Dora NodaA September pricing guide did something rare: it named the number. Past roughly $150 a month in total Railway spend, the guide says, self-hosted Coolify or Dokku on a dedicated VPS beats metered PaaS on price alone. That number has been bouncing around teams deciding whether to stay or go — and it deserves a recompute, because it answers the single-VPS question, not the fleet question.
Here is the recompute up front: against a three-node Hetzner fleet you already run, the crossover is not $150. It is about $30 — roughly one always-on full-stack project. The rest of this post shows the ladder, the itemized math on both sides, and which workload shapes cross first.
The $150 line in the sand
Nayan Kyada's Railway Pricing 2026 guide, published September 7, walks the full ladder from trial credit to Pro. The ladder, as the guide documents it: a one-time $5 trial credit that expires after 30 days, a Free plan with $1 of monthly usage and tight per-service caps (1 vCPU, 0.5 GB), Hobby at $5 a month including $5 of usage, and Pro at $20 a month per workspace including $20 of usage with unlimited members.
The unit rates underneath are what matter, and they are confirmed by Railway's own 2026 pricing writeup: about $20 per vCPU per month, $10 per GB of RAM per month, $0.05 per GB of egress, and about $0.15 per GB-month of volume storage — all metered per second. The guide's worked examples make the mechanism visceral: a single 512 MB service running continuously burns roughly $5 a month in memory alone, the entire Hobby credit, before CPU. A real full-stack project lands at $20–40 a month on Hobby once usage is counted honestly.
The verdict the guide lands on is the honest one for its framing: Pro's ceiling is uncapped, and large teams past roughly $150 a month in total spend should request volume pricing or evaluate self-hosted Coolify or Dokku on a dedicated VPS. That is the baseline being recomputed here — not disputed, just re-run against a different denominator.
The recompute: three columns, three project counts
The $150 figure compares metered spend against one VPS. That is the right comparison for one project, but it quietly assumes the self-hosted side scales the way Railway does — pay more, get more, one box at a time. A small Cluster API fleet does not work that way: its cost is flat, and every project until the fleet fills up rides at zero marginal cost.
So the honest recompute needs three columns, not two, and a fleet spec box instead of a hand-waved number.
The fleet: three Hetzner CX33-class cloud VMs (4 vCPU, 8 GB RAM, 80 GB NVMe each) at about $10 a month per node on Hetzner's current cloud list, for roughly $30 a month all-in. That buys 12 vCPU, 24 GB of RAM, and 240 GB of NVMe, with 20 TB of included traffic per server making egress effectively free at these volumes. The CAPI controllers ride along on the same nodes at a few percent overhead — no separate management cluster in this math, though adding a small dedicated mgmt node costs about $6.50 if you want one.
Why one to three projects fit: a modest always-on stack (app at 0.5 vCPU / 0.5 GB, Postgres at 0.5 / 1 GB plus a 5 GB volume, Redis at 0.25 / 0.25 GB, 20 GB of egress) draws about 1.5 vCPU and 2.25 GB per project. Three projects plus system overhead (CNI, monitoring, controllers) land near 6 vCPU and 10 GB — half the fleet, with headroom left for preview environments. A fourth node enters the picture around project six to eight, or when Postgres data outgrows a single node's local NVMe.
Now the table. Each Railway column itemizes the same modest stack per project (app ~$15, Postgres ~$21, Redis ~$7.50, egress ~$1 — about $44 per always-on project), against a self-hosted Coolify panel ($0) on a $12 VPS-class box, against the $30 fleet:
| Monthly cost | 1 project | 2 projects | 3 projects |
|---|---|---|---|
| Railway (usage far above credit, either plan) | ~$44 | ~$88 | ~$133 |
| Single VPS + self-hosted Coolify | ~$12 | ~$12–24 | ~$24+ and a resize |
| 3-node CAPI fleet on Hetzner | ~$30 | ~$30 | ~$30 |
Three readings fall out. First, one project belongs on the single VPS — $12 beats $30, and anyone telling you to run a fleet for one app is selling you something. Second, Railway's bill passes the fleet's entire fixed cost before the first project is even fully counted: ~$44 of metered spend versus $30 of flat hardware. Third, the guide's $150 crossover overstates the stay-on-Railway zone by roughly 5x for always-on workloads — not because the guide is wrong, but because its denominator is one VPS that you will outgrow, while the fleet's denominator is shared across everything you deploy next.
Which workloads cross first, in dollars
The table above is the always-on case, and always-on crosses first. Any stack with a database that cannot sleep, a worker that runs around the clock, or an API with steady traffic bills 24/7 on Railway's per-second meter — there is no idle discount for a process that never idles. The crossover for this shape is the fleet's fixed cost itself: about $30 of Railway spend, which a single app-plus-Postgres project already exceeds.
Preview environments cross second, and they cross harder than teams expect. The guide budgets $15–30 a month above production for branch deploys; the services-times-days math agrees. One PR environment with an app and a database runs about $35 at full-month rates, so three PRs open a third of the month cost roughly 3 × $35 × ⅓ ≈ $35 — more than the entire fleet, for environments nobody looks at most days.
On owned hardware with headroom, those same previews are marginal-cost zero. Any team running per-PR databases is subsidizing its CI with its hosting bill.
Idle side projects are the honest exception — the shape that genuinely stays cheaper metered. Railway's Serverless setting sleeps a service after roughly ten minutes without traffic, and a sleeping service bills close to nothing: volume storage at $0.15 per GB plus pennies of wake time. Five sleeping side projects fit inside a $5–10 Hobby bill that no owned box can touch, because the box charges for the hours the projects spend doing nothing. Below about $12 a month of Railway spend — the price of the cheapest VPS worth running — metered wins, and sleep is the mechanism that keeps it winning.
Egress is the sensitivity variable worth naming separately. At the 20–50 GB a typical app serves, Railway's $0.05 per GB is $1–2.50 of noise. At a terabyte of image or video traffic served straight from the platform, it is $50 — decisive on its own, and the fix (a CDN in front) is the same on either side. If your app is media-heavy and CDN-less, recompute with your own egress line before trusting anyone's crossover, including this one.
What the $150 number gets right — and the ops-time asterisk
None of this makes the guide's number wrong. For its stated comparison — metered PaaS versus one VPS running Coolify or Dokku — $150 is, if anything, generous to Railway: a single $12 box absorbs the first $100-plus of metered spend before it breaks a sweat, as timed Coolify-on-a-$6-VPS walkthroughs keep demonstrating. The number gets the shape of the answer right. It just stops one step short of the fleet, where flat hardware plus multi-tenant bin-packing moves the break-even point from "large team" territory down to "second always-on project."
The asterisk is ops time, and dollars-only math that hides it is lying by omission. A fleet needs upgrades, node rotation, and incident response; at a loaded cost of $100 an hour, two hours a month of fleet care is $200 — several times the hardware. That cuts both ways, though.
If you would build the fleet for this one project, those hours belong in the crossover and Railway wins longer. If you already run the fleet — and anyone self-hosting a PaaS does — the marginal ops cost of the next project rounds to zero, and the dollar crossover stands as computed. The real question is never "is $30 of hardware cheaper than $150 of metered spend." It is "am I already in the fleet business."
Stay or go: the checklist
Three verdicts, then a link. Stay on Railway if your spend sits under ~$30 a month, your projects sleep most of the day, you have no one to page when a node misbehaves, or you need scale-to-zero more than you need predictable bills. Take the single VPS if you have one or two always-on projects and someone comfortable with Docker — it is the cheapest correct answer at that scale. Take the fleet if you are past two always-on projects, run per-PR environments, or already operate Kubernetes and can bin-pack the next deploy for free.
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.
