One 8-vCPU Cloud VM Now Costs More Than a 24-Core Bare-Metal Server: Hetzner's EX130 and the Post-Repricing Fleet Node Map

One 8-vCPU cloud VM now costs more per month than an entire 24-core bare-metal server. Hetzner's CCX33 — 8 dedicated vCPUs, 32 GB of RAM — bills €138.49 a month after the June 15, 2026 repricing. The EX130 dedicated server — a 24-core Xeon Gold 5412U with 48 threads, up to 256 GB of ECC DDR5, and terabytes of datacenter NVMe — rents for €134. The cloud VM is smaller in every dimension and still costs €4.49 more.

That single comparison redraws the node map for any Cluster API fleet on Hetzner. The auction-vs-cloud math already established that buy-used metal beats repriced cloud by multiples. The EX130 adds the tier above it: buy-new dedicated — current-generation platform, full warranty, orderable on demand rather than won at auction — at a price that still undercuts a mid-size cloud VM. This post works the full comparison: price per core and per gigabyte across the whole menu, which fleet pools move to EX130-class hardware, which stay on elastic cloud VMs, and where dedicated provisioning still loses despite the unit economics.

What the EX130 actually is (and isn't)​

First, an honesty note: the EX130 is not a new launch. Hetzner announced it in October 2023, built around Intel's 4th Gen Xeon Scalable (Sapphire Rapids) platform. The genuinely new 2026 EX-line server is the EX131, launched April 21 with a 32-core Granite Rapids CPU, Gen5 NVMe, and a starting price of €560.70 a month plus a €559 setup fee — roughly four times the EX130's monthly rate, with a seven-times setup fee.

That contrast is exactly why the EX130 matters right now. It is still in the lineup at its old price while everything around it moved: dedicated cloud tripled in June, and its successor launched at 4x. The news is not hardware — it is the new price relationship. The lineup, from the original announcement:

Both ship with an included IPv4 address in Helsinki and Falkenstein, and both scale to 512 GB of RAM with up to six additional NVMe drives if you outgrow the base config. The S skews storage-heavy (7.68 TB raw NVMe for the price), the R skews memory-heavy (256 GB for the price). For fleet-node duty the R is usually the right pick — RAM is the binding constraint on a bin-packed worker — while the S fits storage-dense roles like build caches or local object stores.

The worked math: per-core and per-GB across the whole menu​

All cloud prices below are net EUR per month for Germany/Finland after the June 15, 2026 adjustment, excluding VAT and the separately billed IPv4 address. Dedicated prices are the base monthly rate before the one-time setup fee.

Four ratios fall out of that table, and each one answers a different fleet-planning question.

Cloud parity costs 8–10x. Assembling the EX130-R's 256 GB out of CCX23 instances takes sixteen of them at €85.99 each: €1,375.84 a month, over ten times the dedicated price. Even the densest cloud option, two CCX53s at €533.49 each, totals €1,066.98 — eight times the EX130-R for the same RAM with fewer total threads. Every stable, fillable gigabyte in the fleet has one obvious home.

The EX130-R undercuts even the auction per gigabyte. At €0.52/GB it beats the AX41's ~€0.58 and the AX42's ~€0.72 — with new hardware, a current platform, DDR5 ECC throughout, and none of the auction's stock lottery. The auction still wins on absolute bill (€37–46 vs €134), so it keeps the small-fleet and single-box roles. But the "auction is always cheapest per unit" assumption no longer holds at the top of the density range.

The EX131 repriced new dedicated into cloud territory. At €17.52 per core and €4.38 per GB, the Granite Rapids successor sits almost exactly on top of dedicated-cloud unit pricing. Hetzner claims up to 70% more performance than the previous model, and Gen5 NVMe is genuinely new — but on capacity-per-euro, the EX131 is a performance-tier product, not a fleet-density product. The EX130 is the last box in the buy-new lineup priced like metal rather than like cloud.

Shared cloud still owns the small end. The CX33's ~€2.14 per shared vCPU and ~€1.07 per GB look competitive until you remember the units are shared and burstable, not dedicated. For spiky, low-fill, or tiny workloads the absolute €8.56 bill beats everything. Unit economics only favor metal you can fill.

Which raises the fill-rate question — the one that decides whether this math survives contact with a real fleet. The EX130-R needs remarkably little fill to win. Its breakeven against the CCX33's €4.33/GB is just 31 GB used out of 256 — fill 12% of the box and you have already beaten dedicated cloud per gigabyte. At 25% fill (64 GB used) the effective rate is €2.09/GB, still less than half of cloud. At 50% fill it is €1.05/GB, matching a fully utilized EX130-S.

The honest boundary is not fill rate but absolute need: if your steady-state footprint is one or two small VMs, a €134 box is oversized no matter how good its unit price. Past roughly one CCX33 of steady demand, the EX130 wins on math alone.

Two adjustments before acting on the table. The €79 setup fee amortizes to €6.58 a month over the first year — about 5% on the monthly rate, not ranking-changing, but it rises steeply if you upgrade RAM or add drives, so price your actual configuration, not the "from" line. And dedicated stock is finite per facility: unlike the cloud API, the order form can say no, and unlike the auction, the price does not drop while you wait.

Which pools move, which stay​

With the math settled, the assignment writes itself — as an extension of the hybrid pattern, not a replacement for it. The metal floor gets taller and denser; the cloud margin keeps the jobs only elasticity can do.

The EX130-vs-auction split deserves one explicit line: buy the auction box when the absolute monthly bill is the constraint, buy the EX130-R when per-unit density at scale is the constraint. A three-node baseline pool on AX42s costs ~€138 for 192 GB; a single EX130-R costs €134 for 256 GB in one failure domain instead of three. Fewer, denser boxes mean fewer control-plane members to quorum, fewer machines to patch, and fewer repair tickets — operational consolidation that the per-GB table understates.

Where dedicated still loses​

None of the above changes what a Robot-managed server fundamentally is: a purchase, not an API call. Every limitation in the auction analysis carries over, at a higher absolute price that makes mistakes more expensive.

The autoscaler cannot buy you a server. CAPH models dedicated machines through HetznerBareMetalHost objects in an inventory: a human orders the box, records its Robot credentials, and registers one CRD per server. The controller then drives provisioning — rescue boot, installimage, cluster join — declaratively. But as the CAPH docs put it, the bare-metal controller consumes inventory rather than buying machines. Your elastic tier must stay on HCloudMachine cloud VMs, with the autoscaler enabled only there.

Capacity planning reverts to the old discipline. A €134 box provisioned for peak and idle at night is still €134. Size the EX130 floor at roughly p50 load and let CX/CAX burst nodes absorb everything above it. The June repricing sharpened this split deliberately well: the shared lines you burst on rose only ~33%, while the dedicated capacity you are replacing tripled.

Failure replacement is a ticket, measured in hours. Mirrored NVMe covers the common disk failure, and MachineHealthCheck still evicts workloads off a dead node. But the node itself comes back on human timescales — support ticket, hardware swap, rescue reinstall, rejoin. Never let a quorum or a single-replica stateful set depend on exactly one specific €134 box. The metal is cheap enough per unit to run N+1; buy the redundancy instead of wishing for faster repair.

Networking and geography stay cloud-first. Dedicated servers join cloud private networks through a vSwitch uplink — billable, documented separately, and another topology in the runbook. Managed load balancers and floating IPs remain cloud-native. And EX130 stock is Germany and Finland only, so US or Singapore presence stays on cloud VMs regardless of the math.

The bottom line​

Hetzner's 2026 pricing inverted the fleet-node ladder twice. First the June repricing made a €46 auction box cheaper than a 2-vCPU cloud VM. Then the math above it shifted too: a €134 buy-new dedicated server with 24 cores and 256 GB of ECC now undercuts a single 8-vCPU cloud VM, beats the auction per gigabyte, and sits at one quarter the unit price of its own EX131 successor. The default for stable capacity is no longer "cloud unless proven otherwise" — it is metal unless you need elasticity, granularity, or geography that metal cannot give you.

Run the numbers against your own p50 load: if your steady state clears roughly one CCX33 and you can fill big boxes across tenants, an EX130-R floor with a cloud burst tier is the cheapest architecture Hetzner currently sells. Then budget the real cost honestly — a human in the provisioning loop, hours-long repair, a vSwitch in the network diagram. That is a good trade at a tenth of the cloud price per gigabyte. It is just not a free one.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. Star the repo on GitHub or deploy your first app today.