Microsoft Build 2026: Foundry, FOCUS 1.3, and Agent Cost Trace

Build 2026 (19–22 May in Seattle) was the first conference where Microsoft stopped talking "about AI" and started talking "about the cost of AI". That is a shift I have been waiting at least a year to see – and as the founder of CostSentry.AI I have a strong interest in unpacking it.

Here are the three announcements that genuinely matter from an enterprise deployment standpoint, and three that dominated the headlines without affecting my work much.

1. The Microsoft Foundry Rebrand: More Than a New Name

The key slide of the keynote: Azure AI Foundry → Microsoft Foundry. The word "Azure" was deliberately dropped – the goal is to unify:

• Azure AI Foundry (model catalog, deployments)
• Copilot Studio (low-code agent builder)
• Semantic Kernel (SDK for orchestration)
• The new Foundry Agent Gateway for multi-cloud routing

Practical impact for customers:

My take: the rebrand on its own is marketing, but Foundry Workload Identity and unified billing are real engineering wins that will save us weeks of integration work at Christie's and Nespresso.

2. FOCUS 1.3 GA: Finally Usable

FOCUS (FinOps Open Cost and Usage Specification) has been a nice idea with poor enterprise adoption since v1.0. FOCUS 1.3 changes that:

New mandatory columns vs FOCUS 1.2:
+ ServiceSubcategory       # e.g. "AI/Agent" under "Foundry"
+ CommitmentDiscountStatus # "Used" / "Unused" / "Expired"
+ EffectiveCost            # cost after discounts, credits, and reservations
+ SkuPriceDetails          # structured price breakdown (compute/storage/network)

The practical difference: the same KQL query against Azure, AWS, and GCP FOCUS exports returns a consistent EffectiveCost. No post-processing, no mapping. For our CostSentry.AI work it means the shadow IT detector can run across all three clouds simultaneously without per-cloud adapters.

Microsoft has promised FOCUS 1.3 GA in Cost Management exports by October 2026. Until then it is preview via the Cost Management Data Explorer API.

# Enable the FOCUS 1.3 preview export (May 2026)
az costmanagement export create \
  --name "focus13-monthly" \
  --type "FocusCost" \
  --scope "subscriptions/<SUB_ID>" \
  --storage-account-id "<STORAGE_ACCOUNT_RESOURCE_ID>" \
  --storage-container "focus-exports" \
  --recurrence Monthly \
  --recurrence-period from="2026-06-01T00:00:00Z" to="2027-06-01T00:00:00Z" \
  --format "Parquet" \
  --schema-version "FOCUS_1.3-preview"

Note --format Parquet – FOCUS 1.3 contains nested structures (SkuPriceDetails) that get flattened away in CSV.

3. Agent Cost Trace: the Killer Feature Most People Will Miss

This was the single most important thing for me at the whole Build and the keynote gave it five minutes. Agent Cost Trace is a new telemetry layer in Foundry that maps:

[User request] → [Agent invocation] → [Tool calls] → [Token usage] → [Downstream APIs] → [Cost]

Instead of the previous "Foundry deployment burned USD 3 200 in May" you get:

{
  "traceId": "abc-123",
  "userRequest": "Generate quarterly report for Q1 2026",
  "totalCostUsd": 0.187,
  "breakdown": {
    "agentTokens": { "input": 4500, "output": 2300, "costUsd": 0.041 },
    "toolCalls": [
      { "name": "search-kb",     "count": 3, "costUsd": 0.012 },
      { "name": "sql-query",     "count": 1, "costUsd": 0.003 },
      { "name": "blob-fetch",    "count": 8, "costUsd": 0.001 }
    ],
    "downstreamLlms": [
      { "model": "gpt-5-mini", "tokens": 12500, "costUsd": 0.130 }
    ]
  }
}

Why this matters: as soon as an agentic workload starts to scale, classical cost allocation through resource tags stops working. The cost is not on a Foundry deployment – it is on a specific user pattern. Agent Cost Trace finally addresses this.

Setup is straightforward:

resource foundryDeployment 'Microsoft.Foundry/deployments@2026-05-01-preview' = {
  name: 'agent-prod'
  properties: {
    model: 'gpt-5-mini'
    costTrace: {
      enabled: true
      destination: {
        type: 'LogAnalytics'
        workspaceResourceId: laWorkspaceId
        tableName: 'FoundryAgentCostTrace_CL'
      }
      sampling: {
        rate: 1.0  // 100% for the first month, then drop to 0.1
      }
    }
  }
}

Price: USD 0.05 per 1000 traces plus standard Log Analytics ingestion. For a typical enterprise deployment of 10 000 requests/day that works out to ~USD 15/month – negligible compared to the value of the data.

What Dominated Headlines but Did Not Move Me

Project Magenta (locally-run Phi models for the edge): Nice demo, but irrelevant for most enterprise clients. If you need edge inference you are already running ONNX or NVIDIA Triton, not locking into the Foundry stack.

Copilot Pages 2.0: A SaaS product for "live document collaboration with AI". Pretty, but it does the same thing as Notion AI and Google Docs with Gemini. No unique enterprise reason to migrate.

Azure DevOps and GitHub convergence: A continuation of the Build 2025 trend, but we still lack a clear migration path from ADO Boards to GitHub Issues with feature parity. We have been waiting on this at Nespresso for two years now.

Three Practical Tasks for You After Build 2026

1. Enable the FOCUS 1.3 preview export in a dev subscription and test your existing KQL queries – several columns will most likely be renamed under you
2. If you have a production Foundry deployment, turn on Agent Cost Trace at 100% sampling for at least a month – the data it surfaces will overturn your cost allocation strategy
3. Prepare for the Foundry Workload Identity migration – at Christie's we plan to start in autumn 2026 once it hits GA

Two AKS Announcements That Did Not Get Enough Air Time

LTS for Kubernetes 1.31: Three years of support (until March 2029) for a ~30% premium over standard. For regulated workloads (banking, healthcare) it is exactly what you need. At Creditas we will be moving all production over.

Cilium as the default CNI: Azure CNI Powered by Cilium is now the default for newly created clusters. Existing clusters must be migrated manually (~5 minutes of downtime per node pool). The real benefit: identity-aware NetworkPolicies without a Calico overlay and 2–3x faster conntrack.

Conclusion

Build 2026 was a coming-of-age conference. AI is no longer "the new thing" – it is a product with bills, governance, and ops problems. FOCUS 1.3 + Agent Cost Trace + the Foundry rebrand are three steps that turn enterprise AI into something you can realistically run at scale.

For me specifically: by the end of June I will rewrite the CostSentry.AI Foundry ingest onto the FOCUS 1.3 schema and the Agent Cost Trace pipeline. If you are dealing with a similar transformation, check out our cloud architecture services or reach out for a Build 2026 walkthrough of your stack.