Fabric capacity SKU reference

F2 to F2048 — Capacity Units, the Power BI SKU equivalence, autoscale behavior, and which workloads each tier realistically supports.

A Fabric capacity is bought as an F SKU. The number is the Capacity Units (CU) it provides. Every workload — Spark, Warehouse, Pipelines, Direct Lake, Eventstream, KQL — draws from that shared pool, smoothed over time.

Pair this with the CU cost & capacity calculator to estimate which tier a given workload needs. Prices below are illustrative pay-as-you-go, USD, and vary by region — always confirm in the Azure pricing calculator.

The SKUs

SKUCUsPower BI SKU equiv.Spark vCores (base / burst)Rough PAYG $/hrTypical fit
F224 / 20~$0.36Dev/test, tiny workloads, learning
F448 / 24~$0.72Small team, light pipelines
F88EM116 / 48~$1.44Small prod, a few models
F1616EM232 / 96~$2.88Departmental analytics
F3232EM364 / 192~$5.76Mid prod, mixed Spark + Warehouse
F6464P1128 / 384~$11.52Free Power BI viewing for report consumers; common prod baseline
F128128P2256 / 768~$23.04Large prod, heavy concurrency
F256256P3512 / 1536~$46.08Enterprise, many concurrent Spark jobs
F512512P41024 / 3072~$92.16Very large enterprise
F10241024P52048 / 6144~$184
F204820484096 / 12288~$369

Spark vCore figures are approximate defaults and depend on pool/node configuration; "burst" reflects the higher transient allocation Spark can take.

The F64 line

F64 (and above) removes the per-user Power BI Pro license requirement for viewing content in that capacity's workspaces. Authors still need Pro/PPU. Below F64, every viewer needs a Pro license. This often makes F64 cheaper than F32 + dozens of Pro seats — model both.

Autoscale, bursting, smoothing, throttling

  • Smoothing: background jobs (scheduled notebooks, pipeline runs, model refreshes) are spread over ~24h; interactive operations over ~5 minutes. A short spike doesn't require a SKU sized for the peak instant.
  • Bursting: a single job can transiently use more CU than the SKU nominally provides; the overage is smoothed back against the capacity.
  • Throttling: sustained overuse moves the capacity through interactive delay → interactive rejection → background rejection. The Capacity Metrics app shows how close you are.
  • Autoscale (Spark): optional — bills extra Spark capacity per-second beyond the base SKU instead of throttling. Good for spiky Spark, a cost risk if left unbounded.

Reserved vs pay-as-you-go

  • Pay-as-you-go: per-second billing, pause/resume anytime. Good for dev/test and workloads you can pause outside business hours.
  • Reserved (1-year): ~40%+ cheaper, but you commit and can't pause for credit. Use for the steady-state production baseline; keep a small PAYG capacity for burst/dev.

A common pattern: reserved F64 baseline + a PAYG F-SKU for dev that's paused nightly, + Spark autoscale with a ceiling for month-end spikes.

Choosing

  1. Estimate peak CU for your heaviest concurrent workload with the calculator.
  2. Add headroom for concurrency (reports + pipelines + ad-hoc Spark at the same time), typically 30–50%.
  3. If you have many report viewers, check whether F64 beats a smaller SKU plus Pro licenses.
  4. Start one tier up from your estimate for the first month, watch the Capacity Metrics app, then right-size down.

Stay ahead of Fabric changes

Fabric runtime changes, API updates, and deprecations. No spam, unsubscribe anytime.

On this page