Microsoft Fabric capacity sizing: how to pick the right SKU

How capacity works

A Fabric capacity is a pool of compute that every workload in your tenant draws from: Power BI rendering, Data Factory pipelines, warehouse queries, notebooks, Copilot. You buy a fixed capacity (an F SKU) and the platform smooths usage across the day, so a short spike does not break anything as long as your average sits within the SKU.

The F SKUs at a glance

Capacities run from F2 (entry-level, suitable for evaluation) up to F2048 (enterprise scale). F64 is the practical floor for most businesses: it unlocks Copilot, lets viewers consume Power BI reports without their own Pro licence, and gives enough headroom for a real workload. Below F64 you are usually still paying for Pro licences for every viewer, which erodes the cost saving.

Pay-as-you-go vs reserved

Capacities can be billed by the second (pay-as-you-go) or by annual reservation. Reservation is roughly 40% cheaper but commits you for a year. The right pattern for most businesses: start pay-as-you-go for the first 60 to 90 days while you measure real usage, then move to reservation once you know what you actually need.

Pause to save money

Pay-as-you-go capacities can be paused outside business hours. For a finance team that only reports during the working week, pausing evenings and weekends can cut the bill by half or more. Reserved capacities cannot be paused, which is the main reason not to over-commit on reservation early.

How to size without overpaying

Three steps. First, list the workloads you will actually run in the first six months (Power BI reports, one or two pipelines, maybe a lakehouse). Second, start at F64 if you want Copilot and viewer licences included, or F8 to F32 if you only need Power BI workloads on a small dataset. Third, watch the Fabric Capacity Metrics app weekly: if usage sits comfortably below 70% you can downsize; if it spikes above 100% you need to either resize, schedule heavy jobs differently, or split workloads across capacities.

Common sizing mistakes

Buying the SKU a vendor recommends without measuring; assuming Power BI Premium pricing translates directly to Fabric capacity; forgetting that pipelines and notebooks consume the same pool as reports; and locking into a reservation in week one. Most over-spend on Fabric comes from one of those four.

Smoothing, throttling, and what they actually mean

One of the most useful things about Fabric capacity is that it smooths short bursts of demand across a 24-hour window. If your overnight pipeline pushes usage to 150 percent for an hour but the rest of the day sits at 40 percent, the platform absorbs that without complaining. Only sustained over-use triggers throttling, which slows interactive queries to nudge you towards a bigger capacity. The practical implication: do not panic at the first 100-percent spike. Look at the 24-hour average. If it is comfortable, your sizing is fine.

Splitting workloads across multiple capacities

For bigger businesses, one capacity for everything is rarely the right answer. A common pattern: one capacity for production reporting (sized for stable, predictable use), and a separate, smaller capacity for development and data engineering (where bursts are normal). That isolation means a long-running data science notebook cannot accidentally slow down the executive dashboard. It also makes chargeback to departments cleaner if your business cares about that.

When to resize and how often to review

Treat capacity as something to review, not set and forget. A sensible cadence: weekly during the first quarter (to catch surprises early), monthly once you are stable, and a full re-sizing review every six months. Resizing is a click - capacities can be scaled up or down in seconds with no downtime - so there is no excuse for paying for capacity you are not using. The Fabric Capacity Metrics app is the single most valuable tool you have here, and the team responsible for the platform should be in it every week.

For help sizing and rolling out Fabric, see our Microsoft Fabric service page.