Hetzner Raised Prices Three Times in 2026 — The RAM Shortage Behind Your Next Bill

Hetzner cloud servers cost 30 to 37 percent more than they did on March 31. If you run dedicated vCPU, they cost 113 to 175 percent more. And if you picked the wrong instance family for your node pool, the same workload now costs almost three times what your 2025 spreadsheet assumed — before you changed a single line of application code.

That is not a pricing experiment. It is a memory supply chain passing through to your invoice, and it is the same AI boom everyone celebrates on the demand side showing up on the cost side as a hardware tax.

Here is the answer up front for the team that built its self-hosting math on "cheap Hetzner boxes under Cluster API":

Concrete DE/FI examples from Hetzner's own tables:

In the US region the same families went further — CPX41 38.99 to 120.49 EUR (+209%) and CCX13 18.49 to 50.49 EUR in the US price book. The worst line, CPX52, is up 175 percent: 2.75 times the day before. The headline "Hetzner raised prices 30 percent" describes April. June doubled down on exactly the machines most production fleets actually run.

The rest of this post traces why DRAM did this, what it does to a Cluster API fleet's unit economics when the node pool sits on the wrong family, and why every self-hosting cost comparison you publish in late 2026 should be a trend line, not a frozen number.

Three hikes in ten weeks — and why the third hit differently​

Hetzner's own language across the three notices is unusually consistent. April 1: "drastic increase in hardware procurement costs." April 29: "exceptionally high purchase prices for hardware components." June 15: "massive increase in procurement costs in the hardware market." The docs page for the June adjustment is explicit about scope: new orders and rescales only, existing instances keep their price — a new grandfathering rule that did not exist in April.

The severity map tells you what got expensive. The two families whose price is most sensitive to RAM and high-end silicon — dedicated vCPU (CCX) and AMD shared (CPX) — moved in lockstep at more than double. The families with lighter memory footprints or cheaper cores — ARM (CAX) and Intel shared (CX) — moved roughly a third. Hetzner also used June to standardize dedicated servers into fixed -1/-2/-3 SKUs and introduced a -1-Ltd tier built from lower-cost hardware, which is the usual signal that the old catalog had too many configurations to reprice cleanly.

Why this matters: the node pool is the bill. If your Cluster API MachineDeployment defaults to CCX13 or CPX22 — the sensible picks in 2024-2025 — five nodes that listed at 62 EUR/month on March 31 now list at ~149 EUR if you rescale. The same shape on CAX moves from ~18 EUR to 24 EUR. A 2.4x fork driven by one SKU name.

For context, PrivateDevOps's June 15 analysis called this Hetzner's fourth pricing action of 2026 if you count February setup fees, and flagged the unevenness as the real story: "the lines that lean on the most expensive resources took by far the largest correction." Heise and Tom's Hardware reported the April round at 30 to 37 percent in DE/FI with US servers higher, which matches Hetzner's tables.

Why RAM did this: HBM is eating DDR5's factory​

Server DRAM contract prices rose 43 to 48 percent in Q4 2025 alone, according to the procurement notices Hetzner cited, and forecasts at the time pointed to another 58 to 63 percent in Q2 2026. TrendForce revised Q1 2026 PC DRAM contract prices to a 105 to 110 percent quarter-over-quarter increase. By May 2026 one industry tracker had DDR5 RAM up roughly 110 percent year to date, with spot DDR5 trading around $1.50 per gigabit — below older DDR4 at $2.10 per gigabit, but above HBM3e at $1.70 per gigabit in a brief spot inversion that tells you how violently the mix shifted.

HBM — the stacked memory next to AI accelerators — is built on the same wafers as the DDR5 in your Hetzner node, but consumes roughly three times the wafer capacity per gigabyte. Every wafer Samsung, SK Hynix, or Micron allocates to HBM is three times as much DDR5 that never gets made. As AI demand surged in H2 2025, fabs prioritized the higher-margin stacked part. Textbook crowding-out.

Three numbers pin down the scale:

• 20 percent of global DRAM wafer capacity for AI in 2026. Industry estimates put AI's share of wafer consumption above one-fifth this year, up from single digits in 2023.
• HBM profit parity flipped. TrendForce notes that HBM3e prices were four to five times server DDR5 in early 2025 and expects that gap to compress to one to two times by the end of 2026 as DDR5 profitability catches up — which, paradoxically, lets suppliers push HBM contract prices higher still because the alternative is also expensive. Suppliers are now shifting capacity back toward DDR5 at the margin, but not fast enough to clear the shortage.
• The Stargate letter of intent. In October 2025 OpenAI signed simultaneous non-binding letters of intent with Samsung and SK Hynix for up to 900,000 DRAM wafers per month for the Stargate infrastructure project. Global 300mm DRAM capacity is roughly 2.07 million wafers per month (projected 2.25 million in the near term), so 900,000 is about 40 percent of world output. The number is widely cited and widely misunderstood: it is not a shipped order, it is undiced-wafer LOIs that have not yet converted to binding purchase orders, and critics argue it inflated procurement panic ahead of actual consumption. Whether or not Stargate ever draws 900,000 wafers, the signal alone was enough to trigger defensive forward-buying — Yole Group specifically calls out "fear of future shortage" as a present-tense price driver.

The downstream effects are already visible outside Hetzner. OVHcloud raised public cloud and bare metal roughly 9 to 11 percent in the same window citing RAM up 250 to 300 percent. Micron wound down its Crucial consumer brand to redirect consumer DRAM capacity toward enterprise and automotive. Samsung earmarked about 60,000 wafers per month for HBM4 production by fall 2026. None of this is Hetzner-specific; Hetzner is just the cheapest provider with the least headroom to absorb it before repricing.

What the hikes do to a Bex fleet's bill — three worked fleets​

Take a concrete, boring production shape: a three-node HA control plane plus a two-node worker pool, 20 GB of volume per node, DE/FI region, no autoscaling headroom. The only variable is which instance family the MachineDeployment names. Prices are EUR/month excluding IPv4, DE/FI, post-June 15 list.

A few honest caveats that cut both ways:

• If you already run CCX13 and never rescale, you still pay 15.99 per node — June grandfathering protects you until the next MachineDeployment rollout or volume resize. The moment you roll the pool to a new Kubernetes version or change instance type, you reprice.
• US/SG regions are worse: the same CCX13 in the US went 18.49 to 50.49, so a US-hosted fleet's delta is larger.
• ARM is not a drop-in for every workload. x86-only binaries, older JVM assumptions, and some Helm charts still assume amd64. Moving a worker pool to CAX saves 80 percent of the repaved cost, but you need to verify the container image matrix first — docker buildx multi-arch is not free work.

The sensitivity is the point. "Self-hosted on Hetzner is 10x cheaper than Render" is now a family-dependent line, not a single number. Re-running your spreadsheet with CAX as the default worker pool and reserving CCX only for latency-sensitive stateful nodes is the cheapest move that does not require changing providers.

Second-order: memory add-ons jumped ~575% in April. Scaling vertically by adding RAM to a CX just got punitive. Scale out on ARM beats scale up on Intel this cycle.

Why the honest comparison is now a trend line​

Every PaaS cost comparison published before spring 2026 made an implicit assumption that was safe at the time: the bare-metal floor price is flat. Hetzner's three hikes in five months break that assumption in public, and they trace to a single upstream curve — DRAM contract prices — that is itself moving with the AI boom.

That means the self-hosting pitch in late 2026 has to change framing, not just numbers:

Name the vintage of every price you quote. "Hetzner CX22 at 4.15 EUR/month" is wrong twice — the SKU was renamed (CX22 to CX23), and the old price is pre-April. Mixing a March Railway screenshot against a June Render table and a January Hetzner table silently miscounts bandwidth and compute. Techsy's July 19 same-date audit got this right; apply the same discipline to Hetzner.

Show the range, not the point. Publish the cheapest and most expensive sensible node pool for your shape (CAX vs CCX), not just the family that makes your platform look best. If your PaaS runs a mixed fleet — ARM workers plus a dedicated-vCPU control plane — say so and show the blended number.

Acknowledge the non-binding 40 percent. Stargate's 900,000 wafers per month is an LOI, not a shipment, and Sam Altman has reportedly walked back the implied consumption since. Citing it as destiny overstates the case. Citing it as a demand signal that already triggered defensive buying and pulled forward the price curve — which Yole, TrendForce, and procurement desks all say is happening — is the defensible read.

Do not freeze a post-hike number either. TrendForce expects DRAM output to grow more than 20 percent in 2026 as fabs convert capacity, and the HBM-to-DDR5 margin gap narrowing should ease crowding at the margin. Hetzner itself capped June at new orders and rescales, which is a very different posture from April's "every instance now" repricing. The floor may stabilize, but it will stabilize higher — DDR5 contract prices more than doubling in a quarter do not unwind in a month.

The Hacker News consensus across the 287-comment June thread is revealing: "even after +30 to 50 percent, Hetzner is the cheapest option." That is true against hyperscalers and against per-GB PaaS egress — it remains the central fact of the owned-hardware thesis. What it does not survive is quoting a single frozen pre-hike number as if the thesis has no slope. The thesis still holds; its y-intercept moved.

What to actually do on a Cluster API fleet this quarter​

This is not a panic post. It is a re-default post. Four concrete moves that production teams running on Hetzner via Cluster API Provider Hetzner (CAPH) are actually making:

1. Re-default the worker pool to ARM and gate x86. Set CAX11 or CAX21 as the default MachineDeployment and require a workload label to schedule onto a CCX/CCX pool. Most stateless web workloads do not need dedicated vCPU; they need more replicas on cheaper cores. Keep one CCX pool for Postgres or other latency-sensitive stateful workloads and nothing else.

2. Pin existing CCX instances and defer rescales. June grandfathering is a gift with an expiry date — the day you roll the pool. Batch your next Kubernetes minor upgrade with a deliberate repricing decision rather than letting a routine kubectl rollout silently triple the node bill. If you must add capacity, add ARM workers first.

3. Scale out, not up, and keep memory headroom flat. The 575 percent memory-add-on hike punishes vertical scaling. Prefer one more CAX node over 8 GB more RAM on an existing CX. Hold 15 to 20 percent headroom rather than scaling exactly to demand so a burst does not force an emergency rescale onto new pricing.

4. Keep a second provider as a comparison, not yet as a migration. Netcup dedicates under EUR 9 for 4 cores/8 GB today and has not moved since May, which looks tempting on a CCX-vs-Netcup table. But CAPH maturity, API surface, and regional footprint are not interchangeable — Netcup is a cost footnote, not a drop-in Cluster API provider. Track it, do not replatform to it this quarter. Multi-location and multi-server-type fallback within Hetzner (DE plus FI, CAX plus CX) buys more resilience per hour of operator work than a provider switch.

None of this makes Hetzner expensive. It makes the 2024 assumption — "quote one Hetzner price, forget it" — the thing that expired. The new assumption is the one the AI shortage forces on everyone else too: hardware costs are a curve driven by the same boom you hoped would stay on the demand side.

Hetzner's 2026 is a preview of a weirder 2027 for anyone whose platform owns hardware. The AI buildout does not just compete for GPUs; it competes for the DRAM, SSDs, and power grid your PaaS never planned to outbid. That is not a reason to retreat to metered egress at $0.15/GB — but it is a reason to stop quoting a single price as if the floor cannot move.

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 node pool is a line in your own CAPH config instead of a PaaS plan page, repricing is a diff you review, not an email you receive. Star the repo on GitHub or deploy your first app today.