Blog Post

Azure Compute Blog
4 MIN READ

Direct Virtualization: A New Foundation for Agentic AI in Azure

DarrenM's avatar
DarrenM
Tin Contributor
Nov 18, 2026

Agentic AI and other modern workloads increasingly need to run in strongly isolated sandboxes, with high performance, high scalability, fast lifecycle operations, and efficient access to GPUs.  Traditionally, many of these workloads relied on nested virtualization, trading-off performance and efficient access to GPUs to gain strong workload isolation.  Azure’s new technology, direct virtualization, eliminates that tradeoff.  Instead of nesting a second hypervisor, direct virtualization provides a simpler solution.  Direct virtualization enables services to create strongly isolated sandboxes inside Azure VMs, choose their own lightweight OS for those sandboxes, and run agentic AI and other modern workloads with efficient access to GPUs.

The direct virtualization sandboxing model

By removing the need to nest a second hypervisor, direct virtualization provides a direct path between strongly isolated sandboxes and Azure’s underlying virtualization platform (see the diagram below).  This architectural enhancement reduces virtualization overhead, simplifies the stack, and empowers services to deploy agentic AI and other modern workloads that need strong isolation, high performance, high scalability, fast lifecycle operations, and direct access to GPUs.

 

 

Because direct virtualization is an integral part of Azure’s virtualization platform, a variety of workload hosting platforms can leverage this innovation.  That includes Kubernetes platforms running lightweight isolated pods, enterprise virtualization platforms running custom VMs, data services hosting large numbers of isolated guest environments, GPU-backed services that need direct access to GPUs, and agentic AI platforms that need secure execution sandboxes at high density.

Performance becomes increasingly important at scale

The benefits become increasingly visible as workload density increases. In baseline CPU and memory testing, direct virtualization and the next-best alternative, nested KVM, delivered performance within a few percent of each other.  However, as workload density increased, direct virtualization sustained performance more effectively.

We used Redis as a representative latency-sensitive, high-throughput workload to evaluate how performance holds up as workload density increases. We increased workload density across the same underlying host configuration to compare the throughput of both virtualization techniques. Direct virtualization and nested KVM delivered comparable throughput at baseline density.  At 4x workload density, direct virtualization delivered approximately 32% higher throughput (see the graph below).

 

 

This matters because the goal for many services is not simply to maximize the performance of a single isolated workload.  It is to maintain useful workload performance while consolidating more isolated environments onto shared infrastructure.  The ability to sustain performance as density increases directly affects service scale and economics.

Bringing strong isolation to cloud-native platforms

Kubernetes pod sandboxing, via KubeVirt or Kata, is an important example of the need to maintain strong isolation while sustaining performance and increasing workload density.  When the Kubernetes node itself runs inside an Azure VM, Kubernetes has traditionally depended on nested virtualization for pod sandboxing.

Direct virtualization enables Kubernetes services, such as AKS Pod Sandboxing, to leverage strong virtualization-based pod sandboxing while reducing the overhead and complexity of nesting a second hypervisor.  This creates opportunities to improve workload performance and support higher density as these platforms scale.

The goal is straightforward: preserve the isolation benefits of virtual machines while moving performance and scale closer to native virtualization.

Modernizing enterprise virtualization with Red Hat

Direct virtualization also creates new opportunities for enterprise virtualization platforms.  Microsoft and Red Hat are collaborating to enable Red Hat OpenShift Virtualization to take advantage of direct virtualization on Azure Red Hat OpenShift, a fully managed application platform.  Red Hat OpenShift Virtualization allows organizations to run and manage virtual machines and containers running in pod sandboxes alongside regular containers using Kubernetes APIs and workflows.  For customers modernizing existing VM-based applications, this provides a path to bring both virtualized and containerized workloads together on a common Kubernetes platform.

We used Sysbench to provide a simple, repeatable measure of CPU performance as workload density increased, revealing the scalability of direct virtualization.  The test used identically sized guest VMs while increasing the number of VMs per host, allowing us to compare CPU performance as workload density increased.  Direct virtualization and nested KVM delivered comparable performance at baseline CPU allocation.  As workload density increased, direct virtualization sustained performance more effectively than nested KVM.  At 4x workload density, direct virtualization delivered approximately 15% higher Sysbench throughput (see the graph below).

 

 

This collaboration also highlights an important design goal for direct virtualization: empowering the existing virtualization ecosystem and service platforms to adopt the innovation without requiring them to abandon the tools and operating models they already use.

Expanding the isolated compute landscape

The same architectural model applies well beyond Kubernetes and enterprise virtualization.  Data and analytics services often need to run large numbers of isolated guest environments inside a service-controlled VM.  In these environments, per-VM overhead, networking performance, startup time, and virtual processor scale can directly affect both customer experience and service economics.  Direct virtualization preserves strong workload isolation while reducing the architectural overhead of a nested hypervisor.

GPU-backed services, gaming, media, AI inference, and other accelerator-heavy workloads introduce another challenge: device access.  These services may need to pass GPUs through to workloads while providing strong workload isolation.  Direct virtualization is designed to make these scenarios easier by creating a direct relationship between isolated workloads and GPUs.

Agentic AI workloads increasingly need their own isolated execution environments where they can run code, use browsers, install tools, access data, and interact with external systems.  Those environments need to be secure, lightweight, fast to create, scalable to high densities, and capable of accessing high-performance devices.  Direct virtualization provides a foundation for building these kinds of isolated execution environments without the performance overhead of a traditional nested virtualization stack.

Direct virtualization: a new foundation for agentic AI and modern workloads

With direct virtualization, agentic AI and other modern workloads are empowered to run in strongly isolated sandboxes, with high performance, high scalability, and fast lifecycle operations.  By leveraging direct virtualization, hosting services can create strongly isolated sandboxes inside Azure VMs, use their own lightweight OS, and give their workloads direct access to GPUs.

If you would like your organization to be considered private preview access, then please submit this form: aka.ms/directvirtualizationpreview.

Updated Nov 16, 2026
Version 1.0