Scaling data centers for GPU-as-a-Service (GPUaaS) customers- connecting multiple data halls poses significant challenges- especially as the number of VxLAN endpoints grows within the network fabric. This presentation explores how EVPN VxLAN Multisite can address these challenges by dividing the network fabric into smaller segments and isolating them with split-horizon groups. We'll discuss the standard implementation and how SONiC- combined with Free Range Routing (FRR) and Switch Abstraction Interface (SAI)- achieves this. Additionally- we'll cover important considerations for network design to avoid common pitfalls and ensure a robust- scalable infrastructure for AI clusters.