Nutanix to show AMD-powered enterprise AI infrastructure

Nutanix is combining its cloud and enterprise AI software with AMD processors and accelerators to help organisations run agentic AI workloads on private and hybrid infrastructure.

The approach is designed to give enterprises greater control over model selection, data governance and computing costs as autonomous AI agents move into production environments.

Nutanix said organisations can use a two-tier model that reserves frontier AI systems for tasks requiring complex reasoning. Higher-volume workloads can instead run on optimised models hosted on private infrastructure.

Routing control

Autonomous agents can retrieve information, interact with enterprise applications and execute workflows over extended periods. These activities can generate large volumes of model requests as agents perform routing, validation and execution tasks.

The resulting token usage can increase operating costs when organisations rely primarily on rented infrastructure or external model services.

Nutanix proposes using an agent gateway to direct requests between different models and infrastructure environments. The gateway can apply organisational policies, manage access controls and determine which workloads require external frontier models.

Routine requests can be routed to models running within an organisation's own infrastructure. More demanding tasks can be sent to larger external models when additional reasoning capability is required.

This structure is intended to reduce the proportion of workloads processed through higher-cost models. It also gives organisations a central point for managing model access and usage policies.

"Owning your intelligence doesn't mean completely abandoning frontier models; it means taking control of your routing, your volume, and your costs. This allows you to leverage expensive frontier models for a small fraction of tasks that require complex, edge-case reasoning, while the vast majority of your high-volume tasks are routed to highly optimised models running securely on your private infrastructure," said Debo Dutta, Chief AI Officer, Nutanix.

Private workloads

The architecture combines the Nutanix Cloud Platform and Nutanix Enterprise AI with AMD EPYC processors and AMD Instinct MI355X accelerators.

Nutanix Enterprise AI is used to deploy and operate AI models and applications within governed infrastructure environments. The broader cloud platform provides the underlying compute, storage and management layer for workloads running across private and hybrid environments.

Keeping selected AI workloads within private infrastructure can help organisations retain control over proprietary information. It can also support data sovereignty requirements by limiting where sensitive information is processed and stored.

This may be relevant for organisations operating across jurisdictions with specific data residency, security or governance requirements.

The platform is intended to support long-running agents that repeatedly access models and enterprise data. Policies can be applied to determine which models agents can use, what information they can access and where their workloads are processed.

Shared inference

AMD EPYC processors and AMD Instinct MI355X accelerators provide the computing resources for model inference and related enterprise workloads.

Nutanix said the combined infrastructure can support shared inference environments rather than requiring separate systems for each AI application or business unit.

Shared capacity can allow organisations to allocate computing resources across multiple workloads. It can also improve infrastructure utilisation when demand varies between teams and applications.

The design targets enterprises that want to expand AI deployments without assigning dedicated hardware to every model or agent. Centralised infrastructure can serve several applications while governance and access policies remain managed through the platform.

Organisations can also decide which workloads remain within private environments and which are sent to external services. This provides a way to balance model capability with operating cost and internal control.

Production scale

Enterprise AI programmes are increasingly moving beyond isolated trials towards systems that operate continuously within business processes.

Agentic systems create different infrastructure demands from standalone chatbots or occasional model queries. Agents may remain active for extended periods, consult several sources and make repeated requests before completing a task.

These operating patterns can raise computing and token costs as deployment volumes increase. They can also make model governance more complex when agents use several internal and external services.

Nutanix's two-tier structure separates high-volume processing from tasks that require larger frontier models. Organisations retain the option to use external systems while handling more predictable workloads through privately operated models.

The platform also allows AI workloads and proprietary data to remain within governed hybrid multicloud environments. Access, routing and policy controls are managed through the agent gateway and the Nutanix software stack.

Nutanix said the shared inference infrastructure is designed to maximise the use of AMD EPYC CPUs and AMD Instinct MI355X GPUs while supporting long-lived agents and controlling token-related costs.