Nutanix Boosts Enterprise AI with New Inference, K8s, Governance

Enterprises running agentic AI in production consistently hit the same wall: inference, orchestration and governance span separate toolchains, and the seams between them accumulate cost and operational drag. Nutanix‘s NAI 2.8 and the forthcoming NKP 2.19 are a direct response, pulling inference control, Kubernetes management and multitenancy governance into one platform across hybrid environments.

Inference at Scale

NAI 2.8 adds smart model routing and one-click secure inference endpoints to the Nutanix Private Inference stack. For larger workloads, NAI 2.8 adds scalable multi-GPU serving via tensor parallelism for high-throughput, low-latency LLM inference, along with speculative decoding that Nutanix says can accelerate token generation by up to 2.5x using lightweight draft models. LoRA fine-tuning support lets organizations cost-effectively customize open LLMs under 8 billion parameters on private data using single-GPU compute.

These inference improvements matter most to teams that need consistent throughput without building a separate MLOps layer on top of their existing virtualisation estate. For context on how open-weight models are shaping enterprise inference demand, see our coverage of open-weight inference trends.

MCP Gateway and Governance

The MCP gateway, now generally available within the Nutanix Agent Gateway, is the governance centrepiece of this release. It gives IT a single control point for how AI agents interact with applications and data via the Model Context Protocol: access policies, token consumption monitoring, usage caps and cost-based model routing all flow through one interface rather than across fragmented per-agent configurations.

Thomas Cornely, Nutanix’s executive vice president of product management, described the goal as giving IT a unified management interface for developer and AI builder policies, according to the company. Nutanix Flow Network Security complements this with microsegmentation across workloads and locations, isolating AI components from the broader application estate. Teams managing sensitive data under strict sovereignty or compliance requirements will find the combination of Flow and the MCP gateway covers both the network and the agent interaction layer.

Kubernetes and Bare-Metal Options

NKP 2.19 ships with a built-in AI catalog: notebooks, vector databases and MLOps engines available as curated open-source services. The platform is CNCF-compliant and targets both virtualised and bare-metal environments. NKP Metal handles OS, firmware and container deployment on bare metal automatically, adding enterprise-grade persistent storage for workloads that need direct hardware access without giving up operational consistency.

The dual-native architecture runs VMs and containers side by side on the same platform. That matters operationally: teams can bring AI workloads to where data already lives rather than migrating applications to a separate AI-optimised cluster.

Multitenancy via SP Central

Service Provider Central is now generally available, providing a multitenant control plane covering infrastructure, applications, cloud-native services and AI. On shared GPU infrastructure, SP Central isolates tenants at the resource level, which lets service providers extend their offering beyond raw GPU capacity into Kubernetes-as-a-service and model-as-a-service without rebuilding isolation controls for each tier.

For large enterprises, the same capability applies internally: separate business units running distinct AI initiatives on shared infrastructure each get isolated resource pools and dedicated management, without requiring separate physical deployments. Nutanix Flow handles self-service consumption and workload mobility across the multitenant environment, and the company says the architecture avoids vendor lock-in through its open-source and CNCF-compliant foundation, though how that plays out in practice depends on how deeply teams integrate with Nutanix-specific tooling.

Where This Fits

Existing Nutanix Cloud Platform customers face the lowest friction path here. The dual-native architecture means no forced migration from VM-based workloads, and the MCP gateway plugs into existing governance processes rather than replacing them. The platform is a natural fit where the requirement is to run production AI agents alongside traditional applications under unified policy, particularly in regulated industries where data sovereignty constrains public cloud options.

For service providers, SP Central’s generally available status is the most commercially consequential part of this release. The ability to offer model-as-a-service and Kubernetes-as-a-service on top of GPU infrastructure, with tenant isolation built into the control plane, removes a significant integration burden from providers building AI-as-a-service offerings. Whether NKP 2.19’s AI catalog is broad enough to meet enterprise build requirements will depend on which open-source services make the cut at general availability.