The Real Cost of Building the Nexus

I’ve been at this since Saturday.

Not casually tinkering. Not following a tutorial with someone holding my hand. Thirty to forty hours of staring at a terminal, prompting back and forth, rewriting config files, and hitting walls I didn’t see coming.

This is the story nobody tells you about building your own AI infrastructure stack. And it’s exactly why most people quit before they ever get to the good part.

What I’m Actually Building

The Nexus is a self-hosted AI command center running on a dedicated server in Helsinki, Finland. A Hetzner AX42-U to be specific: 64GB of DDR5 RAM, a Ryzen 7 PRO processor, dual NVMe drives, and $64 a month. That’s it. No cloud subscription eating into your margins. No platform that can pull the rug on you. You own the machine and everything running on it.

The brain of the operation is Hermes, an open-source agent framework built by NousResearch. Hermes is what makes this thing actually intelligent. It doesn’t just respond to commands. It reasons, routes tasks, executes workflows, and operates with a persistent set of instructions I’ve built into a file called SOUL.md. Think of SOUL.md as the operating manual for the agent’s entire personality and decision-making process. Every behavior, every priority, every rule I want it to follow lives in that file. When Hermes wakes up, it reads SOUL.md first. That’s where my business logic lives.

Underneath that is n8n, a self-hosted automation engine that handles the complex workflow logic: Gmail, Google Drive, Calendar, Telegram, and more. And then there’s the NCA Toolkit, an open-source API built by Stephen Pope that replaces an entire stack of expensive subscription tools. Video transcription, captioning, media manipulation: all of it runs locally through Docker at no additional cost per use. What most creators are paying hundreds of dollars a month for in SaaS subscriptions, the NCA Toolkit handles for free on your own hardware.

The goal of the Nexus is to have a single autonomous system managing the operational layer of my entire business: AI consulting, content production, client communication, and more. Not replacing my judgment. Executing on it. Around the clock. Without me having to be the one doing it every time.

That’s what I’m building. Now here’s what it actually took to get there.

The OpenRouter Problem

Here’s the first thing that will slow you down and it’s not glamorous: rate limits.

I built the Nexus on OpenRouter as the LLM backend. Makes sense on paper. One API, access to multiple models, clean routing. What they don’t tell you upfront is that you’ll get rate-limited constantly until you’ve spent enough money to earn your way off the restricted tier. And when you’re in the middle of a build and trying to test whether your agent is actually working, getting rate-limited isn’t a minor inconvenience. It’s a full stop.

And here’s the thing: it’s not a mystery. It’s not a cryptic error message where you’re wondering if it’s Hermes, or your Docker config, or your n8n workflow. Hermes tells you straight up through Telegram that it’s rate-limited and can’t do anything. The tool is working. OpenRouter is the bottleneck. But knowing the problem and fixing the problem are two completely different things.

The real fix is moving off OpenRouter and plugging in individual API keys directly, pointing Hermes to specific models for specific tasks. Gemini for this. Anthropic for that. DeepSeek if it makes sense. But now you’ve added a layer of complexity to your build, because you have to build that routing logic into SOUL.md so it knows which model to call based on what it’s trying to do.

That’s a project inside the project. And if you weren’t expecting it, it’s the kind of thing that makes you close the laptop and walk away.

Why People Quit

This is the part I actually want you to sit with.

People quit because they come into this thinking AI is plug-and-play. They’ve seen the demos. They’ve read the threads. They believe you sit down for 15 minutes, connect a few things, and walk away with an autonomous system running your business.

That’s not how this works. It’s not even close to how this works.

Agentic AI isn’t truly agentic yet, not because the technology isn’t there, but because the setup requires a level of intentionality and patience that most people don’t have. The system is only as smart as the instructions you give it. The routing is only as clean as the logic you build in. The automation is only as reliable as the hours you put into testing, breaking, fixing, and testing again.

It’s still just a tool. An extraordinarily powerful tool, but a tool. And tools don’t build things on their own.

The ones who win are the ones who grind through the frustration, through the anger, through the mind-numbingly tedious back-and-forth of changing one line of code and seeing what breaks. They make it to the other side. That’s their moat. Not their intelligence. Not their tech background. Their willingness to stay in the room when it gets uncomfortable.

The Part That Should Excite You

Here’s what I want you to understand though: a year ago, this would have taken months.

The barrier to entry on building something like the Nexus has dropped through the floor. Because of large language models like Claude and ChatGPT, you can troubleshoot in real time, ask questions, get explanations, and course-correct without needing a computer science degree or a team of engineers. What used to take months of specialized knowledge now takes days of focused effort and a willingness to not give up when it gets hard.

Days. Not months. Not years. Days.

That’s the real story. It’s hard, it’s frustrating, and it’s absolutely worth it.

The Video

If you want to see what this actually looks like in practice, the 40-minute screen share of my initial install is behind the paywall. No commentary. No editing. Just the raw process so you can watch exactly what it takes and decide if you’re ready to build your own.

My name is Adam Peters, and I’m here to unfuck the transition.