DEV Community

niuniu
niuniu

Posted on

I Stopped Paying $25/Month for API Hosting — Cloudflare Workers Runs My Backend for $0

Last month I got a $25.40 invoice from my VPS provider for hosting a tiny JSON API. The API serves ~180k requests/month, averages 12ms of CPU time per request, and sits idle 96% of the day. I was paying for a server that does almost nothing.

So I moved the whole thing to Cloudflare Workers. New monthly bill: $0.00.

The numbers before and after

Item VPS (Hetzner CX11) Cloudflare Workers Free
Monthly cost $4.51 + $20.89 bandwidth $0
Requests included unlimited (but I pay for idle) 100,000/day
Cold start n/a (always on) ~0ms (V8 isolates)
p50 latency (US) 140ms 18ms
p99 latency 310ms 45ms
My monthly usage 180k requests 180k requests (60% of free quota)

The latency drop surprised me most. Workers run in 300+ edge locations, so my users in Frankfurt and Singapore stopped round-tripping to a single box in Falkenstein.

The actual migration (took me one evening)

My API is a thin proxy that validates a token, queries a KV store, and returns JSON. The entire worker:

export default {
  async fetch(request, env) {
    const url = new URL(request.url);

    const token = request.headers.get("X-API-Key");
    if (!token || !(await env.API_KEYS.get(token))) {
      return Response.json({ error: "invalid key" }, { status: 401 });
    }

    const key = url.pathname.slice(1);
    const cached = await env.CACHE.get(key, "json");
    if (cached) {
      return Response.json(cached, {
        headers: { "X-Cache": "HIT" },
      });
    }

    const fresh = await fetchUpstream(key);
    // Cache for 5 minutes, fire-and-forget
    ctx.waitUntil(env.CACHE.put(key, JSON.stringify(fresh), { expirationTtl: 300 }));
    return Response.json(fresh, { headers: { "X-Cache": "MISS" } });
  },
};
Enter fullscreen mode Exit fullscreen mode

Deploy is one command:

npm create cloudflare@latest
npx wrangler deploy
# Free tier: 100k requests/day, 10ms CPU per request, Workers KV included
Enter fullscreen mode Exit fullscreen mode

What I gave up (be honest)

  • No long-running processes. 10ms CPU limit on free tier (30s on paid). My image-resize endpoint had to move to a queue.
  • No raw TCP. No direct Postgres connection — I use Neon over HTTP or Hyperdrive.
  • Node compatibility is good but not perfect. Two npm packages (bcrypt, one ORM driver) needed swaps.

The controversial part

I now think renting a whole VPS for a low-traffic JSON API in 2026 is a legacy habit, not an architecture decision. Between Workers (100k req/day free), Vercel Functions, and Railway's trial, the "I need a $5 droplet" instinct costs the indie dev scene millions a year in aggregate — for compute that sits idle 96% of the time.

If your API does under ~3M requests/month and is CPU-light, you're probably donating money to Hetzner too.

When I'm prototyping these migrations I lean on MonkeyCode (free, open-source AI coding assistant) to generate the worker boilerplate and wrangler config — saved me the docs-diving: https://ly.cyberserval.tech/iIETXiF

What's your cutoff? At what request volume does a real server actually become cheaper than serverless for you?

Top comments (0)