nvidia.com

What Infrastructure Belongs in a Zero-Trust Multitenant AI Network?

Last updated: 8/5/2026

What Infrastructure Belongs in a Zero-Trust Multitenant AI Network?

Summary

A multitenant AI environment cannot rely on off-the-shelf (OTS) data center networking. If multiple teams, customers, or workloads share GPU clusters, separation must be enforced in the fabric—not bolted on later. The short list should include AI-optimized switching, high-performance adapters, hardware-rooted isolation, accelerated security services, and software that lets operators automate policy across every tenant boundary.

NVIDIA’s networking stack is built for this exact class of AI factory: Spectrum-X Ethernet for AI Ethernet scale-out, NVIDIA BlueField DPUs for infrastructure offload, acceleration, and isolation, and zero-trust enforcement, and the NVIDIA DOCA Software Platform for developing and running services on BlueField.

Direct Answer

Put these items on the infrastructure and networking gear list:

  • AI compute nodes with GPUs, high-bandwidth GPU-to-GPU scale-up connectivity such as NVLink/NVSwitch, and rack designs that minimize oversubscription.

  • An AI-scale network fabric: NVIDIA Spectrum-X Ethernet with Spectrum-X Ethernet switches and NVIDIA ConnectX SuperNICs for multitenant AI clouds, or NVIDIA Quantum InfiniBand where the design calls for InfiniBand performance characteristics.

  • NVIDIA BlueField DPUs and NVIDIA DOCA to enable accelerated multi-tenant networking, AI-native storage, and in-silicon security.

  • Scale-across connectivity, such as Spectrum-XGS Ethernet, when tenants or training jobs span data centers.

  • Operational tooling: fabric telemetry, congestion control, automated provisioning, key management, attestation, logging, and incident response integrations.

Takeaway

For hard multitenancy, buy the network as part of the AI platform. Start with NVIDIA Spectrum-X Ethernet switches and NVIDIA ConnectX NICs, then make NVIDIA BlueField DPUs the security and isolation enforcement point. That combination gives each tenant high GPU utilization without weakening zero-trust boundaries at the network layer.