Cloud-Native Platform
The cloud-native platform supplies cluster and runtime capabilities beneath the AI platform. The core remains neutral among enterprise Kubernetes distributions that satisfy the required interfaces and support boundaries.
Explore cloud-native platform ecosystem solutions.
Ownership Boundary
Enterprise Kubernetes reconciles workloads and provides the default pod-scheduling path against declared constraints and allocatable resources. Resource Orchestration supplies higher-level queues, quotas, priorities, fairness, and placement intent. A selected orchestration implementation can extend or participate in filtering, scoring, and binding. Accelerator operators reconcile GPU, networking, and DPU software so Kubernetes can advertise supported resources and services. The Enterprise Hardware Reference Architectures remain authoritative for physical topology and supported scale points.
Required Capabilities
Isolation and Runtime-Security Boundary
Select an isolation class from the workload authority, data sensitivity, tenant boundary, and threat model—not from the packaging format alone. A namespace or ordinary container is not automatically equivalent to a sandboxed runtime, microVM, or dedicated VM.
Isolation and Runtime Services must provide:
- Admission controls that bind workloads to approved images, identities, policies, runtime classes, and device access.
- Default-deny network policy plus controlled ingress and policy-routed egress where workloads can reach external services or tools.
- Supported sandbox, microVM, VM, or equivalent containment boundaries for higher-risk workloads.
- Enforcement that preserves approved identity, autonomy limits, and resource boundaries when agents delegate execution. AgentOps tracks approvals and outcomes; Policy and Enforcement defines the enforceable limits.
- Runtime, network, identity, and policy telemetry correlated with platform and security operations.
- Named ownership for node and runtime patching, vulnerability response, quarantine, recovery, and decommissioning.
When the threat model requires separation from the host, enforcement, monitoring, or containment can run in a separately managed infrastructure domain such as an NVIDIA BlueField DPU. This pattern supplements the workload’s sandbox, VM, network, identity, and policy controls; it does not replace them.
Secure Agent Workspace applies these primitives to persistent agent execution and adds workspace lifecycle, delegated authority, credential, egress, and audit requirements.
NVIDIA Operators
Treat the NVIDIA AI Enterprise Infrastructure Support Matrix as a release gate before selecting or upgrading the platform. Supported combinations depend on GPU hardware, operating system, hypervisor or cloud, Kubernetes distribution, deployment model, and NVIDIA AI Enterprise release branch.
NVIDIA provides operators that can implement parts of the required lifecycle:
- GPU Operator for GPU software lifecycle and telemetry. Verify its current platform support.
- Network Operator for NVIDIA networking components and RDMA enablement. Verify the selected release against the infrastructure support matrix and operator documentation.
- DOCA Platform Framework Operator for supported BlueField DPU provisioning and infrastructure-service orchestration. Runtime isolation and containment policy remain separate architecture responsibilities.
- NIM Operator for NIM service and model-cache lifecycle. Verify its current platform support.
Check current platform support, product compatibility, and upgrade guidance before selecting versions. This design does not reproduce those time-sensitive matrices.
Boundary with the Hardware Foundation
The platform consumes nodes, accelerators, network interfaces, fabric, storage, and availability characteristics supplied by the selected Enterprise Hardware Reference Architecture. It does not redefine physical topology or facility requirements.