For the past few years, much of the enterprise AI conversation has been about experimentation.
Which model should we use? What proof of concept should we build? Where could generative AI save time?
Microsoft Ignite 2026 showed that the conversation is moving on.
AI is entering production. And production AI needs an entire platform underneath it.
That means infrastructure, compute, data, security, networking, monitoring, governance and cost management all start to matter.
Microsoft’s latest Azure announcements reflect that reality.
Infrastructure built for the agentic era
Azure Cobalt 200 virtual machines are now generally available and are planned for 24 Azure regions by the end of 2026.
Microsoft says the new Arm-based infrastructure can deliver up to 50% better CPU performance than Cobalt 100. AI agents are specifically relevant because much of an agent’s work can happen outside the model itself, with temporary workspaces and other CPU-intensive tasks affecting overall performance.
Why does that matter?
Because once organisations start running thousands or potentially millions of AI-driven actions, infrastructure efficiency starts to translate directly into performance and cost.
The economics underneath AI therefore become just as important as the model sitting on top.
From experimentation to production with Microsoft Foundry
Microsoft is also creating a more continuous route between Microsoft Foundry and Azure infrastructure.
New preview capabilities will allow customers to deploy open-weight models and agents from Foundry into their own Azure Kubernetes Service clusters while retaining their existing security, infrastructure and governance policies. Organisations can select GPUs and Azure regions based on requirements including latency, cost and data residency.
That’s particularly important for organisations where a fully managed AI service isn’t appropriate for every workload.
Azure’s AI infrastructure is also becoming more agent-aware.
Agent Grid is designed to secure interactions between agents, models, tools and Azure services through centralised policy enforcement. Meanwhile, an Agent Sandbox extension for AKS will provide VM-level isolation for agent-written code. These capabilities remain in preview.
The underlying message is simple:
An AI architecture needs to be designed for production from the beginning.
Your AI strategy is also a data strategy
Infrastructure is only one side of the equation.
An agent can only understand a business if it has access to the right business context.
That’s why Microsoft’s investment in data is so important to the wider AI story.
Azure HorizonDB is now generally available as a fully managed, PostgreSQL-compatible database designed for mission-critical applications, combining enterprise security and resilience with capabilities including vector search. Microsoft is also previewing AI model management, AI Functions and AI Pipelines that bring AI closer to operational data.
At the same time, Microsoft Fabric is increasingly acting as a governed data foundation underneath analytics, applications and agents. Fabric apps, for example, can reuse existing semantic models, warehouses and lakehouses as well as trusted business definitions and access controls.
That matters because an organisation with fragmented, poorly governed or inaccessible data isn’t going to solve the problem simply by deploying a better AI model.
As agents become more capable, the quality and accessibility of the context behind them becomes a competitive advantage.
Operations are becoming agentic too
AI isn’t only something Azure needs to host.
It’s increasingly helping operate Azure itself.
Microsoft announced expanded agentic capabilities across Azure Monitor, Azure Copilot and Azure SRE Agent. These are designed to give operational agents more context to help investigate problems, identify root causes and recommend or automate remediation while retaining human oversight.
Azure Monitor health models, now generally available, can transform telemetry into health context that represents application architecture, dependencies and business priorities. This can then give agents more meaningful context when investigating a problem.
It’s another example of the movement from AI as a standalone application towards AI embedded throughout the technology estate.
Cost needs to become part of AI design
Getting AI into production introduces another issue: consumption.
Cloud FinOps already taught organisations that the ease of spinning up workloads can create unexpected bills if consumption isn’t visible and actively managed.
AI magnifies that issue.
Every model call, agent action and piece of infrastructure has a cost. Choosing the right model, determining which tasks genuinely require AI reasoning and designing efficient architectures can therefore make a significant difference once usage reaches scale.
So AI FinOps can’t be something added once the bill arrives.
It has to become part of architecture and governance from the beginning.
The ANS angle: making AI production-ready
At ANS, this is where we see the conversation moving.
Customers aren’t only asking “What can we do with AI?”
Increasingly, the question is “How do we make this work securely, reliably and commercially at scale?”
ANS can help organisations build the foundations behind production AI across Azure, Microsoft Fabric, data, security and governance.
And through managed services, we can continue optimising those environments once they’re live, helping customers manage cloud and AI consumption, improve performance, strengthen security and evolve their architecture as Microsoft’s platform changes.
Because getting a successful proof of concept running is only the start.
The organisations that get lasting value from AI will be those that can industrialise it, with the infrastructure, data, controls and operating model to support AI in production.
Talk to ANS about building an Azure, AI and data platform ready to take AI from pilot to production.

