People searching “railway vs aws” are almost always asking a price question. It is the wrong question, and the fastest way to show that is to answer it properly first.
So here is the answer. Price the same small always-on service on both platforms and the gap is either sixteen dollars a month or fifty-eight cents, depending on which AWS you mean. Then read the incident report Railway published in May 2026, and the actual difference between these two platforms stops being about money at all.
Price the same box on both and the gap nearly closes
Take one container that stays up all month with one full vCPU and 2 GB of memory. That is a modest API with a database behind it, the shape most people are actually deploying.
Railway bills by the second. Its published rates are $0.00000772 per vCPU-second and $0.00000386 per GB-second (Railway pricing). Over a 730-hour month that is $20.29 for the vCPU and $20.29 for the two gigabytes, so $40.58 of metered usage. The Pro plan is $20 per month per workspace and includes $20 of usage credits (Railway pricing), so the bill lands at about $40.58. Egress adds $0.05 per GB.
AWS App Runner is the AWS service closest in shape, a managed container you point at a repository. In US East it charges $0.064 per vCPU-hour and $0.007 per GB-hour for active instances (AWS App Runner pricing). Same box, same month: $46.72 plus $10.22, so $56.94, before the build fee the same page describes.
Now the part that undoes the comparison. AWS Lightsail sells a container service plan called Medium with 1 vCPU and 2 GB for a flat $40 per month, with 500 GB of data transfer included (AWS Lightsail pricing). That is Railway’s number, from AWS, with the egress thrown in.
So AWS is 40 percent more expensive than Railway and also within fifty-eight cents of the same price, and both statements are true at once. Which one you get depends entirely on whether you walk into AWS through the door built for people who did not want to make these decisions.
The dramatic savings figures attached to Railway are real but they are not per-unit price differences. Daniel Lobaton of G2X told VentureBeat his infrastructure bill fell from $15,000 a month to about $1,000 after moving, and Kernel’s chief technology officer said the company runs its entire customer-facing system on Railway for $444 a month (VentureBeat). Those are stories about replacing an over-provisioned architecture, not about the rate card. If your AWS bill is mostly load balancers, NAT gateways and instances sized for a traffic peak you no longer have, any platform that forces you to declare what you actually run will look like a miracle.
What you are choosing is the number of decisions
Railway gives you one deployment path. You connect a repository, it builds a container, it runs. There is no decision to make about which of several compute services fits, because there is only one.
AWS gives you a catalog and expects you to assemble. That is a genuine strength when your requirements are specific and a genuine tax when they are not. The strongest evidence for that tax is that AWS sells Lightsail at all: it is the company’s own product for customers who want the assembly done in advance, priced in flat monthly tiers rather than in vCPU-hours (AWS Lightsail pricing).
That is worth saying plainly because it is the thing a small team is really buying. A two-person shop running production on a platform with one path is a team doing work that used to need a platform engineer on staff to make possible at all. The gain is reach, not a smaller payroll, and it is the same pattern we found in AI Coding Tools: What the Slowdown Study Actually Says, where the version that holds up is a two-person shop taking on five-person projects.
The eight hours that tested the entire premise
Railway raised a $100 million Series B led by TQ Ventures, announced on January 22, 2026, with FPV Ventures, Redpoint and Unusual Ventures participating, bringing the company to $124 million raised against a team of 30 people (VentureBeat). The argument behind the round was independence. Railway left Google Cloud in 2024 to build and run its own hardware, now four regions of it in California, Virginia, Amsterdam and Singapore (Railway Metal documentation). Owning the metal cut its own published egress price in half, from $0.10 per GB to $0.05, and disk from $0.25 per GB to $0.15 (Railway Metal documentation).
Four months after the raise, on the night of May 19, 2026, Google Cloud placed Railway’s production account into a suspended status as part of an automated action, and Railway went dark for about eight hours (Railway incident report).
The mechanism is the part to read twice. Railway’s dashboard, API and parts of its network infrastructure still lived inside Google Cloud, and so did the network control plane. Customer workloads on Railway Metal and on AWS kept running. Then the edge proxies, which hold a cached copy of the routing tables from that control plane, hit cache expiry. With no control plane to refresh from, the edge could no longer resolve routes to instances that were sitting there perfectly healthy, and requests to every region started returning 404 (Railway incident report). Railway committed to removing the hard dependency afterward, spreading high availability database shards across AWS and Metal and taking Google Cloud services out of the data plane’s hot path (Railway incident report).
Here is why that matters more than any rate card. Leaving a hyperscaler is not a switch you flip. Railway had already moved the expensive, visible half, the compute, onto hardware it owns. It went down anyway because the small, cheap, invisible half had not moved, and that half is the one that decides whether anything is reachable. Any platform sitting between you and a data center has a control plane somewhere. The useful question is not who owns the servers. It is what happens to your traffic when the thing that knows where your servers are cannot be reached.
Where Vercel sits, since that is the other comparison people run
Vercel is not a cheaper or more expensive Railway, it is a different shape. Its Hobby plan is $0 per month and Pro is $20 per month per developer seat (Vercel pricing), but the included usage is counted in function invocations, edge requests and data transfer rather than in vCPU-months, because the product assumes your code mostly runs in response to requests and then stops.
If what you have is a Next.js frontend with API routes, that model will be cheaper than renting a container that idles all night. If what you have is a process that must stay resident, a queue worker, a websocket server, a database you would rather not hand to a managed service, then Railway and the AWS container products are the real comparison and Vercel is not in it. Comparing the monthly totals across those two shapes tells you nothing.
Three checks before you pick
1. Price your actual shape, not a generic instance. Run docker stats against your service under normal load, or open your current provider’s billing page, and write down the vCPU and memory you genuinely use. Multiply that by each platform’s published rate rather than comparing plan names. The exercise takes twenty minutes and it is the only version of this comparison that is about your application.
2. Read the last two incident reports, not the uptime number. Railway publishes its post-incident writeups openly on its blog, and AWS publishes post-event summaries through the AWS Health Dashboard. You are looking for one thing: when it broke, did the failure stay inside the component that failed, or did it take everything with it. Uptime percentages average that distinction away.
3. Check the exit before the entry. If your application is a container with a Dockerfile and its state lives in a managed database, moving between any two of these platforms is a weekend. If it has grown into a specific provider’s queues, identity system and event triggers, moving is a quarter. Decide which of those you are building on the day you pick, because that is the day the choice is cheap.
Our read, and it is a read
The price comparison is close enough that it should not decide anything, and treating it as the deciding factor is how teams end up on the platform that was 12 percent cheaper and 400 percent more work. Neither company publishes the numbers that would settle the reliability question properly, so what follows is judgment rather than a finding.
Railway’s May incident is a point in its favor, not against it. A 30-person company published a detailed account of being taken offline by an automated action at a vendor it was in the middle of leaving, named the exact dependency that turned a partial outage into a total one, and said what it would change. That is more information than most infrastructure buyers ever get. The failure mode it describes, a control plane in someone else’s account, is present at every platform of this kind, and Railway is the one that told you about its own.
The argument against Railway is not price and not the outage. It is concentration. Thirty people, four regions, $124 million raised, against a provider whose worst week is still someone else’s problem to fix. That is a real risk and it is a reasonable reason to choose AWS. It is just not the reason most people give, and it is not what “railway vs aws” usually means when someone types it.
Frequently asked questions
Is Railway actually cheaper than AWS?
For one always-on service with 1 vCPU and 2 GB, Railway’s published per-second rates work out to about $40.58 a month (Railway pricing), AWS App Runner to $56.94 at $0.064 per vCPU-hour and $0.007 per GB-hour (AWS App Runner pricing), and an AWS Lightsail Medium container plan to a flat $40 with 500 GB of transfer included (AWS Lightsail pricing). Cheaper than one AWS product, level with another. Prices move, so check the date beside the number, a habit we argued for in AI Software for Small Business: Check the Expiry Date.
Did Railway really go offline because of Google Cloud?
Yes. Google Cloud suspended Railway’s production account through an automated action on May 19, 2026, and the platform was down for about eight hours because the dashboard, API and network control plane were still hosted there; workloads on Railway’s own hardware stayed alive but became unreachable once the edge proxies’ cached routing tables expired (Railway incident report).
Does Railway run on its own hardware or on someone else’s?
Both, and the split is the interesting part. Compute runs on Railway Metal, hardware the company owns in California, Virginia, Amsterdam and Singapore (Railway Metal documentation), while the control plane sat inside Google Cloud until the May 2026 outage, after which Railway said it would spread its database shards across AWS and Metal and take Google Cloud out of the data plane’s hot path (Railway incident report).
Should I use Railway or Vercel?
It depends on whether your code runs continuously or on request. Vercel’s Hobby plan is $0 and Pro is $20 per month per developer seat, metered in invocations and data transfer (Vercel pricing), which suits a frontend with API routes. A resident process such as a worker or a websocket server prices better as a container on Railway or an AWS container product.
Before you commit either way, do one thing that costs nothing. Open the status page and the incident history of whichever platform you are leaning toward, and find the last total outage. If you cannot find one, that is not reassurance; it usually means the write-up was never published. Infrastructure spending is increasingly set by demand you do not control, a pattern we traced through storage in Why Hard Drives Are So Expensive When Nothing Is Scarce, and the one thing still fully in your hands is knowing how your provider behaves on its worst day before you find out live.
