Skip to main content

Hetzner's €5.99 Server Is a Number You Can't Buy: What the Missing Cheap Tier Does to the Self-Hosting Price Anchor

9 min readDora NodaDora Noda
Share
On this page

For years, every self-hosting cost comparison started from the same number: Hetzner's €5.99 cloud server. It was the industry's price anchor — the figure that made a VPS look nearly free and every managed platform look expensive by comparison. That anchor is now a number you cannot actually buy.

On September 7, 2026, developer Vincent Schmalbach checked Hetzner's public catalog and found every model in its Cost-Optimized tier marked unavailable: four x86 sizes and four Arm sizes, all eight of them. The labels were still there when he checked again on September 8. The advertised 4 GB server still costs €5.99 a month. The 4 GB option you can actually order, in the Regular Performance tier, costs €19.99. That is 3.34 times the monthly budget — after Hetzner already raised prices twice this year.

Here is the gap at three memory sizes, same-RAM comparisons from Schmalbach's September 9 catalog check (Germany and Finland, including IPv4, excluding VAT):

MemoryAdvertised (unavailable)Buyable (active Create button)Price multiple
4 GBCX23: €5.99CPX22: €19.993.34×
8 GBCX33: €8.99CPX32: €35.994.00×
16 GBCX43: €16.49CPX42: €69.994.24×

Each buyable plan matches the vCPU count, doubles the storage, and runs newer processor generations — but both tiers still share CPU resources, so this is the cost of keeping the same memory, not of buying equal performance. And note the direction of the multiples: the more RAM your workload needs, the further the buyable price has drifted from the advertised one. This post works through how we got here, why the cheap tier is empty, and how to re-anchor your cost math to the catalog you can actually order from.

Two hikes got us here​

The tier split itself dates to October 16, 2025, when Hetzner divided its cloud lineup into Cost-Optimized (CX/CAX, older proven hardware) and Regular Performance (CPX/CCX, newer generations). Then came two rounds of increases in 2026:

April 1: the broad adjustment. Hetzner raised prices across cloud servers, storage, and dedicated servers, citing surging hardware costs — several products up around 30%, with cloud tiers rising roughly 30–37%. Critically, this round hit existing subscriptions as well as new orders. Nobody was grandfathered.

June 15: the cloud repricing. Cloud servers went up again, this time for new orders and instance rescales. The increases were uneven by family: CX/CAX rose a further 30–40%, while the dedicated-vCPU and shared-AMD performance lines took the brunt — CPX22 jumped from €7.99 to €19.49 a month (about 2.44×, or 144% more), and CCX13 went from €15.99 to €42.99 (about 2.7×, or 169% more). Some CCX tiers rose as much as 176%. The Hacker News thread on the June adjustment hit 398 points within hours — the kind of velocity Hetzner news only gets when the number on the invoice actually moves.

To keep the two comparisons straight: the June hike compares the same plan before and after (CPX22 at €7.99, then €19.49). Today's 3.34× gap compares two different plans — the €5.99 CX23 you cannot order against the €19.99 CPX22 you can. There has been no new September price announcement. The cheap plans are simply marked unavailable after those earlier hikes, and the shortage acts like a third increase for anyone deploying today.

Why the cheap tier is empty: RAMpocalypse plus a capacity notice​

The proximate cause is the 2026 memory crisis. TrendForce data shows server DRAM contract prices rising 90–95% in the first quarter of 2026 alone, with enterprise NAND flash up 55–60% and some high-end DDR5 chips up over 300% within six months. AI datacenters are on track to consume around 70% of all high-end DRAM production this year, buying up supply years in advance with deeper pockets than any budget cloud can match. When DRAM dominates server cost — over half the total at Azure and Meta scale — a near-doubling of memory prices flows straight into the price of a RAM-sized cloud instance. Analysts do not expect normalization before 2028.

Hetzner has also been explicit that this is a capacity problem, not just a pricing decision. A capacity notice open since June 26 cites high demand and limited hardware components, stating that restrictions affect new customers and some randomly selected existing customers while capacity is expanded. Hetzner frames the Cost-Optimized tier as "proven hardware generations" — older machines that cannot be expanded at will when components are scarce — and calls the unavailability a capacity limit, not a discontinuation.

Put together, the mechanism is straightforward: the cheapest tier runs on the oldest hardware, the oldest hardware is the hardest to expand during a component shortage, and a provider managing scarcity rations the tier it can least afford to grow. The €5.99 price stays on the page because the plan still exists for grandfathered rentals. It just is not something a new deployment can buy.

Capacity crunch or quiet retirement?​

It is worth asking the skeptical question: is "unavailable" a temporary capacity squeeze, or the quiet retirement of a tier Hetzner no longer wants to sell? The evidence points both ways, so here it is honestly.

On the capacity-crunch side: the June 26 notice is explicit about high demand and limited components, says capacity is being expanded, and restricts only new customers plus a random subset of existing ones — an odd shape for a deliberate kill. Unchanged existing rentals keep their terms, which is consistent with "we cannot provision more of these right now" rather than "we are winding this down."

On the quiet-retirement side: the tier split dates to October 2025, the cheap tier's whole identity is older hardware generations that only get older, and there is no published restock date or recovery timeline. A provider that wanted to sunset its cheapest plans without the backlash of a third price announcement could hardly design a quieter mechanism than indefinite unavailability.

The best-supported reading is capacity pressure — Hetzner's own statements say so, and the component shortage is independently well documented. But here is the uncomfortable part: for budgeting purposes, the distinction barely matters. A €5.99 server you cannot order constrains your architecture exactly as much as a €5.99 server that no longer exists. Until the Create buttons come back, every cost model built on the Cost-Optimized price list is fiction, whatever the cause.

Re-anchor your math to the buyable catalog​

The practical rule is simple: compare alternatives against the €19.99 CPX22 you can order, not the €5.99 CX23 you cannot. Schmalbach puts it well — for his next 4 GB deployment, he will benchmark other providers against Hetzner's buyable option rather than budget for a server he may not be able to get. Two corollaries follow immediately. First, never delete or resize a grandfathered cheap instance without checking: under the June terms, unchanged rentals keep their old prices, but a rescale moves you to the new ones — deleting a €3.99-era CX23 to "recreate it cleanly" is a one-way trip to €19.99. Second, treat any comparison, blog post, or vendor calculator still quoting Cost-Optimized prices as stale until it shows a working order flow.

To make the re-anchoring concrete, here are two typical small stacks, advertised versus buyable:

Small stack — 1× 4 GB app server + 1× 4 GB database server. Advertised: 2 × €5.99 = €11.98/month. Buyable: 2 × €19.99 = €39.98/month. Delta: +€28.00/month (+234%). The classic "two small boxes" self-hosted setup costs nearly three and a half times its brochure price.

Larger stack — 1× 8 GB app server + 1× 4 GB database server. Advertised: €8.99 + €5.99 = €14.98/month. Buyable: €35.99 + €19.99 = €55.98/month. Delta: +€41.00/month (+273%). Because the multiple grows with RAM (3.34× at 4 GB, 4.00× at 8 GB, 4.24× at 16 GB), RAM-heavy stacks — databases, caches, anything memory-sized — drift the furthest from their advertised budgets.

This is also where a self-hosted platform has to reprice its own math, and the mechanism deserves spelling out rather than hand-waving. Bex.co runs tenant apps on owned Hetzner machines with flat-rate pricing: push a git repo, get a running HTTPS service, one predictable price. That flat rate is underwritten by hardware economics — and if the underwriting assumes €5.99-class machines while provisioning actually happens at €19.99-class prices, the platform eats the 3.34× gap on every deployment.

There is no clever scheduling that recovers a 3.34× input-cost miss; either the flat rate is repriced against buyable tiers, or the margin silently goes negative. The repriced rule for any PaaS cost story in this market: quote the tier with a working Create button, show the unavailable tier only as the reason your numbers moved, and re-check the catalog on a schedule — because the next drift will be another quiet label change, not another press release.

The anchor moved​

Hetzner is still cheap relative to the hyperscalers — a buyable 2 vCPU / 4 GB box at €19.99 remains a fraction of its AWS or GCP equivalent, with 20 TB of included transfer that the metered clouds cannot touch. The price advantage narrowed; it did not disappear. What disappeared is the specific number that anchored a thousand "self-hosting costs almost nothing" calculations. Until DRAM normalizes — and analysts say not before 2028 — treat every sub-€10 server price as guilty until proven orderable.

The broader lesson outlasts this shortage. A price list is a marketing artifact; the buyable catalog is the infrastructure reality. Teams that budget from the former get a 3.34× surprise at deploy time. Teams that track the latter — checking availability the way they check prices, pinning cost models to orderable tiers — never have to explain the gap. The Create button is part of the price. Quote accordingly.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, at a flat rate underwritten by hardware you can actually provision. Star the repo on GitHub or deploy your first app today.

Related articles

Run this on infrastructure you own

bex is the open-source, AI-native Render alternative — push a git repo and get a running HTTPS service on your own machines.

Get started with bex