Building the Phase 1 network for a multi-gigawatt AI campus

The starting point

The campus is one of the larger AI and high-performance computing sites being built in North America. When I joined, there was no network: no fabric, no management network, no addressing plan and no connectivity back to the metro. Server and platform teams needed infrastructure they could build on quickly, and it had to scale to a campus many times the size of Phase 1 without being rebuilt.

What was built

An EVPN-VXLAN fabric defined entirely in Terraform

The data center fabric is an edge-routed EVPN-VXLAN design managed by Juniper Apstra. Rather than building blueprints by hand in the Apstra GUI, the whole fabric is defined in Terraform using the Apstra provider, which I had co-developed while at Juniper. Rack types, templates, blueprints, routing zones and virtual networks all live in Git. Adding capacity means a reviewed code change, not a weekend of hand configuration.

Automated DNS, DHCP and IPAM

Infoblox DDI is provisioned through Terraform too, so networks, address ranges and DNS records are created in the same workflow as the fabric itself. When a new network exists in the fabric, it exists in IPAM and DNS.

A Mist-managed out-of-band network

A dedicated out-of-band management network, built on Juniper EX switches managed from Mist, gives server, power and network teams access to every device even when production is down. It was the first thing built, so every later piece of infrastructure could be reached and managed from day one.

Diverse dark fiber to the metro

I worked with the optical vendor on diverse-path dark fiber from the campus back to a metro data center, including the vendor's software for optical network management and automation, so a single fiber cut cannot isolate the site.

A phased plan, starting in the lab

Every major design was validated in a lab before it touched production, and the rollout was planned in phases from lab validation through full redundancy. That kept Phase 1 on a predictable path while the facility around it was still under construction.

Why it matters for other projects

Most organizations aren't building a gigawatt campus, but the same approach works at any size: build the management network first, define the fabric and addressing as code from the start, validate in a lab, and give the server and platform teams infrastructure they can rely on. If you're planning a new data center, a fabric refresh, or automation for a network that was built by hand, I can help.