At Nutanix .NEXT in Amsterdam, we sat down with Mano Bhattacharyya, CTO of Nutanix, about the company’s shift to dual-native infrastructure, the consolidation of IT and platform engineering, GPU scarcity strategies, and the upcoming AI Data Hub. A pipeline that promises to bring enterprise data under unified governance for agentic AI.
Nutanix’s headline announcement at .NEXT is what Mano calls dual-native: the platform now treats virtual machines and containers with the same policy framework. Networking, storage, disaster recovery, replication, snapshots, VPCs, and micro-segmentation can all be configured once and applied uniformly to both VMs and Kubernetes workloads. This is something that was very hard to do before, but Nutanix has now combined both technologies in it’s platform an unified the control layers.
Accordding to Mano it all starts with the approach. Mano states it’s about managing an protecting applications. “Whether the application is written in a monolithic virtual machine or in a containerized form factor, we need to provide support for all these guarantees on both types of applications.”
Nutanix has done the first major step in doing this. The second step is coming soon when micro-segmentation for containers will become available. Accordding to Mano there is no architectural barrier to handle VM’s and containers the same, it’s simply a matter of roadmap execution.
Also read: As IT complexity escalates, Nutanix fights back
The collapse of DevOps in the enterprise
One of the more striking arguments Mano makes is that DevOps, the model where developers own and handle IT operations, has largely failed in enterprise environments. The result of giving developers full operational freedom was an explosion of configuration diversity: different database versions, incompatible tooling, and security postures that IT could not govern.
Nutanix’s response is a three-persona model: application developers, platform engineers, and infrastructure providers. In large enterprises, Mano says, the platform engineering and infrastructure layers are now merging back under IT control. IT defines what services are available; this can be a specific version of Postgres, a prescribed load balancer, a specific web or cache server. Developers simply consume those services without bringing their own.
“We’re going back to one unified configuration basically, where everything runs,” Mano says. This will make managing the IT infrastructure a lot less complex. The analogy he uses is the public cloud: developers can request RDS or Kinesis, but they do not get to provision arbitrary infrastructure. The same model can also be applied on-premises with Nutanix.
Project-based isolation and multi-tenancy
To make it easier to manage the infrastructure, Nutanix is moving customers towards a project-based resource structure. Each project functions like a VPC (virtual prive cloud): it has its own compute, storage, and networking boundary, and users are granted access through role-based access control (RBAC). A user can belong to multiple projects with different permission levels, read-only on a cash register application, write access on an inventory system, for example.
Existing (brownfield) customers are automatically placed into a default project when they migrate. Mano recommends that organizations progressively move workloads into named projects to gain quota control, SLA assignment, and visible isolation. Applications that need to communicate across projects can use VPC tunneling, with both NAT and no-NAT connectivity.
GPU scarcity and the NeoCloud strategy
For enterprises that want to run AI inference workloads, GPU access remains a significant constraint. Buying GPUs outright is expensive; deploying them in existing data centers is even harder because modern GPU racks demand 100 to 150 kilowatts. Compared to the 20 to 30 kilowatts typical of enterprise racks today. Mano cites that Dell has demonstrated running at 400 kilowatts per rack as an illustration of where the hardware is heading.
Nutanix’s answer is to connect its NKP (Nutanix Kubernetes Platform) and NAI (Nutanix AI) software to inference endpoints running wherever GPUs are available, this can be in the public cloud or at NeoClouds. NeoClouds are what Mano calls the new generation of GPU-as-a-service providers. He notes that while tier-1 NeoClouds have their own software stacks, the majority of the 100-plus NVIDIA-certified NeoCloud partners do not. Nutanix is targeting that tier-2 and tier-3 segment, offering its full stack; hypervisor, Kubernetes, and NAI inference endpoint to cloud operators who are currently just renting bare hardware.
For enterprises that do have GPU hardware on-premises, full passthrough virtualization and hardware partitioning via NVIDIA’s Multi-Instance GPU (MIG) technology are already supported. Software-based GPU virtualization is not used for AI that much, according to Mano. The performance requirements push customers toward MIG and bare-metal passthrough instead.
AI software stack: NKP catalog and managed tooling
Nutanix is building a software catalog for the Nutanix Kubernetes Platform (NKP) with many AI-oriented managed services. The planned catalog includes Jupyter Notebook as a service for data engineers, model versioning and MLOps tools including Kubeflow and MLflow, and vector databases. Nutanix will take care of version management en dependency issues, so Nutanix customers don’t have to worry about compatibality between all the software packages.
Mano draws an explicit parallel to what public cloud providers do with open source software: take the components, ensure they work together, and abstract the complexity away from the end user. The NAS storage layer has already been certified by NVIDIA for enterprise AI performance, including GPU Direct support for maximum throughput.
AI Data Hub: the data side of the AI equation
Nutanix is also working on a AI Data Hub that should become available soon. The product addresses a gap that most compute-focused infrastructure vendors have not tackled: curating, joining, and governing the unstructured data that AI systems need to be useful.
AI Data Hub functions act as a pipeline with multiple ingestion points: Google Workspace, Microsoft SharePoint, local NFS file services, and semi-structured sources like Salesforce and Workday. These can all be injected into the AI Data Hub. Once the data is ingested, customers can transform the data as well. They can use AI-driven prompts rather than code to join or reduce data sets, but also apply security controls before the data is exposed to the AI systems.
The output is exposed via MCP (Model Context Protocol), making it compatible with agentic AI systems. What is very important is that the role based access control travels with the data. If a user queries through an AI agent, they see only the documents and records they are authorized to access across all ingestion sources, normalized across systems.
“It is doing basically three things,” Mano summarizes. “One is preparation of data, second is joining of data, and third is RBAC of data.”
Nutanix is its own first customer for AI Data Hub, currently running it across GitHub, Jira, and Atlassian integrations in its engineering organization.
The broader strategic direction
Mano closes by articulating Nutanix’s differentiated position: most infrastructure vendors focus on either compute or data, not both. Nutanix believes it has the DNA, built through years of hyperconverged infrastructure, to attack the AI challenge from both directions simultaneously. The AI Data Hub is the most visible expression of that ambition, and Mano suggests that future Nutanix .NEXT events will be heavily oriented around data and AI themes.