Everyone told me to move to the cloud. I went the other direction.
In 2024 I documented something I'd been quietly running for years: a network of
35+ web domains, all served from a single Intel i5-750 — a processor released in 2009.
Peak response time: 784 microseconds.
Monthly cloud bill: €0.
Here's the architecture that makes it possible.
The Core Problem with "Just Use Cloud"
Cloud providers sell convenience. What they don't advertise is that convenience
has a compounding cost — especially for workloads that are read-heavy, relatively
static, and don't need elastic scaling.
My infrastructure is exactly that: 35 domains spanning IT consulting, local SEO,
aerospace research, and cybersecurity. High content, low write frequency.
A perfect candidate for aggressive caching.
The Stack
Origin: Intel i5-750 (2009), running 24/7, CPU load < 1%
Edge layer: Cloudflare Workers (free tier)
The Worker handles:
- Reverse proxy with 1.5s abort timeout (kills slow requests before Cloudflare times out)
- Cookie stripping (reduces cache fragmentation)
- Anti-flood filtering (path/query length > 100 chars → reject)
- Failover to immortalHTML (semantic HTML with full JSON-LD Person schema)
- robots.txt and llms.txt serving
- Subdomain routing to Soloist.ai microsites
Protocol: HTTP/3 end-to-end where supported
Compression: Zstandard (zstd) — significantly better ratio than gzip at same CPU cost
Cache efficiency: 96.01% bandwidth served from cache (Cloudflare edge),
meaning the origin server handles less than 4% of all requests.
The immortalHTML fallback isn't just an error page — it's a fully valid HTML
document with Schema.org Person JSON-LD, all sameAs links, and CTA elements.
Visitors and Googlebot both get a useful response even when the origin is offline.
The Numbers
| Metric | Value |
|---|---|
| Peak latency | 784μs |
| Sustained latency (35 nodes) | 884μs |
| Cache hit rate | 96.01% |
| CPU load (origin) | < 1% |
| Monthly cloud cost | €0 |
| Hardware age | 17 years |
What This Proves
The bottleneck in most web infrastructure isn't the CPU — it's the architecture.
A well-cached, edge-proxied static-heavy site on a 17-year-old processor
outperforms most cloud VPS setups on latency metrics because the origin
is almost never involved.
The full audit (methodology, raw data, tools used) is published on Zenodo:
https://doi.org/10.5281/zenodo.20706458
https://medium.com/@sterpimarco
I'm happy to share more about the Worker logic, the caching strategy,
or the JSON-LD setup for Knowledge Graph targeting.
Top comments (0)