Emerging Technology Preview
We are delighted to share innovation in the world of AI. At Build, we demonstrated several innovations that showcase where Azure Container Instances (ACI) is heading and how it can become a foundational platform for the next generation of AI workloads.
As AI agents become capable of writing code, invoking tools, performing reasoning, and executing complex workflows, the industry needs a new execution model—one that combines the elasticity of serverless computing, the security of virtualization, and the speed required for agentic workloads.
This is the vision behind ACI AI Sandboxes.
Containers, Reimagined for AI
Containers have become the standard unit of deployment in modern cloud computing. However, traditional process-isolated containers were never designed to be the security boundary for running autonomous AI workloads.
ACI took a different approach by leveraging Hyper-V isolation to provide VM-grade security while preserving the agility of containers.
The next evolution is taking place through Direct Virtualization (L1VH), where isolated container workloads can run much closer to the hardware while maintaining strong isolation guarantees. Direct virtualization drastically reduces overhead while improving performance and efficiency.
AI Sandboxes
AI Sandboxes build on these foundations to provide isolated execution environments for AI agents.
An AI Sandbox can be viewed as a secure runtime where an agent can:
- Execute generated code
- Run tools and plugins
- Access enterprise resources
- Perform reasoning tasks
- Execute independently from other agents
Each sandbox operates within its own isolated environment, enabling secure multi-tenant execution at cloud scale.
Converging Compute Platforms
We are bringing the power of Compute platform together to unlock the potential of agents.
Rather than building separate infrastructure stacks for virtual machines, containers, AI agents, and future workload models, Azure is moving toward a common compute platform and Control Plane capable of hosting multiple execution environments.
The architecture brings together:
- Azure Container Instances
- AI Sandbox Groups
- Virtual Machines
- Direct Virtualization (L1VH)
- L2 Utility VMs
- Azure Host Infrastructure
- A common Azure Compute Control Plane
This enables innovation in one layer of the stack to benefit all workload types.
Demonstrating Scale
Perhaps the most exciting demo showcased the scale characteristics of AI Sandboxes (Cloud Native Innovations with Mark Russinovich - Advances in Serverless Containers and ACI Enhancements)
Using memory snapshotting techniques, ACI can initialize sandbox environments without repeatedly performing full startup sequences. Instead, pre-initialized memory images are reused, dramatically reducing startup latency.
The result is the ability to launch 10,000 AI sandboxes in approximately two seconds, demonstrating both rapid startup and hyperscale elasticity.
For agentic workloads, where thousands of short-lived execution environments may be required simultaneously, this capability becomes foundational.
AI Infrastructure Beyond Containers
The demonstrations also highlighted how AI workloads increasingly consist of multiple cooperating components:
- CPU-based orchestration services
- GPU-accelerated inference engines
- Embedding models
- Tool execution environments
- Agent runtimes
Direct Virtualization enables these workloads to be securely isolated while still receiving direct access to hardware resources, including GPUs.
This creates a powerful model where AI agents, inference services, and supporting workloads can coexist on a shared platform without compromising security or efficiency.
Looking Ahead
ACI AI Sandboxes are still an emerging technology, but they provide a glimpse into where cloud infrastructure is heading.
The future of AI will require millions of isolated execution environments capable of running securely, starting instantly, and scaling elastically. By combining Hyper-V isolation, Direct Virtualization, serverless operations, and Azure's compute platform, ACI AI Sandboxes are laying the groundwork for that future.
As agentic AI continues to evolve, secure serverless sandboxes may become as fundamental to AI applications as containers became to cloud-native applications.
If you like to engage or get early access, please fill out this form: https://aka.ms/ACI-Sandbox-engagement
Microsoft