Dora NodaA couple of months ago, Theo Browne released a video rating the best place to host a new application in 2026. He said nice things about Fly.io — and then concluded that of all the providers he watches, Fly.io was the one he was least confident would still be around by the end of the year. He also noted they hadn't announced a fundraise in several years.
Fly.io CEO Kurt Mackey opened his reply with two words: "Well, fuck." Then came the substance: the best financial months in company history, a fresh pile of raised money, a new iteration of their Sprites agent computers, and Docker's ex-CEO Scott Johnston stepping in as chief executive while the whole company pivots toward computers for agents. Fourteen minutes of reading, and the takeaway the hosting tier lists never grade is right there in the open: the question that matters about your platform isn't the letter grade — it's what happens to you when the platform changes its mind.
So here is the re-score, up front, in one paragraph. Graded on exit cost instead of developer experience: Hetzner plus an open-source PaaS wins outright because the bill is flat and you keep the machines; Fly.io is the cheapest metered exit at roughly $20 per terabyte; Railway and AWS sit mid-curve at $50 and $90; Render's $30 blocks land near $300 per terabyte past the free allowance; and Vercel is free until it suddenly isn't — a $240-a-year seat cliff for commercial projects, then $400 per terabyte past the first one.
The tables below substantiate every number, and no section of this post hands out a letter grade. That was the whole problem.
The same gigabyte, six different bills
Tier lists argue about developer experience — the dashboard, the deploy animation, the docs. Nobody argues about the gigabyte, so price it. Same solo developer, same outbound transfer, three volumes: a side project at 10 GB a month, a growing app at 100 GB, and a working business at 1 TB.
| Outbound transfer / month | Hetzner CX33-class box (~€10/mo flat) | Fly.io (~$0.02/GB) | Railway (~$0.05/GB) | Render (100 GB incl., +$30/100 GB) | AWS (~$0.09/GB) | Vercel |
|---|---|---|---|---|---|---|
| 10 GB | €0 extra | ~$0.20 | ~$0.50 | $0 | ~$0.90 | $0, Hobby (non-commercial only) |
| 100 GB | €0 extra | ~$2 | ~$5 | $0 | ~$9 | $0 Hobby at its ceiling — or $20/mo Pro seat the day the project earns a dollar |
| 1 TB | €0 extra | ~$20 | ~$50 | ~$300 in 100 GB blocks | ~$90 | $20 Pro (first TB included) |
Three things in that table never appear in a tier list. First, at 10 GB every platform costs essentially zero, which means rankings that argue about side-project pricing are arguing about nothing — the differences below a dollar are vibes. Second, Vercel has the nastiest shape in the industry: free at 100 GB, then a licensing cliff, not a usage slope. The Hobby tier is personal and non-commercial only, so the first dollar of revenue moves a solo dev from $0 to a $20-per-seat Pro plan — $240 a year before a single overage byte. Third, the fan-out past 1 TB is where exits actually die: another terabyte costs roughly €0 on the Hetzner box (inside its 20 TB included transfer), ~$20 on Fly.io, ~$50 on Railway, ~$90 on AWS, ~$300 on Render, and $400 on Vercel. A platform that looked free in the tutorial can charge twenty times its cheapest rival for the same traffic, and no letter grade captures a curve.
Bandwidth isn't the whole bill — compute, seats, and builds all count — but it is the least escapable one. You can optimize compute; you cannot negotiate with your users downloading your product. Grade the curve, not the demo.
Platforms pivot; your architecture pays
Every tier list grades the present. Exit cost is about the pivot, and the pivot always comes. The last four years are a metronome:
| Year | Pivot | Who paid |
|---|---|---|
| 2022 | Heroku kills free dynos entirely, citing fraud and abuse | Every side project that assumed free was structural |
| 2023 | Railway removes its free tier, leaving a $5 trial credit | Hobby backends that outlived the trial |
| 2024 | Fly.io ends free allowances for new organizations | New users who arrived on a tutorial's pricing |
| 2025–2026 | Heroku slides into "sustaining engineering mode"; Render free still sleeps after 15 minutes of inactivity | Anyone who built on the assumption of active development |
| 2026 | Fly.io bets the company on agent computers; the community asks out loud whether the apps-and-Machines platform winds down | Every Fly customer re-reading their architecture doc in July |
That last row is the live one. Mackey was explicit that Fly Machines and the PaaS features "aren't going anywhere," and explicit that the company's energy goes to Sprites — drive forking, Connectors, a block-device stack rebuilt to the studs. Both statements can be true at once, and the gap between them is exactly what exit cost measures: not shutdown, but drift.
Deprioritized platforms don't send you a breakup email. They slow the changelog, move the best engineers to the new thing, and let your stack become legacy while it still runs. The community thread asking whether the apps platform "starts winding down" isn't ingratitude — it's customers doing the math the tier list never did.
Note the asymmetry. When Heroku killed free, when Railway metered everything, when Fly.io pivoted to agents — the vendors announced, and the users migrated. The cost of leaving was always borne on one side. A ranking that doesn't ask who pays for the next pivot is a weather report that never mentions rain.
The exit-cost scorecard
Same six contenders, four axes the rankings ignore. Each cell answers one question: if this platform pivots next year, what do you lose?
| Axis | Vercel | Railway | Render | Fly.io | Hetzner + open PaaS |
|---|---|---|---|---|---|
| Egress curve | $400/TB cliff past 1 TB; seat cliff before it | $0.05/GB, linear and predictable | $30 blocks ≈ $300/TB; gentle until it isn't | ~$0.02/GB, cheapest metered exit | Flat: 20 TB included, €0 marginal |
| Pivot history | Hobby fenced to non-commercial; 2022 free-plan deprecation scare | Free tier removed 2023 | Free sleeps after 15 min; paid tiers stable | New-org free removed 2024; 2026 agent pivot in progress | Sells compute by the month; no platform layer to pivot |
| API portability | Proprietary surface (ISR behavior, Edge Config don't travel) | Proprietary API, single vendor | Public REST API, proprietary — but an open-source implementation of it exists | Public Machines API, single vendor; direction now openly debated | Render-compatible API with an open implementation; same manifests, any box |
| Failure modes | Serverless-duration bill shock is documented ($300 to $3,550 in one viral case) | Usage meters on everything; surprise comes as line items | Cold starts after sleep; bandwidth blocks stack silently | 2–5s cold image pulls; the live risk is strategic, not technical | You are the on-call; disks fail at 3am and nobody pages but you |
Read the rows, not the columns. Egress decides the ceiling of your bill. Pivot history tells you how often the vendor has moved the floor. API portability decides whether leaving means re-platforming or re-pointing. Failure modes decide whether the surprise arrives as an invoice, a cold start, or a strategy post. And hardware ownership underlies all four: on every rented platform, the machines belong to someone whose roadmap you don't control. On the Hetzner box, the machine is yours in the only sense that survives a pivot: you can take the image somewhere else.
Where the €10 Hetzner box lands
A CX33-class box — 4 vCPU, 8 GB of RAM, roughly ten euros a month — with 20 TB of included transfer, running an open-source Render-compatible PaaS, wins this scorecard and loses the tier list. It loses the tier list because tier lists grade day one: the dashboard polish, the zero-to-deployed stopwatch, the free tier. Day one is the one day you will never repeat. It wins the scorecard because every axis compounds in its favor over time: the bill stays flat while traffic grows, there is no licensing event at the first dollar of revenue, the deploy API has an implementation you can read, and when anything pivots — the vendor, the market, your own architecture — the machine and the image are already yours.
Now the honest counter-argument, because a scorecard that hides self-hosting's costs is just a tier list with the grades flipped. The box has exit costs too, and they are denominated in your time. Somebody patches the host, rotates the disks, and debugs the 3am outage, and that somebody is you. There is no managed database on the box — that is a deliberate non-goal of the open PaaS layer, not an oversight — so day one you are either operating Postgres yourself or renting it from someone with their own egress meter. If your app needs managed Postgres with point-in-time recovery before it needs anything else, rent the database and don't apologize. If your side project fits in Vercel's free 100 GB and will never earn a dollar, stay there and don't apologize either.
Self-hosting wins for the app in the middle: real traffic, real revenue, and a bill where the metered gigabyte is the largest line item. That app is paying the pivot tax every month, and the box deletes it.
Theo's verdict and Kurt's answer were both honest, and both incomplete. Theo graded survival odds without pricing the escape. Kurt defended the company's future without addressing what the pivot costs the customers who signed up for the old one. The summer's real lesson for anyone choosing where to host: ask the survival question first, price the exit second, and pick the platform whose answer to both is written in infrastructure you can keep — not in a letter grade you'll forget by fall.
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.
Sources
- Theo Browne, video rating the "best place to host a new application in 2026" (youtube.com/watch?v=yfxDdQo2cyI) — Fly.io praised, but named the provider he was least confident would survive the year; no announced fundraise in years. As quoted and linked in Kurt Mackey's reply.
- Kurt Mackey, "Turn And Face The Strange," Fly.io blog, July 24, 2026 (14-minute read) — "Well, fuck"; best financial months in company history; fresh fundraise ("raised a fuckload of money and didn't need more"); Sprites iteration (Sprite Block Device, drive forking, Connectors); Scott Johnston (ex-Docker CEO) incoming as CEO; company focus pivoting to computers for agents; Machines/PaaS "aren't going anywhere."
- Fly.io community thread, July 2026 — "Will the apps and machine platform start winding down?", citing the strategy post as the trigger.
- Egress pricing: AWS $0.09/GB first 10 TB; Vercel Pro $20/seat with 1 TB included, $40/100 GB overage ($400/TB), Hobby 100 GB and non-commercial only; Railway $0.05/GB; Fly.io ~$0.02/GB North America; Render 100 GB included per account with extra 100 GB blocks at $30 (per Render staff, community forum 2024); Hetzner cloud VMs 20 TB included transfer, CX23 ~€5.99/mo (2 vCPU/4 GB), CX33 ~$10.42/mo (4 vCPU/8 GB); dedicated machines uncapped.
- Pivot history: Heroku free-plan elimination, November 2022 (TechCrunch, "fraud and abuse"); Railway free-tier removal 2023 ($5 trial credit); Fly.io free-tier removal for new orgs 2024; Render free services sleep after 15 minutes inactivity; Heroku "sustaining engineering mode" coverage, 2026.
- Vercel bill-shock case: $300/mo to $3,550 after moving client-side requests server-side, serverless-function duration dominated (The Register, July 2025).
