← Armando Rodriguez All posts
Writing

Building a GPU AI Inference Lab

June 25, 2026
Edge AINVIDIAGPUHomelab

I'm building a self-hosted AI inference lab on a bare-metal Ubuntu workstation (RTX 4090, i9, 32 GB RAM) to get hands-on with the infrastructure that runs modern AI in production: the layer beneath the models, where the real operational work lives.

Rather than renting a managed endpoint, I'm assembling the full stack from the driver up, so I understand every layer the way I understand cloud infrastructure today.

The stack

Why I'm building it

I've spent about twelve years in cloud and infrastructure engineering: Azure, virtualization, networking, and secure, regulated environments. AI infrastructure is the same discipline (provisioning, observability, orchestration, security) applied to GPUs. This lab is where I turn that transfer from conceptual, as an NCA-AIIO certified engineer, into operational, hands-on skill.

What it demonstrates

Designing, provisioning, and operating GPU infrastructure end to end: from driver and CUDA validation, through containerization and telemetry, to model serving and Kubernetes orchestration, with the observability and reproducibility discipline of production cloud engineering.

I'll post updates here as each layer lands, with the concrete results (validated driver and CUDA versions, live metrics, and serving benchmarks) as they come in.