FabCon Europe 2026 and Fabric in GCC High Go Live the Same Week: The Seven Microsoft Fabric Decisions to Make Before Your Capacity Renews

What changed in one week, and why the bill is the story

EPC Group's Microsoft Fabric practice reads the Barcelona announcements as a billing-model shift first and a feature release second. Arun Ulag's September 28 post lists the headline items: Fabric IQ available in Microsoft Copilot Chat and Cowork today, with “no additional AI token costs for Fabric IQ in Copilot”; Fabric IQ coming to the Code experience through the Frontier program; agentic app creation in Power BI Desktop arriving in preview “in the coming weeks” for Power BI Pro, Premium Per User and Fabric capacity customers, with “1 GB per app” of Fabric database capability included for Pro and PPU users at no additional cost; a data engineering agent in preview built on the Osmos acquisition; an operations agent in preview with pre-approved actions and root-cause analysis; observability across workspaces in preview; Google BigQuery mirroring generally available; Salesforce Data Cloud 360 bidirectional integration in preview; on-demand billing “expanding across Fabric workloads with automatic scaling”; and Fabric zero-provisioned, described as getting started “without upfront capacity commitment.”

Six days earlier, and easy to miss, Microsoft's US Government blog announced Microsoft Fabric in GCC High: public preview from September 2, 2026, general availability October 1, 2026, with the caveat that “feature availability at public preview and general availability will vary by workload and expand over time” and that “not all Microsoft Fabric capabilities will be available at launch.”

Read together, the two announcements say the same thing to two audiences. To the commercial CIO: Fabric is moving from a fixed capacity you size once a year to a metered platform whose cost follows usage — Copilot questions, agents, Spark jobs, overage — and the levers that control that cost are policy settings you must now own. To the regulated CIO in defense, federal contracting and state agencies: the wait for a compliant landing zone is over on October 1, and the same cost levers apply from day one.

The seven decisions

1. Decide who owns Fabric spend — before on-demand billing decides for you. Fabric capacity was a line item: buy an F SKU, watch the Capacity Metrics app, add capacity when throttling bites. Microsoft's own description of capacity overage — an opt-in that bills excess usage “at 3x the pay-as-you-go rate” rather than throttling, with per-capacity 24-hour limits set by the administrator — and the Barcelona expansion of on-demand billing move Fabric toward the FinOps pattern Azure teams already know: consumption, budgets, alerts, chargeback. Name the owner (usually the data platform lead) and the reviewer (finance) this month. A capacity that nobody owns will be the one that opts into overage “to stop the tickets” and never opts out.

2. Decide your overage posture per capacity, in writing. Overage is the right answer for a production capacity that runs month-end close and cannot tolerate a paused job; it is the wrong default for a development capacity or a self-service analytics capacity where a runaway notebook should be throttled, not paid for at three times the rate. Microsoft lets you opt in per capacity and cap the daily overage; use both. Write the posture into the capacity's name or description so the next admin sees it.

3. Decide whether Copilot questions run against governed semantic models — or against everything. Fabric IQ in Copilot Chat lets a business user “ask business questions in Copilot Chat and get answers grounded in governed Power BI reports.” The value is real only if “governed” means something in your tenant: endorsed (certified) semantic models, sensitivity labels applied, and the Purview controls Microsoft highlighted at FabCon — sensitivity labels and DLP flowing into Copilot experiences, DLP-based access restriction, and an administrator “Restrict from Copilot” control — actually configured. The decision is binary and belongs to the data governance owner: certify the twenty models that matter and restrict the rest from Copilot, or accept that Copilot will answer from whatever it can reach.

4. Decide what the agents are allowed to do. The operations agent Microsoft previewed runs “pre-approved actions” and root-cause analysis; the data engineering agent builds pipelines. Both are useful and both are new principals in your tenant. Treat them the way you would treat a new service account: an approved-actions list, a workspace scope, an owner, and a log you review. The organizations that struggled with Copilot Studio agents in 2025 struggled because nobody wrote the approved-actions list first.

5. Decide your Power BI Pro and PPU strategy now that “Fabric Apps” reach Pro users. Agentic app creation in Power BI Desktop with 1 GB of Fabric database per app at no additional cost for Pro and PPU users means Fabric capabilities will appear in business units that have never had a Fabric capacity. That is good for adoption and bad for sprawl. Decide whether Pro users may publish Fabric Apps freely, only to designated workspaces, or not until the governance owner in decision 3 has certified the underlying models.

6. Decide when — not whether — to consolidate Dataflow Gen1, Power BI Premium P-SKU habits and shadow lakehouses onto Fabric. Zero-provisioned Fabric removes the last excuse for the “we will look at Fabric next fiscal year” posture, because starting no longer requires a capacity purchase. Inventory the Gen1 dataflows, the Premium-era refresh schedules and the departmental lakehouses that grew up outside the platform team, and set a consolidation order by business criticality. EPC Group's consolidation guidance for Dataflow Gen1 to Gen2 migration covers the mechanics; the decision here is sequencing and ownership.

7. If you are a GCC High tenant, decide your first Fabric workload for October 1 — and which ones to wait for. Microsoft is explicit that not every Fabric capability lands in GCC High at GA and that availability expands over time. The right first workload is the one already causing an audit finding or a manual data hand-off between systems of record: a Power BI reporting estate that still refreshes from file shares, or a Sentinel and Defender telemetry archive that needs a governed analytics layer for CMMC evidence. Check the supported-features list Microsoft points to before committing a workload that depends on a feature still in preview.

The 30-day sequence

Days 1–5 — Read the meter. Open the Fabric Capacity Metrics app for every capacity. Record utilization, throttling events, and the top ten items by compute for the last 30 days. Export the workspace list with owners; every workspace without a named owner is a finding.

Days 6–10 — Set the posture. Apply decision 2 to every capacity (overage on or off, daily cap, written reason). Put a budget alert on the Azure subscription that carries the Fabric resource. Name the spend owner and reviewer from decision 1.

Days 11–20 — Govern the answer surface. Certify the semantic models that Copilot Chat and Fabric IQ should answer from; label them; configure the Purview DLP policies and the Restrict-from-Copilot control for the rest. Write the approved-actions list for any agent in preview in your tenant and scope it to one workspace.

Days 21–30 — Decide the consolidation order and the GCC High first workload. Produce the Gen1-dataflow and shadow-lakehouse inventory with a business-criticality rank. If you are GCC High, select the first workload against Microsoft's supported-features list and schedule a 60-day pilot with an exit criterion.

The mistakes EPC Group keeps seeing

• Treating overage as capacity planning. Overage at three times the pay-as-you-go rate is insurance for spikes. If it triggers every day, the capacity is undersized and the fix is a larger SKU or a reservation, not a bigger daily cap.
• Letting Copilot answer from unlabeled models. “Grounded in governed Power BI reports” is only as true as the certification and labeling behind it.
• Approving agents without an approved-actions list. An agent that can run “pre-approved actions” needs someone to have pre-approved them.
• Starting the GCC High move with the hardest workload. Feature availability varies by workload at GA; pick the workload that is fully supported today and already hurts.
• Buying capacity to solve a governance problem. Most throttling complaints we are asked to fix are runaway refreshes and unbounded notebooks, not a shortage of compute.

Frequently asked questions