Zerops Says Its Pricing Is 'Just Hardware.' The Real Bill Runs 2.6x Raw Hetzner

"PaaS is just hardware with a cherry and bow on top, so why would we charge you for anything else but hardware resources?" That's Zerops' own pitch, lifted straight from its docs, and the small Czech-founded developer cloud — which closed a $2 million seed round led by Gi21 Capital in May 2026 — has built its entire pricing page around that one sentence. No plans. No feature tiers. No per-seat fees. Just CPU, RAM, and disk, billed by the minute.

Run the actual 2026 rate card through a real three-service app, though, and the bill lands at 2.6x what the identical CPU/RAM/disk specs cost as raw, self-managed Hetzner boxes. That's not a knock on Zerops — a managed platform is supposed to cost more than bare metal, and the gap buys real things. But "just hardware" undersells what's actually happening on the invoice, and the honest version of that gap is the more useful number for anyone deciding between renting Zerops, renting nothing and running raw VPSes, or running the same hardware under a self-owned Cluster API fleet.

The Actual 2026 Rate Card​

Zerops' pricing page is refreshingly free of the usual PaaS pricing-page fog — tiers named after animals, "contact sales" walls, per-seat multipliers. Here's the current rate card in full, pulled directly from Zerops' own docs:

Two things worth flagging before the worked example. First, the "no per-request or per-GB meter" framing isn't quite right — object storage is metered per GB every month, and egress, backups, and build time all convert to per-unit overage charges the moment a project crosses its plan's included allowance. What's true is that the included allowances are generous enough that most small-to-mid apps never see those line items: 100 GB of egress on the free tier, 3 TB on the $10 tier, is a lot more headroom than Railway's or Render's tighter per-GB egress meters. The metering exists; it's just pushed far enough out that it rarely bites.

Second, "bare-metal" isn't marketing filler here. Zerops runs on its own AMD EPYC servers with NVMe storage and ECC RAM, across datacenters with n+1 redundancy — no reselling AWS or GCP capacity underneath. Every layer above the hardware was built in-house rather than bought off the shelf: Incus for containers (full Linux OS containers, a middle ground between a Docker process and a full VM), a custom VXLAN mesh for per-project private networking, and custom L3/L7 balancers handling entry traffic, routing, and SSL termination. That's architecturally the closest thing on the hosted-PaaS market to what a Cluster-API fleet does for itself on owned Hetzner machines — the difference is who owns the boxes underneath, not how the software layer works.

Same App, Three Ways​

Take the topology this blog keeps coming back to because it's what a real small app actually looks like: one web service, one background worker, one Postgres instance. Size it modestly — 2 vCPU / 4 GB for web and worker, 4 vCPU / 8 GB for Postgres — and price it three ways.

On Zerops, using shared CPUs, the $10 Serious Core (needed for 3 TB of egress headroom on anything real), and a right-sized disk per service rather than a bundled allotment:

On raw Hetzner, the identical specs — post the June 2026 price-hike numbers this blog already recomputed — run a CX23 (2 vCPU/4 GB) for web, a second CX23 for the worker, and a CAX21 (4 ARM vCPU/8 GB) for Postgres: €5.49 + €5.49 + €10.49 = €21.47/mo, roughly $25/mo.

That's a $40.80/mo gap (2.6x) for the same raw compute. What that $40.80 buys on Zerops: managed Postgres with automated backups, the VXLAN-isolated private network and L7 balancer with SSL termination configured out of the box, a GitHub/GitLab-triggered build-and-deploy pipeline, and autoscaling in 0.125 GB / 1-core steps with zero YAML to write. On the raw Hetzner boxes, none of that exists until someone builds it — either by hand, or via a Cluster-API-managed fleet like bex, which provisions the same managed layer (TLS, private networking, git-push deploys, autoscaling) as software running on the hardware you already own, rather than a recurring per-core markup.

The HA Tax Nobody Bundles for Free​

"No plans, no tiers" doesn't mean high availability is free — it means HA is billed exactly as much extra compute as it actually costs, with nothing hidden in a plan upgrade. Zerops' HA database mode runs the cluster across 3 containers instead of 1 (with 2 HAProxy load balancers thrown in at no extra charge). Flip the Postgres tier above to HA and the database line alone triples, from $28.40 to $85.20, pushing the full-stack total to $122.60/mo.

The same HA architecture self-managed — three CAX21 boxes running something like CloudNativePG, plus the web and worker instances — prices out around $49/mo in raw hardware. The gap widens to roughly $73.60/mo once HA enters the picture, because Zerops' per-container markup applies to every replica, not just the primary. That's the honest reading of "flat rate": it's flat per unit of compute, but redundancy still costs three units, on Zerops exactly as much as it would building it yourself.

What Happens at 5x​

A single three-service app is a toy-sized data point. The more useful question is what happens as a team's real product grows into more services inside the same project — five times the compute, still one app, still one Zerops project (so the $10 Serious Core doesn't multiply, only the resources do):

The ratio barely moves — 2.6x at 1x, 2.3x at 5x, slightly better because the fixed $10 core fee dilutes across more resources. What moves is the absolute dollar amount sitting on the table: $40.80/mo isn't worth anyone's engineering time to avoid, but $164/mo — nearly $2,000/year — is real money, and it keeps growing in direct proportion to usage, forever, because Zerops' per-core and per-GB markup never shrinks. A Cluster-API fleet's management layer, by contrast, is a cost a team (or bex, on the team's behalf) pays once as software, not a percentage that compounds with every additional service. That's the actual crossover: not a single dollar threshold, but the point where the growing gap outpaces the fixed cost of owning the platform layer instead of renting it.

Where the Markup Is Worth It — and Where It Isn't​

Zerops' bet earns its keep for exactly the team it's built for: no dedicated infra hire, a handful of services, and a genuine preference for managed HA Postgres, automatic TLS, and git-push deploys over spending a weekend wiring those up. At $40.80/mo of premium for a real three-service app, that's cheaper than an hour of anyone's time, and the "hardware-only" pitch — while not literally true once object storage, egress overage, and HA replication enter the picture — is close enough to honest that it's a fair PaaS to recommend for that team.

It stops being the obvious choice once the gap compounds into hundreds of dollars a month across a growing service count, which is exactly the scale at which owning the boxes under a Cluster-API-managed fleet stops being a DIY chore and starts being the cheaper, more durable default — because the platform-engineering cost that Zerops recovers through a permanent per-core markup gets paid once, in software, instead of forever, on the bill.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, provisioned through Cluster API onto the same class of raw hardware this post priced against Zerops. Star the repo on GitHub or deploy your first app today.