From DPU to Fabric: Multi-Tenant Network Automation with NVIDIA NICo and Aviz ONES

Overview

Building a multi-tenant AI infrastructure requires coordination across two distinct networking domains: the GPU host and DPU layer, where tenant networking originates, and the data-center fabric, where that tenant must reach storage, shared services, and external networks.

NVIDIA Infra Controller (NICo) manages GPU hosts and their BlueField-3 DPUs, establishing the network identity associated with each tenant. Aviz ONES extends that tenant networking model into the broader AI fabric.

The architectural principle is straightforward: NICo remains the source of truth for tenant network identity, while ONES consumes that identity and programs the surrounding network infrastructure consistently. This avoids creating two independent sources of network configuration and provides a common workflow from GPU host provisioning through external connectivity.

Dividing Responsibility Between NICo and the Fabric

NICo operates closest to the workload. When a GPU server is allocated to a tenant, NICo configures the associated BlueField-3 DPU with the networking information required to identify and isolate that tenant. This includes parameters such as:

• VLAN
• VNI
• VTEP information

Because these values originate during workload and host provisioning, NICo naturally becomes the authoritative source for the tenant's network identity.

The important architectural decision is that the external fabric should not independently allocate another set of tenant identifiers. Instead, the network infrastructure should consume the identifiers already assigned by NICo.

This creates a clean ownership model: NICo owns the tenant identity at the host/DPU layer. ONES extends that identity across the network fabric.

Where Aviz ONES Fits

Aviz ONES manages the network infrastructure surrounding the GPU hosts, including the north-south portion of the AI fabric.

In the reference architecture, ONES manages:

• Front-end fabric (leaf/Spine)
• Border leaf switches
• Aviz Service Gateway
• Firewall connectivity
• Storage leaf

ONES connects to NICo as an infrastructure provider and retrieves the networking attributes associated with each tenant. It then uses those values to configure the rest of the fabric.

The objective is not for ONES to create a parallel tenant model. Instead, ONES translates the tenant definition established by NICo into the configuration required by the physical and virtual network infrastructure.

Extending Tenant Connectivity Beyond the GPU Host

The BlueField-3 DPU represents the first enforcement point for tenant networking.

From there, tenant traffic traverses the front-end leaf-spine fabric toward services such as storage, external networks, or other shared infrastructure.

The Aviz Gateway provides the point where tenant overlay connectivity can terminate before reaching external networks. It can also provide functions such as:

• Tenant-aware routing
• NAT & Firewall
• Load Balancer
• External network connectivity

Firewall policy can then determine which external destinations or services a tenant is permitted to access. ONES programs these infrastructure components using the tenant networking information obtained from NICo rather than independently assigning network attributes.

Reference Architecture

The reference topology illustrates the separation of responsibilities between the two control domains.

At the bottom of the architecture are GPU servers equipped with BlueField-3 DPUs. These systems are managed by NICo.The DPUs connect into the front-end leaf-spine network. Tenant traffic then traverses the fabric toward border leaf switches and the Aviz Gateway, which are managed through ONES. Storage is connected through a separate storage leaf network. The important point in the topology is not simply which controller manages which device. It is the continuity of tenant network identity across those domains. A tenant created at the GPU infrastructure layer must retain the same networking context as its traffic moves through the fabric. The diagram in the original blog captures this boundary: NICo manages the DPU and compute domain, while ONES manages the surrounding fabric infrastructure.

Control-Plane Workflow

The integration can be understood as a sequence of control-plane operations.

1. Build the AI Fabric

The operator first creates the network fabric in ONES. This establishes the physical fabric topology and the infrastructure ONES will manage.

2. Register NICo as a Provider

ONES connects to NICo using the NICo controller address and provider credentials.This integration gives ONES visibility into the tenant networking information managed by NICo.

3. Create the Tenant

The operator creates a tenant and assigns GPU server resources. NICo provisions the appropriate networking configuration on the BlueField-3 DPU associated with those servers.

4. Retrieve Tenant Network Identity

NICo returns the network identifiers assigned to the tenant.For example, this may include the VLAN, VNI, and VTEP information associated with the tenant.

5. Extend the Tenant Across the Fabric

ONES consumes these values and configures the corresponding network infrastructure, including:

• Border leafs
• Aviz Service Gateway
• Firewall
• Storage-facing network

The same tenant identifiers therefore remain consistent from the GPU host through the network fabric.

Why the Source-of-Truth Model Matters

Without this integration, two orchestration systems could independently maintain tenant networking information. For example:

NICo has Tenant A → VLAN 100 → VNI 10100 while the network controller separately maintains,

Fabric Tenant A → VLAN 200 → VNI 20200

Even if both configurations are individually valid, the overall system is no longer deterministic.

The better architecture is, Tenant Creation
→ NICo assigns network identity
→ ONES discovers the identity
→ ONES extends it across the fabric

This establishes a single authoritative source for tenant network parameters while still allowing separate systems to manage different infrastructure domains.

Architectural Takeaway

The interesting part of this integration is not simply connecting two controllers through an API. It is establishing a clear control-plane ownership model.NICo manages GPU infrastructure and establishes tenant networking at the BlueField-3 DPU. Aviz ONES consumes that tenant definition and extends it across the AI network fabric.

That separation allows each controller to operate within its natural domain while maintaining a consistent tenant network model end to end. For operators building multi-tenant GPU infrastructure, this provides a practical way to combine GPU infrastructure orchestration with fabric automation without introducing a second, independent source of tenant network identity.