We Ran 17 Apps on a 5 Euro Server and Load-Tested It
The pitch for running your own server is always the same: one cheap box, a stack of apps, a bill that does not move. The question nobody answers with numbers is what happens when real traffic hits that box.
So we measured it. A single Hetzner cx23 (2 shared vCPU, 4 GB RAM, 5.49 euros a month before VAT and the IPv4 fee) running 17 apps at once: Next.js, Astro, SvelteKit, Express, Hono, Vite, Docusaurus, and the Vercel commerce, chatbot and platforms templates. All live, nothing stopped for the test. The box was already using about 750 MB of swap when we started.
How we tested
The load generator ran on a separate machine and hit the apps through a custom domain, through Cloudflare, with HTML deliberately not cached at the edge. Every request reached the server.
Two apps are the interesting ones: a Next.js 16 server-rendered app (the official dashboard tutorial) and a Hono API on Bun.
Next.js 16, server-rendered
| Concurrency | Requests/s | P50 | P95 | Errors |
|---|---|---|---|---|
| 50 | 294 | 160 ms | 263 ms | 41 of 13,022 |
| 100 | 343 | 272 ms | 403 ms | 144 of 15,702 |
| 200 | 351 | 304 ms | 2,617 ms | 0 of 16,032 |
| 500 | 362 | 968 ms | 3,711 ms | 74 of 16,461 |
At 10 concurrent connections the same app's login page served 264 requests a second with a 36 ms median.
Read the shape, not any single row. Throughput flattens around 350 requests a second from 100 concurrent connections on. Adding concurrency past that only adds queueing: the median grows, and the 95th percentile goes from under 0.4 seconds to several seconds. The login page, which does not hit the database, returned zero errors even at 500 concurrent connections.
350 requests a second, sustained, is about 30 million requests a day. On a 5 euro server.
Hono on Bun
| Concurrency | Requests/s | P50 | P95 |
|---|---|---|---|
| 50 | 820 | 55 ms | 81 ms |
| 100 | 786 | 55 ms | 98 ms |
| 200 | 795 | 56 ms | 105 ms |
| 500 | 835 | 361 ms | 415 ms |
This is not the app's limit. The load generator was pinned at 100% CPU while the server sat about half idle. Treat 800 requests a second as a floor, not a ceiling.
What actually limits you
One Next.js process uses one CPU core. The box has two. During the test the Next.js process sat at about 90% of one core while the second core had headroom. A single-process server-rendered app runs out of its core long before the server runs out of capacity, and the fix is a second process, not a bigger box.
The caveats, up front
- These are small template apps. They do not run heavy queries or hold large data in memory. An app with slow database queries will do less.
- The errors in the first table are the server resetting connections under queueing, not a bug in the app.
- The numbers are one server at one point in time, through Cloudflare, on shared vCPU. Your workload will differ.
- This is a measurement, not a benchmark suite.
What this means for you
If you run a handful of client apps or a small SaaS on managed platforms, you do not need most of what you are paying for. One inexpensive server and one deploy tool will serve real traffic for a flat monthly cost. The ceiling is real, and it is higher than you think.
Want a number for your own apps? The server size recommender and the savings calculator take a minute each.