Hetzner Just Tripled Its Best VMs: What the 2.7x Price Shock Does to Self-Hosting Math

Hetzner's best virtual machines cost 2.7 times what they did four months ago. Not a typo — not 27 percent, 270 percent.

On June 15, 2026, Hetzner applied its second price increase in ten weeks. The headline from April — cloud servers up 30 to 37 percent — already stung. The June hike lands differently: dedicated-vCPU CCX instances jumped 2.1x to 2.73x per month in Germany and Finland, shared-vCPU CPX instances jumped 2.4x to 2.75x, and in the US the same CPX line hit 3.1x. Hetzner's cheapest US VPS went from roughly $7 to $20.49 per month, a 193 percent increase, in a single year. The supply-constrained hardware underneath your self-hosted PaaS just got repriced twice while most "self-host on a Hetzner box" comparisons still quote the pre-hike number.

So the honest question is not whether Hetzner is still cheap in absolute terms. It is whether the entire cost-advantage pitch that platforms like Bex.co build on — own the hardware, skip the per-GB meters — survives when the cheapest side of the comparison tripled and the expensive side, Railway, Render, and Fly.io, did not stand still either.

Here is the recomputed answer, line by line, with every price pulled from the same post-hike price sheets.

The Answer Up Front: Even at 2.7x, the Flat Box Still Wins — But the Family You Pick Now Matters​

If you priced a typical small production workload last November, Hetzner won by a laughable margin. If you reprice it today with June numbers, Hetzner still wins — just not by the same margin, and not on every instance family.

Sources: Hetzner price-adjustment notice via docs.hetzner.com, byteiota.com/costgoat.com post-June audits, vpsmaxxing comparison updated June 15, thevibeworks/claude-reads-hn HN thread (175 points, 287 comments). Euro-to-dollar at ~1.08.

Two things jump out before we touch any competitor:

1. The hike was surgical, not uniform. Intel-shared CX and ARM CAX lines rose a comparatively mild 30 to 40 percent — the same neighborhood as April. AMD-shared CPX and dedicated CCX lines more than doubled. That split is the whole story for a Cluster API fleet's default node pool.
2. The cheapest box moved the most in the US. Hetzner's US regions (Ashburn, Hillsboro) tracked the steepest CPX increases, up to 3.1x, pushing the entry price every US-vs-EU comparison leans on from $7 to $20.49.

The cost-advantage thesis does not collapse. It gets narrower and more family-dependent, which is exactly the nuance most pre-hike blog posts omitted.

What Actually Changed: Two Hikes, Two Different Rules​

Hetzner's 2026 pricing has three chapters, not one. Collapsing them into "Hetzner got expensive" misses the mechanics that decide what you actually pay.

April 1, 2026 — the broad hike. Announced February 24, effective April 1, Hetzner lifted cloud servers in Germany and Finland 30 to 37 percent, US and Singapore cloud up to 38 percent, dedicated servers 3 to 21 percent, and storage volumes 30 percent. The critical detail: it hit new and existing customers together. Monthly-contract tenants woke up to the higher invoice with no way to lock the old rate. Hetzner cited DRAM, NAND, and SSD contract prices by name — the same AI-driven memory squeeze everyone else was quietly absorbing.

June 15, 2026 — the targeted hike. Announced May 27, effective June 15, Hetzner standardized its line into -1/-2/-3 tiers and repriced again — but only on new orders and rescales. Existing contracts were grandfathered at their April rate. This time the increases diverged wildly: CCX dedicated-vCPU lines rose 113 to 176 percent depending on size (CCX63 famously went €374.49 to €853.49 per month), CPX shared-AMD lines rose a similar band, while CX and CAX rose roughly 30 percent. Independent audits put the family-level multipliers at 2.1x to 2.73x for CCX and 2.4x to 2.75x for CPX in EU regions, with the US CPX tail at 3.1x.

In practical terms:

• If you provisioned before June 15 and never rescaled, you still pay the April price.
• If you provision today, you pay the June price — two hikes stacked.
• The cheapest monthly bill for a new tenant is therefore not "30 percent more" but two to three times the number every 2025 benchmark quoted.

That grandfathering is why two teams can report completely different Hetzner bills in the same week and both be telling the truth.

The Workload We Price: One Boring, Representative Stack​

Cherry-picking a micro VM or an enterprise monster makes any comparison lie. The fair test is the workload most solo-to-small-team PaaS tenants actually run:

• App: Node.js API (2 vCPU, 4 GB RAM) — the median container before you add a beefier build step
• Database: managed Postgres equivalent (2 vCPU, 4 GB RAM, 20 GB NVMe volume) — not SQLite, not a 32 GB monster
• Storage: 20 GB persistent volume for the database
• Egress: 100 GB per month to the internet — modest for an image-heavy marketing site, low for a SaaS with API consumers
• Runtime: 730 hours per month (always on) — no scale-to-zero, because the API and database stay warm

Why this shape is representative: it is exactly the stack Railway, Render, and Fly.io price calculators default to, it is what a Hetzner CX22/CPX31/CCX13 fleet actually bin-packs without overprovisioning, and it surfaces the meter that decides most bills (egress) without requiring a viral launch to matter. Sensitivity — where the answer flips — comes next.

Line by Line: The Same Stack on Four Bills​

All numbers below use vendors' published July 2026 pricing, post-Hetzner-hike, with June 15 Hetzner sheets for EU regions. Dollars are monthly, volumes are per-GB-month, egress is internet-facing.

How to read this: the Hetzner column is the price of the box the workload lives on. The other three columns are the sum of every line a real invoice adds up — not just the compute tile the marketing page highlights. Railway's Hobby plan burns its $5 credit against the meter, so you still pay the meter. Render's egress cut — 100 GB free became 5 GB free on Hobby on April 23, 2026 — turns every image-heavy deploy into an overage line. Fly.io's per-second precision is real, but NAT and metrics sit on separate lines that a flat box bundles away.

What the gap says now:

• Cheapest to most expensive: Hetzner CX23 at $6.49 versus Railway at $93 — a ~14x spread on the same logical workload. Even the most expensive post-hike Hetzner box in this table (CPX31 at $53.90 or CCX13 at $46.40) lands at or below Render and Fly.io's fully loaded totals once egress and volumes are counted.
• Where the meter moves first: egress. At 100 GB per month, Render's egress overage alone ($14.25) is more than double the entire Hetzner CX23 box. Double egress to 200 GB and Render adds another $15 while Hetzner adds $0.
• Where per-second stops winning: always-on Postgres. Fly.io's per-second billing shines for bursty, low-average-load workers. A database that never sleeps erases that advantage — you pay for every second of 730 hours, plus the volume it sits on.

The pre-hike version of this table had Hetzner at roughly $4.35 and $16 to $20 for the same families — the spread was wider, but the ordering was identical. The hike narrowed the margin; it did not invert it.

Where the Cheap Box Stops Being Cheap​

A flat box is not always cheaper. The honest recompute names where it loses, because a team that adopts Hetzner for the wrong workload will feel cheated twice — once by the hike, once by the fit.

The low-utilization edge. A workload that runs five minutes per hour (8 percent utilization) — a nightly crawler, a webhook handler invoked a few hundred times per day — pays for 55 idle minutes on a flat Hetzner box but only five billed minutes on Fly.io or Railway per-second meters. At $0.02 to $0.04 per vCPU-hour billed per second, that sparse job can land under $2 per month on a metered platform while even the cheapest CX23 bills $6.49 whether it idles or not. If your fleet is mostly sparse jobs, meters beat flat hardware.

The family-choice edge. The June hike made "which Hetzner line" a cost decision, not a brand decision. A Cluster API fleet defaulting to CPX (shared AMD) or CCX (dedicated) today pays 2.4x to 2.75x the price its autoscaler was tuned against six months ago. The same workload on CX (Intel shared) or CAX (ARM) pays only 1.3x to 1.4x. For stateless web tiers and build runners that tolerate shared vCPU, migrating the default MachineDeployment from CPX to CX/CAX reclaims most of the hike overnight. For latency-sensitive Postgres primaries that need dedicated cores, the CCX premium is now a real budget line — budget it, do not pretend it is still 2024 pricing.

The sensitivity is not theoretical. Raise egress to 500 GB — a modest CDN-backed SaaS without Cloudflare in front — and Render's overage hits $74.25, Fly.io's hits $10, and Hetzner still hits $0. The flat box's advantage grows with traffic; it shrinks with idleness.

Why It Happened and Why It Does Not Un-Happen Next Quarter​

Both Hetzner hikes cite the same root cause, and it is not a Hetzner-specific story.

AI infrastructure is eating the world's DRAM. Samsung, SK Hynix, and Micron have tilted wafer capacity toward high-bandwidth memory (HBM) for GPU clusters — HBM needs roughly three times the wafer capacity per gigabyte as standard DDR5, per Micron. SK Hynix told investors its HBM, DRAM, and NAND capacity is sold out through 2026. Samsung raised 32 GB DDR5 module pricing 60 percent between September 2025 and January 2026. Industry trackers put the numbers bluntly: DRAM contract prices rose 43 to 48 percent in Q4 2025 alone, TrendForce projected a further 58 to 63 percent in Q2 2026, and one synthesis put the year-over-year DRAM increase at roughly 171 percent. One report flagged that 70 percent of high-end DRAM production in 2026 is already spoken for by AI data centers.

That is the supply shock Hetzner's price sheets are downstream of. When the input — server RAM, NVMe — costs two to three times more and is physically constrained, a provider that actually owns its hardware cannot absorb the increase forever without raising prices or running out of stock. Hetzner did both: it raised prices and, separately, hit "limited availability" on cloud SKUs constrained by the same component shortage.

Other EU providers confirm the pattern is industry-wide. 365i and Straithead coverage from February to July 2026 notes OVHcloud flagging 250 to 300 percent projected RAM cost increases into late 2026, and Hetzner and OVHcloud raising in tandem. The Hacker News consensus on the June thread still landed at "even after 30 to 50 percent, Hetzner is the cheapest option" — a thinner margin, but not an inverted one.

The implication for any self-hosting cost comparison written after Q2 2026: quote a hardware-cost trend line, not a frozen point. The pre-April CX22 at ~€3.79 is not coming back while wafer capacity stays tilted toward HBM. The honest framing is "we provision from a family that rose 30 percent, not 175 percent, and we grandfather contracts that did not reprice" — not "self-hosting is always ten times cheaper."

Even at 2.7x, the bandwidth math still carries the pitch. A Hetzner CX/CAX box ships 20 TB of included transfer. At Render's $0.15 per GB overage, 20 TB would be billed at $3,000 if metered. At Fly.io's $0.02 per GB, it would be $400. That single line — the one Hetzner bundles and hyperscalers meter — is worth more than the entire hike on any instance size a small fleet actually provisions.

What This Means for a Bex Fleet​

If you run a self-hosted PaaS on Hetzner through Cluster API, the practical response is not to argue the hike did not happen. It is to provision like you believe it:

• Default node pools to CX or CAX, not CPX/CCX, for stateless tiers. Web servers, build runners, and preview environments tolerate shared vCPU. Keep dedicated cores for Postgres primaries and build that cost into the database line — do not pay the 2.7x premium where the 1.3x family suffices.
• Use the grandfathering window. Machines provisioned before June 15 that have not rescaled keep the April rate. That makes in-place scaling and node-pool immutability a cost decision: prefer adding new nodes from the cheaper family over rescaling an existing expensive family and triggering repricing.
• Price egress honestly in every comparison you publish. The two-to-three-times hardware increase is real, but so is the zero on the bandwidth line. Any comparison that shows compute without showing egress understates the flat box's advantage by more than the hike took away.
• Hold capacity headroom or a second provider candidate. Hetzner has also shown "limited availability" on cloud SKUs under the same DRAM squeeze. A fleet that can fall back to a second region or a second provider profile is not hedging against price — it is hedging against sold-out inventory.

The bottom line recomputed with post-June numbers: a representative always-on Node.js + Postgres stack at 100 GB egress costs roughly $6.49 on a Hetzner CX23, $46 to $54 on the larger Hetzner families, $37 on Fly.io fully loaded, $52 on Render fully loaded, and $93 on Railway fully loaded. At 500 GB egress the Hetzner column barely moves while the managed columns add $10 to $74. The "own cheap Hetzner hardware" argument is less cheap than it was. It is still the cheapest column on the page — if you pick the right family and count the bandwidth.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Cluster API manages the fleet, Hetzner (or your provider) bills only for the boxes, not per GB or per second. Star the repo on GitHub or deploy your first app today.