Capacities and licensing

The second hub session, on Wednesday, October 21, was about money. Dana Whitcombe, the chief financial officer, had asked for a monthly figure before the pilot started in January, and Nikolai wanted Ines and Keiko to understand how Fabric is paid for before he gave her one. Fabric has two separate things that cost money, and confusing them is the most common licensing mistake: capacity, which pays for compute, and per-user licenses, which decide what each person can do with Power BI content.

Capacity units and F SKUs

A Fabric capacity is bought as an Azure resource. In the Azure portal, an administrator creates a Microsoft Fabric capacity in a subscription, a resource group and a region, chooses a size, and names one or more capacity administrators. Sizes are called F SKUs, and each provides a fixed number of capacity units, written CU. The number in the name is the number of units: an F2 has 2 CU, an F64 has 64 CU, and the range runs through F2048 and, at the very top, larger sizes up to F8192. Each step doubles the size of the one before.

A capacity unit is an abstract measure of compute. Every operation consumes CU-seconds: a report query might use a few, a large Spark job many thousands. The capacity's size sets how many CU-seconds are available in each second. An F16 can deliver 16 CU-seconds every second, which is 57,600 in an hour and 1,382,400 in a day. Lesson 11 explains how Fabric spreads the cost of an operation over time, so that a short burst above the size does not immediately fail.

F SKUs are billed per second, with a one-minute minimum, and there is no commitment: a capacity can be scaled up for a busy month and down again, or paused when nobody needs it. While a capacity is paused it is not billed for compute, but the workspaces assigned to it cannot run anything, and their content cannot be viewed. Storage in OneLake is billed separately from compute, by the amount stored. For steady workloads, Azure offers a yearly reservation of capacity units that costs noticeably less than paying as you go. The older P SKUs, bought through Microsoft 365 as Power BI Premium, are being retired in favor of F SKUs.

The trial capacity

Fabric offers a trial capacity that lasts 60 days and provides 64 CU, the same size as an F64. Nikolai started one on October 19, so it will end on December 18, 2026. A trial is attached to the person who starts it, who becomes its capacity administrator, and workspaces can be assigned to it like any other capacity. It is ideal for exactly what the hub sessions are doing: learning, building a design and testing load times. It is not meant for production. When it ends, workspaces on it must be moved to a paid capacity, or their Fabric items stop working.

Per-user licenses

Separately from capacity, every person who uses Fabric has a per-user license. There are three.

The key line in that table is the F64 threshold. On an F64 or larger capacity, people with only a free license can view Power BI reports shared with them, so a company with many report readers pays for capacity but not for a Pro license per reader. On anything smaller than F64, every person who views a Power BI report needs a Pro or PPU license. People who create and publish Power BI content need Pro or PPU on any capacity. Licenses for engineers who only build lakehouses and pipelines are less of a concern: a free license is enough for non-Power BI items on an F capacity.

Regions and data residency

A capacity lives in one Azure region, and the data in workspaces assigned to it is stored in that region. The tenant also has a home region, where tenant metadata is kept; for Harrow & Finch's Microsoft 365 tenant it is in the United States. A company can create capacities in other regions, called multi-geo, to keep data close to users or to meet residency rules. Moving a workspace with Fabric items between capacities in different regions is restricted, so the region should be decided before data is loaded.

For the hub, the capacity will be in Central US, matching the rest of the Azure estate. There is no general federal rule on where data must be stored, but payment card data follows PCI DSS and state privacy laws such as Iowa's consumer data protection act expect the company to know where customer data goes; choosing one region keeps the conversation simple.

Sizing the pilot

Nikolai's approach to sizing was to measure, not guess. During the trial he ran the nightly load, a full refresh of the Sales Overview semantic model and a simulated morning of store managers opening reports, and watched consumption in the Fabric Capacity Metrics app, covered in lesson 11. The busiest hour, around 7 a.m. when store managers check yesterday's sales, used a little under 16 CU on average once smoothing was taken into account. The nightly load ran comfortably in the background. On that evidence an F16 would carry the pilot, with an F32 as the fallback if the busy hour grew.

The harder question was readers. The pilot will have 140 people reading reports: 69 store managers and assistant managers across 23 stores, and 71 at head office. On an F16 every one of them needs a Pro license; on an F64 none of them does. That makes the comparison a simple sum, and the break-even number of readers is worth knowing before the pilot grows.

# Illustrative prices only, not Microsoft's price list
CU_HOUR = 0.20       # dollars per capacity unit per hour, pay-as-you-go
PRO_SEAT = 14        # dollars per Pro user per month
HOURS = 730          # hours in an average month

def capacity_cost(cu):
    return cu * CU_HOUR * HOURS

readers = 140
f16 = capacity_cost(16) + readers * PRO_SEAT
f64 = capacity_cost(64)
print(f"F16 capacity alone: {capacity_cost(16):,.0f} dollars per month")
print(f"F16 plus {readers} Pro users: {f16:,.0f} dollars per month")
print(f"F64, readers need no Pro seat: {f64:,.0f} dollars per month")
break_even = (capacity_cost(64) - capacity_cost(16)) / PRO_SEAT
print(f"Readers at which F64 costs the same: {break_even:.0f}")

F16 capacity alone: 2,336 dollars per month
F16 plus 140 Pro users: 4,296 dollars per month
F64, readers need no Pro seat: 9,344 dollars per month
Readers at which F64 costs the same: 501

At these illustrative prices the pilot costs about 4,296 dollars a month on an F16 with Pro licenses, against about 9,344 on an F64, and the F64 only becomes cheaper beyond about 500 readers. The real prices differ by region and change over time, and the Azure pricing calculator and Microsoft's Fabric pricing page are the places to check them. The method, though, is the one Dana asked for: compare capacity plus licenses against the larger capacity, at the number of readers you actually have. Two other things push toward F64 even below the break-even: some features, such as Copilot in Fabric, have their own capacity requirements, and a larger capacity gives more headroom. Neither applied to the pilot.

Assigning workspaces to capacities

A workspace runs on exactly one capacity at a time, and it can be moved. Moving a workspace from the trial to the paid capacity in January will be a setting change in each workspace, provided both capacities are in the same region. Capacity admins can allow specific users or groups to assign workspaces to their capacity, so not everyone can put a workspace on the company's paid compute. A small company might run everything on one capacity. A larger one often separates development from production, so that an engineer's runaway notebook in development cannot slow the production reports. Nikolai noted that idea for later: during the pilot, the cost of a second capacity would be hard to justify, but once production reports were in daily use he would propose a small F4 or F8 for development workspaces.

Pay as you go or reserve

Pay-as-you-go pricing suits a pilot because nothing is committed: if the pilot is canceled in March, the capacity is deleted and billing stops. A reservation suits a capacity that will run all day, every day, for at least a year, because the discount is earned only if the reserved units are used. A reservation applies to a number of capacity units in a region rather than to a named capacity, so a company that reserves 16 units can apply them to any F16 in that region, or to two F8s. Pausing a reserved capacity saves nothing, since the reservation is already paid. The usual pattern, and the one Nikolai recommended, is to start pay as you go, watch real usage for a few months, and reserve only the baseline that production clearly needs.

What to tell finance

Nikolai's note to Dana, sent on November 10, kept to three points. First, the pilot will run on an F16 capacity in Central US, paid as you go from January 4, 2027, so it can be scaled or paused at no penalty. Second, the 140 pilot readers need Pro licenses, which many already have through the company's Microsoft 365 plan; the rest will be bought. Third, if readers grow past about 500, or if the company wants features that need a larger capacity, an F64 with a yearly reservation should be compared again. Dana approved the pilot budget the same week.