Hetzner Nearly Tripled Bare-Metal Prices: What the AX102's €124-to-€454 Jump Means for Self-Hosting Math

On June 15, 2026, at 8 AM CEST, the cheapest line in half the internet's infrastructure budgets tripled. Hetzner's AX102 — a 16-core Ryzen 9 7950X3D dedicated box that rented for €124 a month — repriced to €454 for every new order. The 256 GB AX162 went from €244 to €844. Some US cloud plans rose 107 to 204 percent, with the 16-core CPX51 jumping from $92.49 to $279.49 a month. Hacker News noticed within hours, and the reactions ranged from disbelief on HN ("25-50%, sure… but 3x is wild") to gallows humor on X ("My Hetzner VPSes are now appreciating assets").

Here is the bottom line up front, because this post exists to recompute it honestly: if your fleet is already rented, nothing changed — existing contracts keep legacy pricing. If you are pricing new capacity, the old "bare metal is 4–5x cheaper" rule of thumb is dead, but the new math still favors owned hardware for anything with real bandwidth. On a typical box-plus-egress workload, the multiple compressed from roughly 15x to roughly 5x. On raw compute with no traffic, it compressed to near parity. Everything below is the worked evidence for those two sentences.

The recompute: what the hike does to real workloads​

Take a concrete, typical workload: a 16-core box with 32 GB of RAM and roughly 640 GB of NVMe, pushing 20 TB of egress a month. That is a mid-size API tier, a busy media upload path, or a small multi-tenant node — not a cherry-picked monster.

Currencies are mixed on purpose — providers bill that way — but at any plausible EUR/USD rate the story is the same. Before June, the Hetzner box cost about one-fifteenth of the AWS equivalent for this workload. After June, it costs about one-fifth. The multiple got cut by two-thirds, and the answer is still "self-host by a mile," because bandwidth was always the larger half of the cloud bill, and the hike did not touch the bandwidth story at all: Hetzner dedicated servers still include effectively unlimited traffic on the standard 1 Gbit/s port, and cloud instances still include 20 TB a month with overage around €1/TB.

Now the sensitivity case the honest version of this post has to include: strip the bandwidth out. A compute-only workload — batch jobs, internal tooling, anything that barely touches the network — compares ~€454 against ~$550 of EC2. That is roughly parity, and once you add the operational cost of running your own box, the cloud may genuinely win that narrow shape. The repricing did not kill the self-hosting case; it narrowed it to workloads where bandwidth, storage, or steady-state utilization dominate — which, to be fair, is most of what small teams actually run.

And the third sensitivity case matters most of all: grandfathered fleets. The June adjustment applies to new orders and cloud rescales. If your boxes were ordered before June 15, you still pay €124, and every pre-June TCO model you built is still exactly right — until you rescale, replace, or expand. That splits the world into two populations with different math: incumbents sitting on appreciating assets, and newcomers pricing from the new list.

What changed, what didn't, and the year's price history​

Precision matters here, because the backlash thread conflated three separate adjustments. The June 15 change, per Hetzner's own price-adjustment notice:

• Affected: all dedicated servers and cloud plans at all locations, for new orders and cloud instance rescales, effective 8 AM CEST. Orders placed before June 15 but delivered after still get the old price.
• Not affected: the Server Auction (Dutch-style, demand-priced), web hosting products, managed servers, IPs, storage products, load balancers, volumes, and snapshots, per the notice text.
• The trapdoor: certain changes to legacy-priced cloud servers — rescales, in particular — switch the instance to current pricing. Touch a grandfathered box and it reprices.

This was also not the first hike of the year — it was arguably the third. Hetzner announced an April 1 round back in February (flagged as "drastic price increases in various areas in the IT sector"): cloud servers up 30 to 43 percent, object storage up 30 to 53 percent, and memory add-ons up by an eye-watering roughly 575 percent. It flagged the June round separately on May 27, blaming supply-chain tightness. Anyone modeling Hetzner costs off a 2025 blog post is therefore two repricings stale, not one.

Why it happened: the AI memory crunch is real​

It is tempting to read a 3x hike as opportunism, but the component market behind it is genuinely on fire — and the fire has one arsonist: AI infrastructure buildouts. Three numbers tell the story:

1. Conventional DRAM contract prices rose about 55–60 percent quarter-over-quarter in Q1 2026, per TrendForce, with a further 58–63 percent projected for Q2 — because suppliers keep reallocating advanced process nodes and new capacity toward server DDR5 and HBM.
2. HBM production consumes roughly three times the wafer capacity of standard DRAM per gigabyte, per a Micron executive — every wafer feeding an AI accelerator is three wafers not feeding a server DIMM.
3. SSD prices were forecast to rise over 40 percent on the same reallocation dynamic.

Samsung publicly warned of memory shortages driving industry-wide price surges through 2026. Hetzner's stated reason — "extremely high procurement costs for new hardware" — matches what every other buyer of server DIMMs is reporting. The April memory-add-on spike (~575 percent) is the tell: when RAM is the component repricing fastest, a provider whose value proposition is "cheap boxes stuffed with RAM" has nowhere to hide. This is not a Hetzner story; OVHcloud pushed through 9–11 percent increases on the same DRAM/flash pressure, and the whole budget-hosting tier is downstream of the same fabs.

A step function you can plan around beats a meter you can't​

Here is the part the backlash thread mostly missed, and the part that matters for a self-hosting strategy: a 3x step-function hike on new orders is a categorically different risk than a usage meter. Compare the two failure modes:

• Hetzner reprices: your existing fleet costs exactly what it cost yesterday. Your next server costs more, and you learn that from a price list before you order it. You can respond by keeping old boxes longer, shopping the Server Auction, buying fewer bigger boxes, or deferring expansion. The shock is bounded, announced, and grandfathered.
• A usage meter reprices you: nothing announces it. A deploy ships a chatty service, a customer scrapes an endpoint, a retry storm loops overnight — and the bill arrives after the fact. Every deploy is a small unpriced option on next month's invoice. AWS has never needed to triple list prices to triple someone's bill; traffic and API-call growth do it silently.

One risk is plannable — read the list, model the step, decide. The other is ambient — it compounds while you sleep. Supply-chain-driven hardware inflation even has a mean-reverting property meters lack: DRAM is a notoriously cyclical commodity, and Hetzner's own notice gestures at cheaper "limited" offerings returning "when we are able to source hardware components at a lower cost." Nobody has ever announced that your per-GB egress fee will mean-revert.

That is not a defense of the hike. It is the reason the hike, however painful, does not invalidate the architectural bet on owned hardware: fixed-price capacity with step repricing is still the only cost model a small team can capacity-plan against.

What to do Monday morning​

Four concrete moves, in priority order:

1. Freeze legacy-priced cloud instances. Audit which instances still carry pre-June pricing and treat rescaling one as a purchase decision at 3x — because that is literally what the trapdoor clause makes it. Scale out with new instances where the math works rather than resizing old ones reflexively.
2. Shop the Server Auction first for new dedicated capacity. Auction pricing is demand-driven and was explicitly left out of the adjustment. Older generations at auction prices are the closest thing left to the pre-June list.
3. Re-baseline every TCO model to the post-June list. Any "migrate off AWS, save 10x" spreadsheet computed on 2025 Hetzner prices now overstates savings by roughly 3x on the compute line. Re-run the numbers with the bandwidth line doing what it has always done — the case usually still closes, but quote the new multiple, not the old one.
4. Bin-pack harder. At €454 a box, the cost of an idle server tripled too. If you run Kubernetes on dedicated hardware, this is the year to take bin-packing, autoscaling, and node utilization seriously rather than treating cheap boxes as disposable. The providers' margin moved; your scheduler is where you get it back.

The through-line: nothing about June 15 changed how to self-host cheaply. It changed the constant factor, and it rewarded the teams that were already grandfathered, auction-literate, and good at packing.

The floor moved, not the ceiling​

Zoom out and the 2026 story is almost symmetrical: AI demand made DRAM expensive, which made cheap servers expensive, which compressed — but did not close — the gap between owned hardware and the meter. The teams hurt worst are the ones pricing their first fleet today against blog posts from the €124 era. The teams hurt least are incumbents whose contracts grandfathered them into 2025 economics.

If there is a forward-looking lesson, it is that the "owned hardware has no meter" slogan was always shorthand, not physics. Owned hardware has no usage meter. It has a procurement cycle, a refresh cadence, and — as June proved — step repricing when the fabs pivot to feeding accelerators. Plan for steps, lock prices when they are low, keep the auction bookmarked, and keep modeling bandwidth honestly. Do that and the next repricing notice reads as a planning input, not a betrayal.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. When the hardware floor moves, a declarative fleet that bin-packs tenants onto every box is how you absorb the step instead of eating it. Star the repo on GitHub or deploy your first app today.