Nutanix is releasing version 2.8 of Nutanix Enterprise AI. The release introduces MCP server management in Agent Gateway and expands Private Inference with fine-tuning, air-gapped Nvidia NIM, and multi-node inferencing.
AI agents rarely operate in isolation. A single autonomous agent can launch multiple other agents, each of which in turn connects to enterprise applications, internal data sources, and external tools. According to Nutanix, this is precisely where oversight gaps arise. Who is allowed to access what, and how do you maintain visibility at scale?
With Nutanix Enterprise AI (NAI) 2.8, the company aims to answer that question. The release is now available to existing NAI customers and affects the Nutanix Agent Gateway and Nutanix Private Inference components.
MCP server management out of preview
The most significant addition is the general availability of MCP server management in Agent Gateway. The Model Context Protocol has now evolved from a convenient integration layer to critical enterprise infrastructure, with over-privileged service accounts, shadow servers, and missing audit trails posing the primary risks.
Agent Gateway acts as a single secure gateway between agents and the tools they invoke. Users can run MCP servers locally within their NAI environment, ensuring that sensitive data remains within their own infrastructure. A single endpoint bundles multiple MCP servers, eliminating the need for separate connectors per tool. Permissions can be finely tuned, for example, read-only versus write, on a per-user or per-API-key basis. Local MCP servers also support rolling updates.
On top of that, header-based rate limiting with token budgets is currently available as a tech preview. This allows administrators to limit token consumption per user without disrupting existing user management.
Smaller models, greater GPU efficiency
On the inference side, efficiency is key. Fine-tuning makes it possible to train models with fewer than 8 billion parameters on proprietary domain data, allowing them to approach the accuracy of large LLMs for specific tasks. This reduces GPU costs.
Additionally, Nvidia supports NIM microservices in air-gapped environments, intended for highly regulated sectors with an Nvidia AI Enterprise license. Two tech previews round out the package: multi-node and multi-GPU inference for models with 100 billion parameters or more, plus kvCache offloading, which moves cache context from GPU to CPU memory once capacity is reached.
At the platform level, Nutanix is introducing a Custom Role Builder. Administrators can use it to create custom roles from a catalog of more than thirty permissions, covering user management, licenses, models, endpoints, API keys, and observability.
Tip: Nutanix doesn’t give AI free rein: infrastructure remains a human endeavor