Local when it matters. Cloud when it counts. | THRSP1065
Ajay Mungara presents Intel and Microsoft’s approach to combining local AI execution with Azure, using local models for speed, privacy, and efficiency while drawing on cloud capabilities for more sophisticated reasoning and enterprise scale.
Choosing where agent workloads run
The session focuses on routing agentic AI workloads rather than treating local and cloud execution as an either-or decision. Routing choices account for three factors:
- Privacy: How data-handling requirements affect where a workload can run.
- Cost: How model selection and token consumption affect overall spending.
- Capability: Whether a task needs a smaller model or more sophisticated reasoning.
Mungara discusses complexity, semantic, elastic, and escalation routing. Escalation routing starts with smaller models and moves to more capable models when needed, helping reduce token costs.
Agent workflows and cost optimization
An AI-generated presentation example uses 59 agents and includes cost tracking. The session also examines why total token spending can rise even as token prices fall, making workload optimization important.
Other topics include:
- Why coding tasks are well suited to AI agents and verification workflows.
- Why most agent workflows run on CPUs rather than large reasoning models.
- How enterprise agents orchestrate work across multiple systems.
Enterprise deployment considerations
The closing segment addresses enterprise policies, security, confidential computing, and practical deployment guidance. These considerations sit alongside privacy, cost, and model capability when deciding how to combine local execution with Azure services.
Chapters
- 00:00 — Introduction to AI agents, token costs, and workload routing
- 01:03 — AI-generated presentation example using 59 agents and cost tracking
- 02:50 — Rising token spending despite falling token prices and the need for optimization
- 03:41 — Why coding tasks are ideal for AI agents and verification workflows
- 04:53 — Most agent workflows run on CPUs rather than large reasoning models
- 06:03 — Enterprise agent workflows and orchestration across systems
- 07:16 — Routing decisions based on privacy, cost, and capability
- 08:10 — Complexity, semantic, elastic, and escalation routing
- 11:27 — Reducing token costs by using smaller models first
- 14:19 — Enterprise policies, security, confidential computing, and deployment guidance
Session details and resources
- Event: Microsoft Ignite 2026
- Session: THRSP1065
- Track: Windows
- Format: Theater
- Level: 300 — Advanced
- Language: English (US)