half of new server deployments at top hyperscalers are set to be Arm-based, which is a forecast, not a count.

Our reading is that arm64 hosting saving 20 to 25 percent is credible where a provider prices Arm below x86. A three-fold saving is a quirk of one price list. AWS also reports more than 120,000 customers using Graviton, which says the move is safe more than it says it is a bargain.

Almost certainly, and that is the good news for arm64 hosting. Docker's official Node.js and PostgreSQL images both list arm64v8 among their supported architectures, so a typical Next.js app with a Postgres database has no missing base image. The risk sits in what you add on top: native addons compiled at install time, and any tool that downloads an amd64-only binary in a build step.

If you build images yourself, build both architectures once and let the host pick. Docker's multi-platform documentation gives the command.

docker buildx build --platform linux/amd64,linux/arm64 .

Building an image on your laptop for the other architecture is a place people lose an afternoon, so our guide to dockerising a Next.js app covers the multi-arch route in more detail. Run your own test suite on the Arm server before you move anything real.

Fewer places than x86. In Hetzner's price adjustment tables, CAX appears under Germany and Finland only. The US and Singapore lists carry CPX and CCX, and no CAX. So if your users, your compliance position or your latency budget needs a US region, the arm64 hosting discount is not on the menu at all.

For a European team the constraint is friendlier. Germany and Finland are where many EU teams want their data anyway, as EU data residency explains.

Worked example, using the table above. A Next.js app and Postgres on 4 vCPU and 8 GB. On CX33 it costs €8.49 a month. On CAX21 it costs €10.49. Arm only wins on price if it does more work per vCPU, and the break-even is 10.49 divided by 8.49, which is 1.24. Arm needs to deliver about 24 percent more throughput per core than that x86 server before the extra €2 a month pays back.

• Against CPX32 at €35.49, CAX21 saves €25 a month, or €300 a year, for the same vCPU and RAM but half the disk.
• Against CX33, CAX21 costs €24 a year more. Move only if your benchmark shows the throughput gain above.
• Measure your own workload. A vCPU is a billing unit, not a performance unit, and we have not found a published benchmark that settles it for a web app.

1. 1Build for bothRun docker buildx build with --platform linux/amd64,linux/arm64 so one tag serves both servers, and the only thing that changes is the hardware.
2. 2Rent an hour of eachHetzner bills by the hour, so trying a CAX21 and a CX33 side by side costs pennies. Put the same image and the same database size on both.
3. 3Replay real trafficUse a load tool such as k6 or autocannon against your slowest real page, and note requests per second at the same 95th-percentile latency. Ignore the peak number; it flatters everyone.
4. 4Do the divisionIf Arm serves 1.24 times the requests or more at the same latency, it wins on price. If not, the x86 box is cheaper for you, whatever the headline says.

If you would rather not do this arithmetic on a server bill at all, that is a fair reason to look at a platform. Our Hetzner Cloud pricing explained covers what each price includes, and our pricing is a minimum monthly spend with usage drawn against it.