From $0.01-a-Day HTML to Per-Second Billing: How Website Hosting Splits in 2026

Two things happened in the same stretch of July 2026: a longtime independent-web advocate renewed the case for running a website with nothing but a text editor and a $0.01-a-day host, and three of the largest deployment platforms kept publishing pricing pages built around metering compute down to the second. Neither side is wrong. They're answering different questions.

Neatnik's Hardcore IndieWeb Makes the Case for Static HTML and $0.01-a-Day Hosting

In a guide published July 18, 2026, Adam Newbold argues that real independence on the web requires two things: a copy of your published HTML sitting on your own hard drive, and a website that exists in that same publishable form at all times. His recommended stack is deliberately minimal — a text editor, a file-transfer tool, and a web host — with no framework, build step, or subscription in between. For hosting, he points to NearlyFreeSpeech.net, a pay-as-you-go static host he has used since 2008, which he says can run a small site for roughly a cent a day with an account fundable for as little as a quarter.

The argument isn't that frameworks or managed platforms are bad. It's narrower: that every layer added between a person and their published files is a layer of dependency, and that the fastest way to make a website portable is to keep it in a format — plain HTML — that needs no service to interpret it. That's a meaningfully different environment than the one Railway, Vercel, and Cloudflare are building for, and the gap between them is easier to see with actual numbers next to each other.

Railway and Vercel Now Bill by the Second Instead of the Box

Railway's pricing page frames its entire model around a single idea: charge for what an app actually uses, not for a pre-sized box sitting idle. The Free tier gives new accounts a 30-day trial with $5 of credit, after which usage reverts to about $1 a month for a small service. The Hobby plan includes $5 of monthly usage credit and then charges $0.00000772 per vCPU-second and $0.00000386 per GB-second of memory once that credit runs out. Pro raises the included credit to $20 and the resource ceiling to as much as 1,000 vCPU and 1 TB of RAM per service. Customer quotes on the pricing page describe monthly bills dropping from the thousands on Heroku or AWS to a few hundred dollars after switching, though those are self-reported figures from named customers rather than independently audited numbers.

Vercel's pricing page runs on a similar logic but a different unit. The Hobby tier is free and includes 4 hours of Active CPU time and 1 million Edge Requests a month. Pro costs $20 per user per month, bundles $20 of usage credit, and then meters "Fluid compute" by Active CPU time — the seconds a function is genuinely executing, not the seconds it's alive and waiting — starting at $0.128 per hour once credit is exhausted. Both platforms are pitching the same pitch Railway makes to its Heroku-refugee customers: stop paying for provisioned capacity you aren't using.

Cloudflare's Workers Pricing Charges for CPU Time, Not Wall-Clock Waiting

Cloudflare's homepage makes the same wall-clock-versus-CPU-time distinction Vercel makes, but pushes it further as a headline pitch: the company says it charges only for compute, not for time a Worker spends waiting on a slow API, an LLM call, or a hibernating WebSocket. The Workers free tier includes 100,000 requests a day, 1 million KV reads, and 10 GB of R2 storage. The paid tier starts at $5 a month and then meters usage directly — $0.30 per million requests, $0.02 per million CPU-milliseconds, $0.15 per million Durable Object requests. Cloudflare's separate site-security bundle, aimed less at compute and more at CDN, WAF, and DDoS protection, runs Free, $20-a-month Pro, and $200-a-month Business tiers, with custom Contract pricing above that.

The distinction matters operationally. A function that calls out to a slow third-party API and waits two and a half seconds for a response costs very little on Cloudflare's model if it's only doing a few milliseconds of actual computation during that wait — but a similarly idle function on a platform billing wall-clock time pays for the entire wait. As AI-agent workloads that spend most of their time waiting on model responses become more common, that distinction is becoming a real cost variable rather than a footnote, a shift that runs alongside Anthropic's own moves to tighten its grip on the agent ecosystem.

Where Google's Stitch Fits Into a Faster Front End

Stitch, Google's AI-assisted UI generation tool, sits at a different layer than any of the hosting platforms above — it's aimed at the design step before code exists, generating mobile and web interface ideas for iteration. Its own landing page is a client-rendered application shell that returns almost no static content: no pricing table, no technical specification, no usage limits visible without signing in. That's worth naming plainly rather than working around. Where the hosting platforms publish detailed, line-item pricing, Stitch's public page currently discloses only that it exists, what problem it's aimed at, and that it's in beta. Anything more specific about cost or capability would have to come from actually using the product, not from what's published.

The Spectrum From Hand-Written HTML to Serverless Edge Compute Is Wider Than It Looks

None of this makes the Hardcore IndieWeb approach and the metered-cloud approach competitors in any real sense. A static blog authored in a text editor has no background jobs, no edge functions, and no reason to pay for CPU-second billing at all — the entire premise of Neatnik's guide is that most of what modern platforms charge for, a simple site never needed. The metered platforms, in turn, exist for the workloads static HTML can't touch: databases, authenticated backends, and the kind of framework-driven interactivity covered in the broader case for moving content-first sites toward Astro's static-by-default model. What's changed by July 2026 is less the existence of that spectrum than how precisely each end of it now prices itself — down to the byte for static hosting, and down to the CPU-millisecond for everything built on top of it, a pattern of squeezing runtime cost that also shows up in efforts like Bun's ground-up rewrite of a million lines of Zig and Rust aimed at cutting overhead lower in the stack.