Hetzner Raised Prices Three Times in Six Months: The "Just Use a Cheap Hetzner Box" Math, Recomputed

The cheapest number in self-hosting economics just moved — for the third time this year. On June 15, 2026, Hetzner repriced its entire cloud and dedicated lineup for new orders and instance rescales, and some families didn't creep up, they jumped: the dedicated-vCPU CCX13 went up 169% per hour, and dedicated AX servers now list at 3–4x their old prices. This blog has spent months citing Hetzner's price list as the concrete floor under "self-hosting is cheaper" — a €3.99 box here, a €7.99 dedicated-vCPU instance there. Those numbers are now wrong, so here is the recompute, up front:

The headline, honestly stated: the self-hosting gap survives the hike everywhere — even the worst-hit mainstream tier is still roughly a quarter of the managed price — but it did not survive evenly. If your fleet lives on the shared-vCPU CX line, you absorbed a 33–38% bump. If it lives on CPX, CCX, or bare metal, your new-order prices went up 2.3–3.7x, and the durability of "own the hardware, it's cheaper" now depends on grandfathering rules and how your automation orders servers. The rest of this post walks the numbers, the trap, and what a claim like "it's cheaper" has to mean when the supplier underneath it reprices three times in six months.

Three Hikes in Six Months: What Actually Happened​

The June 15 adjustment wasn't a one-off. It was the third repricing event of 2026:

1. Early February — Hetzner raised one-time setup fees on dedicated servers, a move management acknowledged (per heise's reporting) didn't fully cover its rising costs.
2. April 1 (announced February 23) — an across-the-board increase of roughly 30–40% on cloud and dedicated products. The stated causes: drastically higher hardware procurement costs, a RAM market "dominated by few manufacturers" with pricing Hetzner called hardly comprehensible, sharp NVMe SSD increases, and unreliable supplier quotas. Unlike the 2022 energy-cost adjustment, this one came with no exceptions for existing customers on long-term contracts. The announcement hit the Hacker News front page at 553 points.
3. June 15, 8 AM CEST — a re-tier of the whole lineup for new orders and cloud instance rescales, documented on Hetzner's own price-adjustment page. Existing orders — including orders placed before the cutoff but delivered after — keep their prior pricing.

The June round is where the families diverged. Real before/after numbers, from Hetzner's documentation and community threads:

Notice the pattern: the increase scales with how much dedicated silicon and RAM the product hands you. Shared-vCPU plans, which oversubscribe cores across tenants, took a moderate hit. Dedicated-vCPU and bare-metal lines — where Hetzner's own RAM and CPU procurement costs pass through directly — took the 2.5x-and-up hits the HN thread (273 points) was built around. That's consistent with the stated cause being component costs, and it matters for the recompute: which line your stack sits on determines which hike you got.

One rule from the fine print worth bolding: a rescale is a repricing event. Resize a grandfathered instance and it moves to the current price list. Keep that in mind for the fleet section below.

The Recompute: Same Stacks, New Price List​

This blog's earlier cost posts priced a canonical indie stack — one web service, one background worker, one Postgres — against managed platforms using pre-hike Hetzner numbers: a CX22 at €3.79, a CPX22 at €7.99. Here are those same comparisons on the June 15 list, at roughly current exchange rates (€1 ≈ $1.17).

Scenario A: the hobby stack (vs. Railway)​

All three containers on one shared-vCPU CX23 (2 vCPU, 4 GB RAM, 40 GB NVMe, 20 TB traffic included):

The hike costs a hobby self-hoster €18 a year. The gap is still roughly 4x, because the gap was never mostly about hardware — it's the difference between flat pricing and per-resource metering. Verdict: the pitch survives untouched here.

Scenario B: the production stack on shared vCPU (vs. Render Pro)​

Our earlier Render analysis priced the same stack on Render Pro at 1 TB of monthly egress: $25 web + $25 worker + $20 Postgres + $25 workspace fee + $146.25 in metered egress = $241.25/mo. Against a shared-vCPU CX43 (8 vCPU / 16 GB, 20 TB traffic included):

A 33% supplier hike moved the monthly delta by about five dollars. When the managed bill is dominated by metered egress and platform fees, the hardware line is almost noise.

Scenario C: the production stack on dedicated vCPU — the tier that actually got hit​

Stopping at Scenario B would be cherry-picking, because plenty of production Hetzner users don't run Postgres on shared, oversubscribed cores — they pay for dedicated vCPU precisely to avoid noisy neighbors. That's the tier the hike hammered. Same stack on a CPX22 or a CCX13:

This is the honest center of the recompute. The worst-hit mainstream cloud tier went from eleven times cheaper than the managed equivalent to four times cheaper — a real compression, in one repricing event, on the family a production reader is most likely to be standing on. The pitch survives — $58 versus $241 is not a close call — but "self-hosting is 90% cheaper" quietly became "self-hosting is 75% cheaper" for this tier, and nobody's old blog post got updated on hike day. (This one now has.)

Scenario D: the bare-metal fleet — where budgets actually break​

A three-node Kubernetes fleet on AX102 dedicated servers — a popular shape for a self-managed bare-metal cluster — used to budget at 3 × €124 = €372/mo. New orders now price at 3 × €454 = €1,362/mo (≈$1,600). Nothing about the managed comparison changes the sting here: if your capacity plan was written in May, your expansion cost tripled in June, even though every node you already own still bills at the old rate.

The Grandfathering Trap: Cattle Pay New Prices​

The June 15 rules read like protection for existing users: everything you already ordered keeps its price. And for a pet server — the long-lived box you SSH into and never touch — it genuinely is. Your €124 AX102 stays a €124 AX102.

But modern self-hosted platforms are built on the opposite philosophy. A Cluster API-style fleet — the architecture underneath bex and any Kubernetes platform using Cluster API Provider Hetzner — treats servers as cattle: nodes are provisioned declaratively, replaced on failure, rolled during upgrades, and added by autoscalers. Every one of those operations is, from Hetzner's billing perspective, a new order at the current price list. And resizing an existing instance instead doesn't dodge it — a rescale reprices too.

Follow that to its uncomfortable conclusion:

• MachineHealthCheck remediation replaces a broken node → the replacement pays June 15 prices.
• Immutable node rollovers (the standard way to upgrade Kubernetes or the OS image) replace every node in the fleet → the whole fleet reprices, one node at a time.
• Autoscaling up on a traffic spike → new-order prices, even if you scale back down.

The better your automation, the faster you forfeit grandfathering. A hands-off, self-healing cluster on the AX line converges toward the 3.7x price over one or two upgrade cycles, while the neglected pet box next to it keeps its 2025 pricing indefinitely. That's a genuinely perverse incentive, and it's worth naming because most "existing customers are unaffected" summaries of this hike miss it entirely.

Practical mitigations, in rough order of usefulness:

1. Know which line you're on. The CX shared-vCPU family took +33–38%, not +169%. For many web workloads it's the value pick post-hike — don't pay CCX prices out of habit for a stack that doesn't need pinned cores.
2. Keep a grandfathered baseline. Let long-lived control-plane or baseline-capacity nodes age in place where operationally sane, and pay new prices only for elastic capacity.
3. Don't rescale grandfathered instances. A rescale reprices; adding a separate node often doesn't reprice what you already have.
4. Make the fleet portable. This is the real hedge, and it's structural: Cluster API's provider model means the same declarative cluster spec can target Hetzner today and another provider — or your own colo metal — tomorrow. Repricing risk you can walk away from is a negotiation; repricing risk you can't is a tax.

Is "Own the Hardware, It's Cheaper" a Claim or a Snapshot?​

Three repricings in six months is a fair reason to ask whether the self-hosting pitch was ever a durable claim or just a screenshot of a price list. The recompute suggests a precise answer.

As a snapshot, it degraded. Every number this blog cited before June 15 is stale, and on the dedicated-vCPU tier the advantage compressed from ~11x to ~4x in one event. Anyone who repeated "a €3.79 Hetzner box" after hike day was quoting fiction.

As a structural claim, it held — for identifiable reasons. Run the sensitivity the other direction: for the managed bill to meet the self-hosted one at 1 TB of egress, the CX43 would need to cost about 12x its post-hike price, and even the freshly-2.7x'd CCX13 would need another ~4x. The gap isn't hardware margin — it's the structure of the bill: metered egress ($146 of Render's $241 at 1 TB, versus 20 TB included on the Hetzner box), platform fees, and per-service pricing. Component inflation moves both sides' costs; it doesn't change the shape.

And the other side is repricing too. The same six months saw Render cut included bandwidth to 25 GB on Pro and meter overage at $0.15/GB, and OVH raise prices roughly 30%. RAM that costs Hetzner more also costs every PaaS's cloud provider more; managed platforms just deliver the increase with a lag and a different label. Pinning your architecture to any provider's January price list — managed or not — is the actual mistake.

So state the durable version of the pitch precisely: it was never "Hetzner is cheap." It's flat, meter-free pricing on capacity you control, from a supplier you can replace. The June 15 hike stress-tested all three clauses. Flat pricing passed (the CX23 is still one number with 20 TB of egress in it). Capacity-you-control passed with a bruise (grandfathering rewards exactly the wrong operational style). Supplier-you-can-replace is the clause doing the most work now — and it's the one that requires your platform layer, not your hosting provider, to be the thing you're committed to.

Watch that last clause over the next year. A provider that reprices three times in six months while citing a RAM market it can't predict will likely reprice again — and the teams that shrug will be the ones whose deployment layer treats the provider as a slot, not a foundation.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, on a Cluster API fleet that can outlive any single provider's price list. Star the repo on GitHub or deploy your first app today.

Sources​

• Hetzner Docs — Price Adjustment (June 15, 2026) — official before/after prices for CX, CPX, CCX, CAX cloud lines and AX/EX/SX/GEX/DX dedicated lines; new-order and rescale rules
• heise — "Hetzner increases prices from April 1" — the April 1 across-the-board increase, stated component-cost causes, no existing-customer exceptions
• Hacker News — "Hetzner Prices increase 30–40%" — 553-point thread on the February 23 announcement
• Hacker News — "Hetzner increased dedicated server prices 3–4x" — AX102 €124→€454, AX162 €244→€844, and the June 15 discussion at 48540844
• Render — Pricing — the managed-side bill inputs (Pro workspace fee, Standard instance pricing, and the $0.15/GB egress meter above 25 GB)