Northflank Ranked 6 Railway Alternatives — I Priced the Same App on All of Them

In June 2026, Northflank — a vendor that sells bring-your-own-cloud infrastructure — published 6 best Railway alternatives, grading each platform on credit-free uptime, deployment visibility, and BYOC support. The six: Northflank itself, Fly.io, Coolify, DigitalOcean App Platform, Render, and Heroku. It is a good roundup as vendor comparisons go, and it makes two concessions worth taking seriously: Railway's beginner-friendly UI is genuinely best in class, and its credit-based uptime plus limited build and log visibility are the reasons people leave.

But a roundup written by a BYOC vendor answers "which dashboard should I rent next" and never "what does the same workload actually total on each, and what does owning the machines cost instead." So I priced one concrete, always-on workload — a web service (1 vCPU / 2 GB), a background worker (0.5 vCPU / 512 MB), and Postgres — on all six, plus a Hetzner box you own. The full table is two sections down; the punchline is that the spread runs from roughly €12 to roughly $77 a month for workloads a reader would call equivalent, and the entire spread is explained by four billing shapes.

The roundup's case against Railway, in brief​

Northflank's critique is specific enough to steelman. On Railway, apps stop running once the $5 trial credit runs out; builds go through Nixpacks detection with a Dockerfile escape hatch that puts more on you; secrets are env vars with the official docs pointing at Doppler for anything serious; volumes exist but are young; background workers are not first-class (the suggested workaround is a manually added second service); and deploy visibility is thinner than the alternatives'.

The roundup also cites an outage record: a December 2025 incident that paused builds across plan tiers in EU West, and a May 2026 platform-wide outage, attributed to Google Cloud suspending Railway's production account, that took the API, control plane, and databases offline for about eight hours.

Some of that has moved since June — Railway's own docs now position Railway Metal, Railway-owned hardware, against Northflank's Kubernetes-or-BYOC story — but the billing-shape critique is structural, not temporal. Credits that gate uptime, minute-metered compute, and a managed-only footprint are the product, not a bug in one quarter. That is the part worth pricing.

The head-to-head: one workload, seven bills​

The workload: an always-on web service at 1 vCPU / 2 GB, an always-on worker at 0.5 vCPU / 512 MB, Postgres with 10 GB of storage, running the full 730-hour month. Every number below is computed from published unit rates or list prices current in mid-2026, and every assumption is stated so you can re-run the math.

How each row is built:

• Railway: $20/vCPU/month and $10/GB RAM/month, minute-metered, per Railway's pricing as quoted in Northflank's own August 2026 comparison. Web: 20 + 20 = $40. Worker: 10 + 5 = $15. Postgres as a template service shaped 0.5 vCPU / 1 GB plus a 10 GB volume at $0.15/GB: 10 + 10 + 1.50 = $21.50. Usage totals $76.50; the $5 Hobby plan covers the first $5, so the bill lands at ~$77.
• Northflank Cloud: $0.01667/vCPU/hour and $0.00833/GB/hour, per-second billing, no seat fees. Web: (0.01667 × 730) + 2 × (0.00833 × 730) ≈ $24.35. Worker ≈ $9.15. Postgres shaped 0.5 vCPU / 1 GB at the same unit rates plus a volume ≈ $15 — computed from published rates, not a list price, since Northflank prices resources rather than database tiers.
• Render: Standard web service $25 (1 CPU / 2 GB), Starter worker $7, Basic-256mb Postgres $6, on a $0 Hobby workspace. Render's own comparison confirms the shape: flat-rate instances beat fully-utilized metered compute at these sizes.
• DigitalOcean App Platform: Professional-xs container $12 (dedicated CPU, the tier DO recommends for production), a Basic $5 container for the worker, managed Postgres Basic $15. Flat, predictable, no BYOC.
• Northflank BYOC: you pay your cloud provider directly for the infrastructure, and Northflank charges a disclosed platform meter of $0.01389/vCPU/hour and $0.00139/GB/hour. For this workload that meter is (10.14 + 2.03) + (5.07 + 0.51) + (5.07 + 1.01) ≈ $24/month, before a cent of AWS/GCP/Azure spend — which runs another ~$25–40 for the same shape. The meter is the price of the control plane; keep that number in mind, because the whole post converges on it.
• Fly.io: shared-cpu-2x with 2 GB ≈ $12, a shared-cpu-1x 512 MB worker ≈ $3.50, Postgres on Machines ≈ $7 plus ~$2 in volumes. Caveat, stated plainly: shared CPUs, not dedicated — the row earns its low price with a weaker unit.
• Heroku: Basic web dyno $7, Basic worker dyno $7, Postgres Mini $5. Same kind of caveat, louder: Basic dynos carry 512 MB of RAM against the 2 GB everywhere else. Price the Standard-1X path for closer parity and the row jumps to ~$55. Heroku is paid-only and has no BYOC story, exactly as the roundup says.
• Hetzner + Coolify: a 4 vCPU / 8 GB cloud instance at roughly €12/month runs all three workloads on one box with headroom. (The entry ~€5–6/month 2 vCPU / 4 GB size exists but is tight once web, worker, and Postgres share it; €12 is the honest all-in figure.) No per-vCPU meter, no per-GB RAM meter, terabytes of included transfer, no control-plane fee. The cost you do pay is your own maintenance time.

Two rows deserve their asterisks repeated: Fly.io and Heroku win on price partly by selling a smaller unit. Everything else in the table is the same workload, and the ranking it produces — metered managed highest, flat managed middle, shared/small cheapest among rentals, owned box cheapest overall — is the billing shape showing through, not a sale.

Four billing shapes explain the whole table​

Strip the logos away and every platform here charges in one of four ways:

1. Plan minimum plus metered consumption (Railway). $5–20/month buys a seat at the table and an equal usage credit; every vCPU-hour and GB-hour above it meters at $20/vCPU and $10/GB-month. This shape punishes always-on and rewards spiky: an idle service costs nearly nothing, a saturated one costs more than the flat equivalent. Render's comparison makes the point with Northflank's own numbers agreeing — a fully-used 1 vCPU / 2 GB service is $40 metered against $25 flat.
2. Flat per service (Render, Heroku, DigitalOcean). You rent named sizes at named prices. Predictable to the dollar, wasteful at low utilization, and the cheapest honest way to run always-on among rentals — DO's $32 row is the table's value winner for full-time workloads on dedicated resources.
3. Per-second metered with no plan floor (Northflank Cloud, Fly.io). Same incentive as shape 1 but finer-grained and, on Fly.io, paired with scale-to-zero auto-stop. Northflank's Cloud rates ($0.01667/vCPU/hr, $0.00833/GB/hr) work out to roughly $12/vCPU and $6/GB per month — cheaper per unit than Railway's meter, which is why its row lands $28 lower for identical shapes.
4. BYOC: your cloud bill plus a control-plane meter (Northflank BYOC). The infrastructure invoice goes to AWS/GCP/Azure — where your credits, commitments, and negotiated rates still apply — and the platform takes ~$24/month for this workload on top. This is the shape the roundup is really selling, and it is genuinely better than shape 1 for teams with existing cloud commits. But notice what it is: a second meter layered over the first. You have escaped the PaaS compute markup and kept a PaaS control-plane subscription.

The owned box is the zeroth shape: fixed rent, no meter at all, maintenance time as the only variable cost. Nobody in the roundup prices it, including Coolify's own section — which is the tell. A comparison written by a BYOC vendor will walk you right up to "bring your own infrastructure" and stop one step short of "own the machines," because its meter only survives the comparison if the zeroth shape stays off the table.

Sensitivity: utilization flips the winner​

A single always-on data point would be cherry-picking if the answer didn't move with the variables, so here are the two that matter.

Utilization. Run the same workload as a side project at ~15% average utilization and Railway's usage falls to roughly $11.50 — the $5 plan plus a few dollars of overage — while Render still bills $38 flat and Hetzner still bills €12. The metered shape wins idle; the flat and owned shapes win busy. The breakeven for this workload sits around 40–50% utilization: below it, Railway is the cheapest managed option in the table; above it, every flat row passes it.

Know your duty cycle before you pick a shape — the roundup's "credit-free uptime" axis is really a proxy for "my app is always on," and always-on is exactly where metering loses.

Scale. Add four more web services at the same shape and watch the meters multiply: Railway climbs past $230, Northflank Cloud past $145, Render to $138, DO to $80 — while the Hetzner row stays €12 until the box fills, then steps to the next size, not the next multiple. Egress sharpens the same point: Railway meters it at $0.05/GB, Render bundles 100 GB against workspace limits, and the owned box includes terabytes. The more services and the more traffic, the more every metered row looks like a tax on growth and the flat box looks like a cap on it.

The honest summary across both variables: metered wins for idle experiments, flat wins for steady small production, and owned wins for steady production at any scale past one box — with maintenance time as the real price tag on the last one, not a footnote.

What the roundup concedes, and the step it stops short of​

Credit where due: the roundup's non-price axes are substantive. Railway's Nixpacks ceiling, basic secrets story, young volumes, and second-class workers are fair criticisms, and its comparison table correctly marks Render, Heroku, and DO as having no BYOC path at all. Its advice on choosing — 24/7 uptime needs, deploy visibility, exit cost — is the right checklist. And the Upwork Lifted case study it cites, running production inside its own VPC under Northflank, is real evidence that self-serve BYOC without a sales process exists in 2026.

But follow the money the roundup never totals. Northflank's BYOC meter on this workload is ~$24/month — assessed on infrastructure you already pay AWS or GCP for, every month, scaling with every vCPU you add. Against the ~€12 all-in owned box, the control plane costs double the computers. That gap is the price of never touching Kubernetes yourself: governed self-service, per-second billing, log sinks, and someone else's pager.

For a regulated team spending its existing cloud commits inside its own VPC, it can be worth it. For a team whose infrastructure is a handful of services that fit on machines it could own outright, the roundup's own logic — no credit shutdowns, full visibility, no lock-in, predictable pricing — points past every row in its table to the shape it never priced.

That is the one step short. "Bring your own cloud" still rents the cloud and rents the control plane. Own the machines and there is no meter left to optimize — only the box, the deploy pipeline, and your time.

Running the numbers on leaving metered PaaS bills behind? Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own, with no per-vCPU meter and no control-plane subscription. Star the repo on GitHub or deploy your first app today.