Fair warning: the blog has been quieter than I'd like lately, and the reason is entirely self-inflicted. I decided to lean hard into the enthusiast side of this whole experiment. The home lab started life as a gaming PC, and I'm in the middle of converting it to a fully custom water-cooled loop, dedicated blocks for both the CPU and GPU. New case, new everything. I was a little overambitious about the timeline, so now I'm waiting on parts and the lab is effectively offline. There will be before-and-after thermal comparisons once it's all buttoned up, for those of you who are here for the hardware geekery and not just the AI discourse.
In the meantime, being lab-less has forced me to slow down and actually think. So you're going to see more posts like this one, call it the view from the cheap seats. What am I actually observing in the market right now? Here's what's on my mind.
Vibe Coding Just Crossed the Enterprise Threshold
Not long ago, vibe coding carried a slightly dismissive connotation, a toy, a hobby, something the citizen developer crowd played with on weekends. That framing is dissolving fast. What I'm seeing now, across large enterprises, is that tools like Claude Code, Codex, and Cursor are being rolled out not just to engineering teams but to all knowledge workers. That's a fundamentally different moment.
This is no longer about developers getting a faster autocomplete. This is companies making a deliberate bet that anyone who works with information can use AI to build something, a script, a report, an internal tool, without going through a formal dev cycle. That's a genuinely new thing.
And here's why I find it interesting beyond the obvious: the conversation has quietly shifted away from job displacement and toward value creation. The early AI discourse was dominated by anxiety, who's going to lose their job? What I'm seeing now is companies asking a different question: where are we leaving value on the table that AI could unlock? Vibe coding, of all things, might be the first real answer to that question at scale.
If you're working somewhere that's already building out a structured path for enterprise vibe coding, I'd genuinely love to hear how it's going. That's the whole spirit of this blog. This phenomenon is here to stay, and I want to understand it better.
Governance Is the Immediate Priority
Follow the logic chain and the next question becomes obvious: how do you make this safe? When you extend AI-assisted development to thousands of non-engineers, the governance surface area explodes. Who's reviewing what gets built? Where does the output go? What data is being fed into which models?
I think we're squarely in a governance mindset era right now. The enterprises that are moving fastest on vibe coding rollouts are simultaneously the ones most anxious about guardrails: acceptable use policies, model access controls, audit trails. That tension is real and it's not going away. I've written before about why regular audits matter and what happens when you actually run one on a live app. The same discipline applies at enterprise scale, just with a lot more stakeholders.
Token Economics Is the Wave Right Behind It
Here's where I think things get interesting from a technical and infrastructure standpoint: cost is about to become the dominant conversation. Right now, most organizations haven't fully felt the bill because adoption is still early and contained. Once vibe coding is genuinely enterprise-wide, the token economics become impossible to ignore.
This is a big part of why I've been so focused on local and hosted models in my own experiments. My working theory is that we end up with a tiered model architecture that looks something like this:
- A self-hosted or private cloud model handling the bulk of agentic, repetitive, and high-volume tasks, the stuff where cost per token really adds up
- A lower-tier frontier model doing most of the content generation and summarization work
- A high-end frontier model reserved for genuine value-creation tasks: strategic planning, complex reasoning, and probably the higher-stakes coding work
That last tier is where you're willing to pay frontier prices because the output actually justifies it. The middle tier is where you're optimizing. The self-hosted layer is where you're driving cost to nearly zero for the high-frequency, lower-complexity workload.
It's not a novel idea in the abstract, but I don't think most enterprises have operationalized it yet. They're still treating model selection as a one-time IT decision rather than a dynamic cost-optimization problem. That'll change.
The lab will be back online soon, and when it is we'll get back to the hands-on experiments. But I wanted to get these observations down while they're fresh: vibe coding going enterprise-wide, governance as the immediate challenge, and token economics as the tidal wave right behind it. That's the arc I'm watching play out in real time.
Are you seeing the same governance-first pattern in your organization, or has cost already jumped to the top of the priority list?
