Microsoft Fabric is no longer “the new thing next to Power BI”. It’s the centre of Microsoft’s data and AI strategy, and what was announced in front of roughly 5,000 people at FabCon Europe in Barcelona (28 September – 1 October 2026) made that very clear.
In this post I pull together what Fabric is today, the most important FabCon announcements, what changed on cost and governance, and where a BI team can realistically start. I’ve marked every feature as GA or preview, because when you’re planning anything for production, that distinction is everything.
What is Fabric?
Fabric is Microsoft’s SaaS platform that brings every analytics workload, from data integration to reporting, together on a single storage layer called OneLake. According to the numbers Microsoft shared at FabCon, it now has more than 40,000 customers (Kanerika).
I find it easiest to think about the platform in layers:
- OneLake: one lake where all data lives in Delta Parquet format. External sources connect through shortcuts (no copy) or mirroring (continuous replication).
- Data preparation: Data Factory (pipelines, Dataflow Gen2, Copy Job), Spark-based Data Engineering, Data Warehouse and SQL database.
- Real time: Real-Time Intelligence, with Eventstream, Eventhouse and Activator reacting to data as it arrives.
- Business context: Power BI semantic models, plus the Fabric IQ ontologies that now sit on top of them.
- Consumption and action: Power BI reports, Copilot, data agents and Fabric Apps.
Power BI is the analytics shop window of all this. An organisation that buys Fabric capacity gets its old Power BI Premium entitlements through F SKUs.
The FabCon Europe 2026 announcements that matter
The common thread was this: instead of hand-writing every line, you describe the outcome, set the guardrails, and let AI do part of the implementation. The table below is my own selection, with status as of October 2026.
| Announcement | Status | What it does |
|---|---|---|
| Fabric IQ in Microsoft Copilot Chat and Cowork | GA | Copilot answers using the business definitions in your Power BI semantic models, respecting existing permissions |
| Fabric IQ MCP endpoint | GA | Exposes trusted Power BI insights to agent-building tools outside Fabric |
| Fabric Runtime 2.0 | GA | Brings together Spark 4.1, Delta Lake 4.2 and Python 3.13 |
| Google BigQuery mirroring | GA | Continuously replicates BigQuery data into OneLake |
| Dataflow Gen1 → Gen2 Upgrade Wizard | Preview | Upgrades Gen1 dataflows to Gen2 in place; there’s no way back |
| Data Engineering Agent (Osmos) | Preview | Plans and runs ETL, migration and tuning work from a described goal |
| Fabric IQ ontologies | Preview | Defines business concepts like customer, product and store, and how they relate, in one shared layer |
| On-demand billing and F0 | Preview announced | Pay for what you use, with no capacity reserved up front |
| Policies in Fabric | Preview | Fine-grained rules by user, item type and workspace |
| Deployment plans | Preview | Ordered, validated deployments with YAML as the source of truth |
| Unified Copilot in Fabric | Private preview | One Copilot that remembers project context across every workload |
Sources: James Serra’s FabCon Barcelona 2026 roundup and Kanerika’s summary.
Note: I’m not covering Direct Lake again here. If you want the details, see my post Direct Lake in Fabric: The Third Way Between Import and DirectQuery.
AI and agents: the new face of Fabric
The real message from FabCon was that AI can only give the right answer if it first has governed data and shared business definitions. Fabric is positioning itself as the layer that provides that context.
Fabric IQ and ontologies. An ontology defines concepts like “customer”, “product”, “store” and “order”, and the relationships between them, in one place. The detail that matters most: you can reuse the DAX measures in your existing Power BI semantic models without rewriting them. You can also describe an ontology in plain language and let AI build it; any change is shown to you for approval before it’s applied.
Data Engineering Agent (Osmos). The engineer describes the outcome they want, the sources and the validation criteria. The agent writes and runs Spark notebooks, creates Delta tables, and adjusts its approach based on the results. You could give it something like “build bronze, silver and gold layers from the sales files, deduplicate the records and check that revenue totals reconcile”. For now, Microsoft recommends it for evaluation, not production.
Data agents. Fabric data agents in Copilot Studio are now GA. With “Build agent with AI”, you no longer hand-write the agent’s instructions and example queries; you describe what it should do.
Fabric Apps. Launched in June 2026, Fabric Apps combines data, functions, storage and hosting so you can ship an app with Entra authentication built in. The open-source framework underneath is called Rayfin. Power BI Pro and PPU users will get Fabric Apps and a database of up to 1 GB per app at no extra cost.
My takeaway: semantic model quality is now an AI problem. If two departments calculate margin differently, Copilot will pick one of them and answer with complete confidence.
Capacity and cost: from F SKUs to F0
Fabric is priced by capacity: you buy an F SKU and every workload shares its Capacity Units (CU). An annual reservation saves 40.5% compared with pay-as-you-go (Microsoft).
The thresholds worth knowing before you decide:
- F64 is the key threshold. Viewing Power BI content without a paid per-user licence is only possible on F64 and above. Below that, viewers need Pro licences too.
- Publishing still needs Pro. Whatever capacity you buy, the person publishing Power BI content needs a Pro licence.
- Bigger options. F4096 and F8192 are now GA.
The cost-related announcements at FabCon:
| Feature | Status | What to watch |
|---|---|---|
| On-demand billing and F0 | Preview announced, rollout soon | “Zero” refers to pre-provisioned capacity, not free compute; consumption still shows up on the bill |
| Capacity overage | GA | Bills temporary overloads instead of throttling; excess CU hours cost three times the pay-as-you-go rate |
| Workspace-level surge protection | Preview | Stops a single workspace from eating the whole capacity |
| Capacity Metrics app improvements | GA | Heatmaps, top consumers and chargeback support |
The real significance of F0 is on the OneLake side: you’ll be able to read and write to the lake without keeping a capacity running. That lowers the cost of starting a pilot. But don’t decide for sustained workloads without comparing against provisioned capacity; a runaway Spark job will now land straight on your invoice.
Governance and DevOps
On a platform where anyone can be a developer, order only survives through rules. This year’s governance features move administration out of settings screens and into code.
- Policies in Fabric (preview). You can write rules based on attributes such as user or group, item type, workspace and capacity. For example, you could allow only the data engineering group to create Lakehouses, Notebooks and Pipelines in production workspaces. Policies are managed from the OneLake Catalog and can be wired into CI/CD through the API.
- Deployment plans (preview). Deployment steps are arranged on a drag-and-drop canvas, with YAML as the single source of truth. Data loads and validation become pre- or post-deployment steps in the plan, so you no longer need a separate manual checklist.
- Git integration (GA). Compare and commit, selective branching, branched workspaces and bulk import/export APIs are generally available. File-level commits are in preview.
- OneLake catalog (GA and preview). Table- and column-level discovery, data preview and the central “Govern” experience are GA. Catalog discovery from Excel and the Search API are in preview.
- OneLake security. The “View users in role” experience lets you audit who can access a table from one screen; inbound and outbound network protection and DLP restrict-access are GA.
My practical advice: before your first agent or Copilot pilot, write down your decisions on workspace naming, metric ownership and who can create what. Policies are the tool for enforcing those decisions, not the decisions themselves.
Where to start: a roadmap for retail
The typical picture in a retail company looks like this: sales and stock in SQL Server, campaign data in Excel files, external data (official statistics, price tracking) in separate scripts. The healthiest way to move to Fabric is not a big bang but a series of small steps that each deliver value.
- Take inventory. List source systems, SSIS packages, dataflows and the reports that depend on them. Agents and migration tools only work properly with a complete inventory.
- Pick your first workload. Something everyone looks at but that needs frequent updates, like daily sales and stock, is a good candidate. Bring the data into OneLake with mirroring or Copy Job and build a medallion (bronze, silver, gold) structure.
- Clean up the semantic model. In your most-used models, remove duplicate measures, give every metric an owner and tighten the naming. Copilot and Fabric IQ will use those definitions exactly as they are.
- Grow capacity based on measurements. Start with a small F SKU or F0 and watch consumption in the Capacity Metrics app. Consider F64 once viewer licensing starts costing more.
- Try previews in a sandbox. Give the Data Engineering Agent, ontologies and Policies their own workspace. Don’t build production plans on anything that isn’t GA.
- Set up governance from day one. Use Git integration and branched workspaces from the start; adding them later is much harder.
Wrapping up
In 2026 Fabric has grown into a mature enterprise platform where data, business context, AI and applications come together under one roof. But a good share of the announcements are still in preview, and AI is only ever as good as the definitions behind it.
My own view: the people who make a difference over the next year won’t be the ones who try every new feature first, but the ones who get their semantic models and governance in order today. Where are you on your Fabric journey?
Sources
- Announcements from the Microsoft Fabric Community Conference — Barcelona 2026 (James Serra, SQLServerCentral)
- FabCon Europe 2026: Major Announcements and What They Mean for Enterprises (Kanerika)
- FabCon and SQLCon 2026 in Barcelona (Microsoft Azure Blog)
- Fabric September 2026 Feature Summary
- Build, deploy, and govern Microsoft Fabric at scale
- Data Engineering Agent overview (Microsoft Learn)
- Fabric IQ ontology overview (Microsoft Learn)
- Policies in Fabric (Microsoft Learn)
- Capacity overage (Microsoft Learn)
- Power BI pricing (Microsoft)
STAY IN THE LOOP
New guides, sorted by topic in a single email.
Every Monday, the week's new posts arrive grouped by topic. No new posts, no email.
1 comment