Multi-Tenancy In AI Clouds

This paper presents an architectural model for hard multi-tenancy — true multi-tenancy built on hard isolation (enforced on networking hardware) — in AI factories built on NVIDIA® DGX™ systems. Tenant intent is defined once and applied consistently across heterogeneous connectivity domains: East–West fabrics based on NVIDIA Spectrum-X™ or NVIDIA Quantum InfiniBand™, Ethernet-based North–South fabrics, NVL72 rack-scale NVLink® Multi-Node fabrics, and NVIDIA BlueField® DPU-augmented domains.

The paper examines aligning Ethernet, InfiniBand, and NVLink partition boundaries; coordinating lifecycle updates across fabrics; and maintaining consistency through topology-aware validation — and shows how the Netris® NAAM (Network Automation, Abstraction, and Multi-Tenancy) platform implements this model using supported NVIDIA management APIs.

This guidance is intended for CTOs, AI/cloud architects, NVIDIA Cloud Partner (NCP) operators, and enterprise infrastructure teams deploying shared DGX-based clusters, including NVIDIA Mission Control-managed environments.