P1 to F64: A Right-Sizing Guide Before You Migrate
F64 is the same size as P1 — but same size is not the same as right size. How to measure, price and plan the move off Power BI Premium.
Core Engineering Takeaways
- •When a P SKU term ends there is a 30-day grace period, throttling from day 31, and rejected operations from day 91.
- •F64 matches P1 in size and keeps the free-viewer benefit, which is why it is the default mapping.
- •A 1-year F64 reservation lists at about $5,003/month vs $8,410 pay-as-you-go in US East — roughly 40.5% less.
- •Measure peak interactive and background load before choosing: many P1 estates were sized for licensing, not load.
1. What happens when your P SKU ends
Microsoft no longer sells Power BI Premium P SKUs. Each existing P subscription runs until the end of its current agreement term — and then the clock starts:
- Days 1–30: a grace period. Everything keeps working.
- From day 31: access is throttled — interactive operations are delayed.
- From day 91: all operations are rejected. Your data is retained but inaccessible until workspaces are moved to a Fabric F SKU or the capacity is deleted.
If you have an active Enterprise Agreement, you can keep renewing existing P capacity annually until the EA term ends. Customers with expiring Enterprise Agreements or Microsoft Cloud Agreements cannot add or buy new P capacity.
There is no automatic conversion. Workspaces are reassigned to the new F capacity — in the Fabric admin portal, or with Microsoft's capacity migration notebook for larger estates.
2. Why F64 is the default — and when it is not enough or too much
F64 has 64 capacity units and the same Power BI capacity size as P1 (8 v-cores). Crucially, it is the smallest SKU where users with a free Fabric licence and the Viewer role can consume content, just as they could on P1. Below F64, every viewer needs a Pro ($14/user/month) or PPU licence.
That licensing fact is usually what decides the size. At list price, a reserved F64 costs about the same as 358 Pro licences (601 at pay-as-you-go). If you have thousands of viewers, F64 is the floor regardless of load. See F64 vs Pro licences for the full break-even table.
Two situations change the answer:
- Too small: F64 allows 25 GB of memory per semantic model, including refresh overhead. A model that needs more than that during a full refresh — or a capacity that already throttles on P1 — needs F128 or optimisation, not a like-for-like move.
- Too big: if you have relatively few viewers, a mostly idle P1, or separate dev/test workloads, part of the estate may run cheaper on a smaller, pausable pay-as-you-go capacity.
3. Measure what you actually use
The Fabric Capacity Metrics app (it also covers P capacities) is the source of truth. Pull several weeks of data that include your busiest periods — month-end, Monday mornings, quarterly reporting — and look at:
- Interactive peaks: the highest utilisation from report queries, and when it happens.
- Background load: refreshes and other scheduled work. Fabric smooths background operations over 24 hours, so a few heavy refreshes can quietly occupy a large share of the capacity all day.
- Throttling events: any interactive delay or rejection already happening on P1. Moving these problems to F64 moves them unchanged.
- Largest models: memory of the biggest semantic models during refresh, against the per-model limit of the target SKU.
My free capacity estimator gives a first estimate from model sizes, refresh schedule and peak users, which is useful before you have metrics access or to sanity-check what the metrics show.
4. Fix what would force a bigger SKU
The cheapest capacity is the one you no longer need. Before choosing a size, check the usual causes of inflated load:
- Refresh collisions. Many models scheduled at 06:00 create a spike that sets your peak. Staggering them often flattens it dramatically.
- Oversized models. High-cardinality columns (GUIDs, datetime stamps), unused columns, and Auto Date/Time tables inflate memory and refresh time. VertiPaq Analyzer shows where the memory goes.
- Expensive DAX. A handful of measures that force row-by-row evaluation can dominate interactive load. Server Timings in DAX Studio shows which.
- Full refreshes that should be incremental. Reloading five years of history nightly to add one day.
It is common for these fixes to take a capacity from "needs the next size up" to comfortably inside its current size.
5. Price it: reserved vs pay-as-you-go
In US East, Fabric capacity lists at $0.18 per CU-hour pay-as-you-go, or $938 per CU per year reserved. For F64 that is about $8,410/month pay-as-you-go or $5,003/month reserved — roughly 40.5% less. Prices vary by region and agreement.
A reservation is the right choice for an always-on production capacity — but only after right-sizing, because it commits you to the size for a year. Pay-as-you-go capacity can be paused, which makes it attractive for dev/test or batch capacities that are idle much of the day. Reserved vs pay-as-you-go works through the break-even.
6. Migration checklist
- Confirm your P SKU term end date and agreement type (EA or not).
- Pull several weeks of Capacity Metrics data and identify peaks, background load, throttling and largest models.
- Fix refresh collisions and the worst model/DAX issues.
- Choose the F SKU and pricing model; buy the capacity in the right region.
- Check dependencies: gateways, Power BI Report Server licensing (still available with a Fabric capacity reservation or SQL Server Enterprise with Software Assurance), embedded apps and service principals.
- Reassign workspaces in waves, lowest risk first, and validate refreshes and reports after each wave.
- Monitor the new capacity closely for the first few weeks before releasing the P SKU.
If your term end date is close, the Premium → Fabric migration sprint covers measurement, right-sizing, migration and validation at a fixed price.
Leaving Premium? Size it before you move it
Get a first recommended F SKU from your model sizes and refresh schedule, or hand the migration to a fixed-scope sprint.
Get the next Engineering Note
Practical DAX, semantic model and Fabric notes. 1–2 a month, unsubscribe any time.