Railway's $100M Bet: What Per-Second Pricing Really Costs When the Container Never Sleeps

Railway just raised $100 million to bury AWS — and the same always-on API, worker, and Postgres that costs about $148/month on Railway's per-second meter runs on two flat-rate Hetzner boxes for roughly $20. That 7x gap is the lead, not the footnote, because the raise changes nothing about the meter: every second the container exists still bills at $0.00000772 per vCPU-second and $0.00000386 per GB-second, with no idle discount. There is a shape where Railway wins — about three hours a day or less — and this post works it, tests the "50% under hyperscalers" claim against an actual AWS baseline, and then reads the asterisk file, including the May 2026 outage that undercuts the invincibility narrative.

The raise in 30 seconds​

On January 22, 2026, Railway announced a $100 million Series B led by TQ Ventures, with FPV Ventures, Redpoint, and Unusual Ventures participating. The numbers in the announcement: two million developers, 100,000 paid users across 25,000 businesses, and sub-second deployments from a platform that runs its own rack-scale hardware design and bare-metal orchestration — no Kubernetes underneath — across four colocation presences (two in the US, Amsterdam, Singapore). The pitch, per CEO statements carried by VentureBeat: deployment is the bottleneck now that AI writes code in seconds, and Railway's vertically integrated metal lets it price roughly 50% under the hyperscalers while staying online through outages that took down the big clouds.

Two of those claims can be checked with arithmetic. One can be checked against the public record. Let's do all three.

The meter, decoded​

Railway bills usage per second on top of a plan fee (Hobby $5/month including $5 of credit, Pro $20/month plus usage). The unit rates:

The translation to remember: an always-on vCPU costs $20/month and an always-on GB costs $10/month. Per-second billing with no idle discount is mathematically identical to monthly billing for anything that never sleeps. The meter only discounts one thing: absence.

The worked recompute: one topology, three bills​

Reference workload, kept deliberately boring: a steady-traffic API (1 vCPU, 2 GB), a background worker (1 vCPU, 2 GB), and Postgres (1 vCPU, 2 GB, 20 GB volume), running 24/7 for 30 days, 100 GB egress. List prices observed September 2026; re-check live pages before budgeting.

Two verdicts fall out of that table. First, the Hetzner leg: the identical topology costs about one-seventh of Railway — roughly $128/month kept, every month, forever, with no promotional period to expire. Second, the AWS leg: on this small always-on shape, Railway's "50% under hyperscalers" claim doesn't just miss — it inverts. $148 versus ~$120 of Fargate-plus-RDS list price puts the meter about 20% above the hyperscaler, before AWS Savings Plans (which would widen the gap further) and before Multi-AZ (which would narrow it). The 50% claim may well hold on larger shapes where owned metal beats hyperscaler list harder — but on the starter topology most teams actually run, it is Railway's math to show, not yours to assume.

When the meter wins: the three-hour rule​

None of that means per-second billing is a bad deal. It means it is a conditional deal, and the condition is quantifiable. The three-service topology above burns about $0.167 per hour on the meter (3 vCPU × $0.0278 + 6 GB × $0.0139). Against a ~$20 flat fleet on a $5 Hobby plan, breakeven lands at roughly 90 running hours a month — about three hours a day:

• Preview environments that live for the length of a PR review: dozens of container-hours a month, pennies on the meter, full price on a box you sized for peak.
• Bursty builders and cron-shaped workers that run minutes at a time: the meter charges minutes; the box charges the month.
• Idle side projects you touch on weekends: ~$2–5/month metered versus the same flat $20 whether you git push or not.

The rule of thumb: if the whole stack averages under ~3 hours a day, Railway's meter wins. Past that, the flat box wins, and the gap compounds every month. Steady production traffic — the API that serves users at 3 a.m. because users exist in other time zones — is the worst case for a no-idle-discount meter by construction.

The asterisk file​

Three footnotes belong next to any "challenge AWS" narrative:

The invincibility claim has a May 2026 exception. Railway's "remained online through hyperscaler outages" line dates to earlier incidents. But on May 19–20, 2026, Google Cloud suspended Railway's own account without warning, and because Railway's edge proxies depended on a control-plane API hosted in GCP, the failure cascaded onto workloads in Railway's Metal and AWS environments — dashboard, API, builds, deploys, and eventually active routing all went down, with non-enterprise deploys paused for hours. Your compute being healthy is not the same as your compute being reachable, and reachability was Google's to revoke. Owning the metal demonstrably did not mean owning the blast radius.

There is no idle discount, and egress is metered. Containers that sit at 1% CPU overnight bill the same as containers at 90% — the meter charges existence, not work. Add $0.05/GB egress on top, and a media-serving or webhook-relaying workload pays a second meter the flat box's 20 TB of included transfer never notices.

Credits are a timer, not a tier. Trial credits cover the meter for months and then stop covering it all at once — the month-13 bill for the topology above is the full ~$148, not a glide path. Budget the run-rate, not the runway.

What $100M of venture funding means for your bill​

One more line belongs on the sheet: the $100 million itself. Venture funding is a temporary subsidy on the meter — it pays for the four colocation presences, the custom hardware program, and the go-to-market motion, which means today's per-second rates do not have to carry the full cost of the company yet. That cuts both ways. In the short term, the meter is the best price it will ever be; in the long term, a Series B priced on "challenge AWS" growth has to return that growth, and there are exactly two levers — more seconds billed, or more dollars per second. Lock-in to a meter you cannot audit against owned hardware is how the second lever gets pulled quietly. Price the workload against the flat box now, while the subsidy is flowing, so you recognize the month the math changes.

What "AI-native" actually changes about the bill​

Here is the honest verdict the title owes: "AI-native" changes the speed of the bill, not the shape of it. Sub-second deploys, agent-friendly workflows, and a platform agent billed at pass-through token cost are real velocity gains — an agent that provisions ten preview environments an hour gets ten bills an hour, each metered to the second. But the meter still charges for every second the container exists, and AI agents are exceptionally good at provisioning containers nobody remembers to delete. Railway's own economics give this away: zero margin on agent tokens, full margin on the compute the agent spins up. The AI-native cloud monetizes exactly what agents over-provision.

That is not a reason to avoid Railway. It is a reason to match the workload to the meter: bursty, ephemeral, agent-spawned work belongs on per-second billing; the steady-state API, worker, and database that never sleep belong on hardware whose price was settled before the month began.

Prefer the flat part of that sentence? 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-second meter on containers that never sleep. Star the repo on GitHub or deploy your first app today.

Sources​

• VentureBeat, "Railway secures $100 million to challenge AWS with AI-native cloud infrastructure" (Jan 2026): Series B amount, TQ Ventures lead, 2M developers, sub-second deploys, ~50%-under-hyperscalers pricing, outage-resilience claims.
• PR Newswire / Morningstar, "Railway Raises $100 Million Series B" (Jan 22, 2026): investor list (TQ, FPV, Redpoint, Unusual), "infrastructure invisible" framing.
• TechTarget, "Upstart cloud provider Railway turns heads with speed" (Jan 2026): 100,000 paid users, custom rack-scale hardware, bare-metal orchestration without Kubernetes.
• The Register, "Google Cloud suspended major customer Railway.com without cause, causing outage" (May 20, 2026): GCP account suspension, cascade to Metal/AWS workloads, deploys paused.
• The Register, "Railway pauses lowest tiers after onrush of GCP customers" (Jun 2025): four colo presences, prior GCP-outage exposure.
• Railway community pricing references (2025–2026): $0.00000772/vCPU-second, $0.00000386/GB-second, $0.00000006/GB-second volumes, $0.05/GB egress; $20/vCPU/month + $10/GB/month equivalents.
• Hetzner cloud pricing references (2026): shared CX-class 2vCPU/4GB from ~€3.79–4.75/month, 4vCPU/8GB CX32-class ~€6.80–9/month; ARM CAX alternatives ~€6/month.