This morning I asked Klaus to pull my AWS bill. Not because I was worried — I had a rough sense of what I was spending — but because I’d never actually sat down and added it all up across every service, every subscription, every line item since I started this whole thing in January.
The number surprised me. Not because it was high. Because it wasn’t.
The full accounting
Seven and a half weeks of running a fully autonomous AI agent across two machines, coordinating tasks, closing GitHub issues, monitoring Discord, running cron jobs at 3 AM, building tools, testing code, generating ideas — all of it.
Total out-of-pocket: $311.54.
Here’s where that breaks down:
- AWS: $0. AWS gave me $120 in credits when I signed up, and they’ve absorbed every EC2, S3, and data transfer charge so far. The gross usage is $78, all covered.
- Anthropic Claude Max: $300. One month at $100, one month at $200 after I upgraded.
- ricoordonio.com via Cloudflare: $11.54, including Washington state and Tukwila city tax.
- OpenClaw, Tailscale, N8N, GitHub: $0. All free or self-hosted.
That’s $5.90 a day. Less than a coffee.
The part that almost wasn’t free
The AWS story has an asterisk. In early March, the EC2 instance crashed.
Not a graceful restart. A full OOM panic — out of memory, kernel panic, auto-reboot. I was running OpenClaw¹ at around 530MB and N8N² at around 300MB on a t3.small with 2GB of RAM and no swap. There was nothing left. The kernel did the only thing it could.
The fix was two lines:
sudo fallocate -l 2G /swapfile
sudo swapon /swapfileAnd then a resize from t3.small to t3.medium, which doubled my RAM to 4GB and added about $15 a month to my eventual bill. I say eventual because the credits still cover it. Right now, sitting at 95% CPU idle and 45% RAM utilization, the instance is technically over-provisioned. But after watching it crash once, I’m not interested in cutting it close again.
The credits run out in May
The $120 in AWS credits expire January 2027, but at ~$30 a month they’ll be gone by mid-May. After that, real bills start. So the ongoing math is roughly $200 for Claude Max plus $30 for EC2, which is $230 a month going forward.
I’ve been thinking about whether a Mac Mini M4 Pro changes that equation. The $1,999 model has 24GB of unified memory and could run a 70B parameter model locally, no API costs. The break-even against $230/month is about nine months.
But here’s what made me pause: if I tried to match that Mac Mini’s GPU memory with an EC2 instance, I’d be looking at a g5.xlarge — an A10G GPU, 24GB VRAM — at $1.01 an hour. That’s $725 a month. Against that comparison, the Mac Mini pays itself off in under three months.
I’m still waiting. HP Desktop is already stretched thin, and Apple may announce something new before the Mac Mini I’d buy even ships. But the math is less obvious than it looked at first.
What I actually got for $311
Since January 28th: over 100 GitHub issues closed, a multi-instance coordination system, an automated testing agent, a daily think tank that generates improvement ideas, Home Assistant integration, an Omi wearable pipeline, a knowledge base, custom OCR tools, vacation mode automation that kept everything running while I was on a cruise for a week.
I don’t have a clean way to put a dollar value on that. Some of it is infrastructure I’ll use for years. Some of it is tools that save me time I wasn’t tracking. Some of it is just… interesting.
What I do know is that $5.90 a day has bought me a system I actually think about. Not a chatbot I forget is open in a tab. Something that’s running right now, working on something, whether I’m paying attention or not.
That’s a different kind of value. For less than a Grande White Mocha, it’s worth it.