HomeAboutThe BuildProject HistoryWork With Me

The April Build

I built a private AI stack on an NVIDIA GH200 in 26 hours.

Not for a client. Not for money. To understand what enterprise AI actually feels like from the inside.

How it happened

Sunday, April 13 · 1:00 PM

Laid off

Laid off from a 4.5-year role. Solar industry policy changes hit the company hard. It stung — but I'd been ready for a change. I'd grown the business significantly, built things that made real money, and never felt like any of it was truly valued.

Sunday, April 13 · 8:20 PM

NVIDIA says yes

Seven hours after getting laid off, I got an email from NVIDIA telling me I'd been accepted into their Connect program. I took that as a sign and kept moving.

Friday, April 17 · The build begins

Started building

Started building Friday night. 26 hours of building ahead. NVIDIA NIM. Docker. Cloudflare Zero Trust. AnythingLLM. Grafana. All of it from scratch on a GH200 superchip.

Monday Morning, April 20 · Done

Full stack deployed

Full stack deployed — 26 hours later. A working private AI gateway running Llama 3.3 70B. The kind of infrastructure companies pay enterprise money for. Built to understand it from the inside.

Friday, April 25 · Pivot

Stop selling. Start sharing.

After sending a blast to 5,500 Chamber contacts and getting one reply, the message was clear. Small businesses aren't ready yet. The real opportunity is inside companies where trust already exists and someone can actually move things forward.

What I actually built

The full stack

A full private AI gateway — the kind of system companies pay $50,000+ to stand up. Running on NVIDIA GH200 hardware. Open source tools. Built by one person in 26 hours.

NVIDIA NIM

Running Llama 3.3 70B — enterprise-grade inference on GH200 hardware

AnythingLLM

Frontend interface for the private AI gateway — clean, functional, deployable

Docker

Containerized the entire stack for portability and reproducibility

Cloudflare Zero Trust

Secure access layer — no VPN, no exposed ports, enterprise-grade security

Grafana + Prometheus

Real-time monitoring and observability for the full infrastructure

NVIDIA GH200

96GB HBM3e, 72 Arm cores, 900 GB/s bandwidth — the superchip underneath it all

What I learned

Four things that changed how I think

Infrastructure is the moat

The companies that own their own AI infrastructure aren't just saving money on API calls. They're building a competitive advantage that compounds over time. Data sovereignty matters.

The gap is bigger than I thought

Most businesses are using ChatGPT through a browser. The distance between that and a private inference stack is enormous — and most people don't know what they're missing.

Curiosity is the skill

I'm not a DevOps engineer. I figured this out by being willing to be confused, ask questions, read documentation, and try things until they worked. That's the actual skill.

You have to build to understand

No amount of reading about AI infrastructure prepares you for actually running it. The latency, the resource management, the failure modes — you only learn these by doing.

"I built it not to sell it but to understand it. It changed how I think about everything."
— Matt Gamble, April 2026
Let's talk about what this means for your team →