Hetzner's 2026 Price Hikes and the 20TB-to-0.5TB Traffic Trap: Recomputing Self-Hosted Fleet Unit Economics

Hetzner raised cloud prices twice in 75 days this year — and the second hike tripled some plans while leaving your neighbor's identical server on the old rate. If your fleet's unit economics still assume 2025 Hetzner pricing, every capacity plan downstream of that assumption is wrong.

The short version, with the numbers up front:

• June 15 repriced new cloud orders hard. The dedicated-vCPU CCX13 (2 vCPU, 8 GB) went from €15.99 to €42.99 a month, a 169% jump. The AMD CPX line rose 2.4–2.75x in Germany and Finland, and up to ~3x in the US. The shared Intel CX and Ampere ARM CAX lines got off lightest at +30–40%: a CAX11 (2 Ampere cores, 4 GB) now lists at €5.99.
• Existing instances kept old prices — until touched. The June increase applies to new orders and rescales only, so resizing a grandfathered instance adopts the new rate.
• The traffic headline is regional. Cloud servers include 20 TB of outbound traffic in the EU, 1 TB in the US, and 0.5 TB in Singapore — with overage at €1.00/TB in the EU and US but €7.40/TB in Singapore. A €6 node pushing 5 TB/month out of Singapore pays more in overage than for the server itself.

The rest of this post recomputes the three inputs every self-hosted fleet budget needs now: the post-hike compute ladder, the Server Auction alternative, and the regional overage math — then states where "just add another Hetzner box" stops working.

What actually changed, and when​

Act one: April 1, everything, everyone. Announced February 23, the first increase hit new orders and existing products alike. Hetzner's statement blamed "drastic price increases in various areas in the IT sector" and said operating and hardware-buying costs had both "increased dramatically" (statement). Reporting put cloud servers at +30–37% in Germany and Finland depending on tier, dedicated hardware at roughly +3–21%, with load balancers and storage swept in too (Fudzilla). Concrete April moves: the CX23 went €2.99 to €3.99 and the CCX13 €11.99 to €15.99 (compiled April figures), and the LB11 load balancer €5.39 to €7.49 (docs price table).

The backdrop was the AI datacenter boom draining the DRAM and SSD supply Hetzner buys from. OVHcloud's founder warned within days that servers produced in 2026 could cost 15–35% more than a year earlier (365i). This was not a Hetzner margin grab; it was a hardware-cost pass-through the whole budget-hosting tier was about to feel.

Act two: June 15, cloud only, new orders and rescales only. Seventy-five days later Hetzner repriced cloud servers again — but this time existing instances kept their rate unless resized (Northflank breakdown). The shape of the increase was deliberately uneven:

Sources for the ladder: Hetzner's docs price table as compiled here, plus the post-June-15 plan comparison and SaaSCity's hike summary. The pattern is unmistakable: shared lines took a bad-but-absorbable third, while anything with dedicated or AMD silicon roughly doubled or tripled. The ARM line took the smallest increase in the catalog — which matters because it was already the cheapest compute per euro.

The post-hike compute ladder​

Here is the recomputed ladder for Germany/Finland list prices after June 15, reduced to the two unit costs a fleet planner actually multiplies: euros per vCPU and euros per GB of RAM.

Three things stand out. First, the cheapest way to buy a vCPU is now a mid-size ARM box: the CAX21/CAX31 at €2.62 per core undercut everything else on the menu. Second, the Intel CX23 slightly undercuts the equivalent ARM CAX11 at the very bottom (€5.49 vs €5.99) — "ARM took the smallest increase" does not mean ARM wins every row. Third, dedicated vCPU is now in a different universe: the CCX13 costs over 7x per core what the ARM shared line charges, so any workload that does not provably need pinned cores should justify why it is not on shared.

Two caveats before you standardize on CAX. The ARM line is EU-only (Nuremberg, Falkenstein, Helsinki) — there is no Ampere capacity in Ashburn, Hillsboro, or Singapore (server-type notes), so a multi-region fleet cannot buy its way out of the traffic asymmetry below with cheap ARM nodes. And the cheapest lines have had stock problems: the CX/CAX lines were reported stock-limited starting early September 2026, with CAX intermittently unavailable to new orders (fleet design notes). Budget on the €5.99 figure, but architect so a stockout degrades to CX rather than blocking provisioning.

The auction-box alternative​

Hetzner's Server Auction — refurbished dedicated boxes, usually with no setup fee — has always been the quiet alternative to cloud list prices, and the hikes widened the gap. A late-September snapshot of the cheapest EU auction rows (source; auction prices fluctuate daily):

A €38 auction box with 64 GB undercuts the cheapest cloud RAM by 60% per gigabyte — and the low end of the auction dips toward €30 for 32–64 GB configs when older i7 rows are in stock (example). For steady-state base load — a Kubernetes node pool that never scales to zero, a build cache, a database primary — one auction box replaces sixteen CAX11-equivalents of RAM for less than half the price.

The discount has honest strings attached. Auction CPUs are older generations (a 2015-era i7 against 2024 Ampere cores is not a fair fight per thread). There is no hourly billing and no cloud-API resize: capacity planning is monthly and manual, through Robot rather than the Cloud API your Cluster API provider speaks. And auction rows list EU datacenters (Falkenstein, Helsinki), so the auction does not help a latency story outside Europe either. The right mental model is base-plus-burst: auction metal for the flat bottom of your demand curve, cloud for everything above it.

The traffic trap, worked​

Now the part of the price list that breaks budgets. Hetzner's cloud traffic terms (hetzner.com/cloud):

Inbound and private-network traffic is free everywhere; only outbound to the internet counts. The EU allowance is genuinely generous — 20 TB covers nearly any small-fleet node. The trap is that nothing about the server's spec sheet warns you the allowance collapses 20x in the US and 40x in Singapore, while Singapore overage runs 7.4x the EU rate.

Hold compute constant at a ~€6/month small node (the class exists in every region on the x86 lines; regional list prices vary by a euro or two, which is noise next to what follows — the overage figures are exact) and vary egress:

Read the Singapore column twice. At 5 TB of egress — a single moderately trafficked app server, a container registry serving a busy CI, an image-heavy marketing site — overage (€33.30) exceeds the server's own price more than fivefold. At 25 TB, one small node burns €181.30 in overage: thirty servers' worth of compute, chargeable to one. The US column is gentler only because overage stays €1/TB; the 1 TB allowance still means a 25 TB node costs 5x its sticker price.

Two honest qualifications. First, this is a trap relative to Hetzner's own EU pricing, not to the market: hyperscaler egress near $90/TB makes even Singapore's €7.40/TB look like a rounding error (comparison). Hetzner remains one of the cheapest egress sellers on earth. Second, most app servers never push 5 TB; APIs, dashboards, and internal tools live comfortably inside every region's allowance. The workloads that detonate are exactly the ones founders forget are bandwidth-shaped: registries, CDNs-that-aren't-CDNs, video and image origins, backup and artifact stores.

Where "just add another box" breaks​

The recomputed rules for a Hetzner-backed fleet, in priority order:

1. Keep egress-shaped workloads in the EU, or behind a CDN. The 20 TB allowance plus €1 overage means an EU origin absorbs 25 TB for €11 total on a small node. The same node in Singapore costs €187. If users need Asian latency, serve them through a cache (Bunny, Cloudflare, or anycast object storage) so origin egress stays small — never by pointing a fat origin at Singapore and hoping.
2. Do not rescale grandfathered instances casually. Anything provisioned before June 15 keeps its old rate only while untouched; a rescale adopts current pricing. A CCX13 resize can turn a €15.99 instance into a €42.99 one overnight. Prefer creating new right-sized nodes and draining the old ones, so the repricing is a deliberate decision rather than a side effect.
3. Buy base load as auction metal, burst as cloud. Steady-state RAM at €0.59–0.67/GB (auction) vs €1.31–1.50/GB (cloud) is a structural ~60% saving that survives every hike so far. Reserve cloud for elasticity, fast provisioning, and regions auction hardware cannot reach.
4. Default EU compute to ARM, with an x86 fallback. CAX holds the best per-core price on the menu after taking the smallest increase — but it is EU-only and intermittently stock-limited, so fleet automation should degrade to CX rather than fail.
5. Meter egress per node before it meters you. The billing surprise is never the server line; it is the traffic line on a node whose workload quietly changed shape. Alert on outbound volume per host in every region, with the Singapore threshold set an order of magnitude lower.

None of this dethrones Hetzner as the default answer for owned-hardware economics — a CAX21 running a whole small PaaS for ~€11/month remains absurd value (see one team's real usage data). What died in 2026 is the idea that one price list describes the whole fleet. There are now at least three Hetzner pricings that matter — legacy grandfathered, current EU, and current US/Singapore — plus a fluctuating auction market underneath, and the "cheap box" reflex only survives if it checks which pricing it is standing in first.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.