CU cost & capacity calculator

Capacity Units (CU) are Fabric's single currency. Every engine — Spark, SQL, pipelines, Direct Lake — draws from the same capacity. The hard part of planning is that the same logical workload costs differently depending on which engine runs it and how you size the compute.

Use the calculator to get a first-order comparison, then validate against the Capacity Metrics app for your tenant.

The model

CU-seconds/day = effectiveVCores × activeSecondsPerDay × cuPerVCore × concurrency
avgCUs         = CU-seconds/day ÷ 86,400
peakCUs        = effectiveVCores × cuPerVCore × concurrency
min F SKU      = smallest F capacity with CU count ≥ peakCUs
$/month (PAYG) = skuCUs × ratePerCUHour × 730

Default assumptions

How to read the result

• Average CUs tells you how loaded a capacity feels over 24h — useful for "will this fit alongside my other workloads".
• Peak CUs drives the minimum SKU. Fabric smooths background jobs over 24h but interactive spikes over ~5 minutes; a high peak with low average still needs the bigger SKU unless the work is purely background.
• Concurrency multiplies peak. Ten parallel Spark jobs on a 32-vCore pool do not fit on an F16.

Worked example: nightly batch vs. warehouse ELT

You have 500 GB/day of transforms. Option A: a Spark notebook on 4× medium (32 vCores) running 90 min. Option B: T-SQL INSERT…SELECT in a warehouse bursting to ~16 vCores for 90 min.

1. Set Active compute / day to 90, Concurrency to 1.
2. Spark: Total vCores 32. SQL warehouse: Effective vCores 16.
3. Compare the monthly figures. Spark's larger pool costs more per second but may finish faster in practice — shorten its active minutes and re-check.

The lesson the calculator makes concrete: right-size the pool and cut active time (fewer, bigger files; the Native Execution Engine) before you go shopping for a bigger capacity.