Nine apps, one €14 VPS: what moving off Railway actually cost me

Amit Ranjan · 13 August 2026

I had four applications on Railway costing about $17/month, and five more I wanted to put somewhere. Napkin maths said $25–40 once everything was up — and Railway's pricing scales with every service you add, so "somewhere between" was the honest answer, not a number.

For a business with revenue that's noise. For a solo founder with a runway, it's a line item you look at.

I moved everything to a single VPS running Coolify. One flat bill, ~€14/month. Here's what that actually involved, including the part where a provider refused my money.

What was running

Four Railway projects, and the per-project costs are worth showing because they weren't what I expected:

The interesting row is the third. My personal trading dashboard — the thing with zero users but me — was more than half my hosting bill, because of a persistent volume holding about a gigabyte of market data. My most complex application, the one running headless Chrome in a background worker, cost a quarter of that.

That is the thing about usage-based pricing: the cost lands where the *resources* are, not where the *value* is. You don't notice until you itemise it. I'd have guessed the Playwright worker was my expensive one and I'd have been wrong by 4x.

Sizing it honestly

Before comparing providers I wrote down what I actually needed to run, because the temptation is to pick the cheapest box and discover the OOM killer at 2am.

Two things drove the number up. Playwright is the expensive tenant — a headless browser will happily take a gigabyte per instance, and it spikes rather than sitting flat. And Docker on a multi-app box eats disk faster than you expect: images, build cache, and a Postgres volume per app. I wanted ≥200 GB so I'd never be pruning images at midnight to finish a deploy.

If you take one thing from this post: size for your worst tenant's peak, not your average. The average is a comfortable lie.

The provider comparison

I compared eleven options on vCPU, RAM, disk, bandwidth, price, and — this is the part most comparison posts skip if they're not written from India — whether I could actually pay them.

Watch the renewal prices. Hostinger's headline numbers are excellent and their renewals are 2–3x that. A first-year price is a marketing number; budget on year two.

DigitalOcean at ~$48/month for 8 GB would have cost more than the Railway bill I was leaving. Managed-cloud pricing in 2026 has drifted far enough from bare VPS pricing that it's not a close call at this scale.

The provider that said no

Hetzner rejected my account. Verification failed for an Indian signup, which is common enough that it's worth stating plainly — Hetzner is the default recommendation in every "cheap VPS" thread, and if you're signing up from India there's a real chance you'll spend a day on it and end up nowhere. I'd budgeted them as first choice and had to write "do not retry" in my own notes.

Payment method is a real constraint, not a footnote. Half the "best VPS" lists on the internet are written by people who have never had a signup declined.

Chosen: Contabo Cloud VPS 8 — 8 vCPU, 24 GB RAM, 300 GB SSD, ~€14/month, Ubuntu 24.04, EU region.

The migration

Coolify installs with one command and gives you a Traefik reverse proxy, automatic Let's Encrypt certificates, git-based deploys, and a UI for environment variables. It's the closest thing to Railway you can run yourself, and that similarity is the whole point — the mental model transfers.

The actual sequence:

1. Provision the box, install Coolify, add an SSH key
2. Recreate each app as a Coolify resource — pointed at the same GitHub repo, same Dockerfile, environment variables copied across by name
3. pg_dump every Railway Postgres, restore into the new one
4. Copy the volumes. ~934 MB for the NSE desk's data directory, ~18 MB of artifacts from the regression tool's object storage
5. Re-point DNS — A records to the VPS IP, Railway CNAMEs removed. I automated this through my registrar's API rather than clicking through twenty records.
6. Wait for certificates, then smoke-test every domain over HTTPS
7. Rebuild the scheduled jobs. One app had a Railway cron hitting an endpoint every five minutes; that became a host crontab entry.
8. Only then, stop the Railway services

Point 8 matters. Keep the old thing running until the new thing has been serving real traffic for a few days. Migration posts always end at "DNS propagated 🎉" and skip the week where you find the thing you forgot.

The things I forgot: the object storage (Railway's Tigris bucket had artifacts I'd never thought about, now on a local volume), and one API subdomain that had never had its own A record because Railway was handling it implicitly.

Four applications, databases restored, ~1 GB of volume data moved, and all five custom domains serving HTTPS. One working day. Making those deploys reproducible from git a few days later found four production bugs that uptime checks had never seen, including four containers answering to the name postgres.

What it actually cost me

The savings are real: ~$17–40/month down to ~€14 flat, and — more importantly — the marginal cost of the next app is now zero. That changes behaviour. I stopped thinking about whether a demo was "worth" deploying.

But the bill is not the whole trade, and I'd be selling you something if I pretended it was.

What Railway was doing that I now do myself:

• Backups. Railway's managed Postgres backed itself up. Mine doesn't. Four databases on one machine with no backup is a strictly worse position than I was in before, until I fix it — and "until I fix it" is how people lose data. If you make this move, set up automated pg_dump to *another provider's* object storage before you cut over. Not after.
• Security. The platform's attack surface is now mine. An admin panel, an SSH port, and every container on one host. Coolify's dashboard defaults to a plain HTTP port, which is fine on a laptop and not fine on a public IP — put it behind TLS and a firewall on day one.
• Uptime. No platform team. If the box goes down at 3am, it's down until I notice. External uptime monitoring is not optional; it's the thing standing in for the ops team you no longer have.
• The blast radius. Nine apps on one machine is nine apps that share a fate. Railway's isolation was worth something.

Latency. I'm in India; the box is in the EU. That's roughly 150ms of round-trip I've accepted because everything on it is a demo or an internal tool. Contabo's India datacentre is about $4/month extra and would be the obvious call the moment a real customer is served from it.

"Unlimited" bandwidth carries a fair-use asterisk, as it always does. Fine for this workload; read the terms if yours is different.

When this is the right move

Do it if you're running several small always-on services, your platform bill scales with service count, you're comfortable with Docker and a Linux box, and you'd rather own the operational work than pay to avoid it.

Don't if your bill is under ~$10/month (the migration costs more in time than you'll save), you have real customers and no on-call story, or you'd be doing it for the aesthetics of self-hosting rather than a number you actually looked at.

The honest summary: I traded roughly $25/month for a few hours of ongoing operational responsibility. At my current stage, where the marginal-cost-of-a-new-demo being zero changes what I build, that's a good trade. When there's a paying customer on that box, I'll re-run the sum — and the answer might well be different.

I build AI systems — retrieval platforms, in-app assistants, automated regression testing — and deploy them on infrastructure like this. If your team wants something in production rather than in a notebook, the AI Readiness Audit is the shortest way to find out what's actually running.