Power BI Capacity Throttling and CPU Spikes: Causes and Fixes on Fabric
Utilization above 100% is not throttling; the stages are.
Power BI capacity throttling is one of the most misread problems in enterprise reporting. Dashboards slow down, refreshes fail and the first instinct is to buy a bigger SKU, when the real cause is usually a few heavy items or a misread chart.
This guide, refreshed for Fabric, explains how smoothing and the throttle stages work, how to confirm throttling in the Capacity Metrics app, what to fix first and when resizing is the right call. It also covers what the Premium retirement changes, with figures taken from Microsoft Learn at the time of writing.
What Power BI Capacity Throttling Is and How Fabric Smoothing Works
Capacity is measured in capacity units (CUs). Throttling occurs when operations use more CU seconds than your SKU allows. Fabric applies it at the capacity level, so one capacity can be overloaded while another runs normally.
Two features soften spikes. Bursting lets an operation use more compute than the SKU provides for a short time. Smoothing then spreads the CU cost over later timepoints, each 30 seconds long. Interactive operations, such as a user opening a report, are smoothed over a minimum of five minutes and up to 64 minutes. Background operations, such as scheduled refreshes, are smoothed over 24 hours.
That corrects a common simplification. Smoothing does not use a 30-second window. The 30 seconds is the size of one timepoint, and the smoothing period is minutes for interactive work and a full day for background work.
Throttle Stages: When Power BI Capacity Throttling Starts
Throttling does not start the moment utilization passes 100%. A capacity has overage protection that lets it consume 10 minutes of future capacity without throttling, and then it progresses through three stages.
The three throttle stages.
| Smoothed usage ahead | Stage | What users see |
|---|---|---|
| Up to 10 minutes of future capacity | Overage protection | No throttling |
| 10 to 60 minutes | Interactive delay | New interactive operations are delayed by 20 seconds |
| 60 minutes to 24 hours | Interactive rejection | New interactive operations are rejected |
| More than 24 hours | Background rejection | All new requests are rejected |
Throttling only affects operations requested after it begins. Operations already running finish. Microsoft also notes that it may change the policy, so check the current page. For a deeper explanation of the mechanics, see our guide to Microsoft Fabric throttling, bursting and smoothing.
What the Premium Retirement Changes for Power BI Capacity
If you came here from a Premium P SKU, the framing has changed. Microsoft is retiring P SKUs at the end of each agreement term and no longer sells new ones. After a subscription ends there is a 30-day grace period, access is throttled from day 31 and all operations are rejected from day 91. Buy the F SKU first, reassign workspaces, validate, and only then cancel the P SKU.
The practical effects on spikes are these. F SKUs scale up and down and can be paused, which gives you levers P SKUs did not have. Power BI viewers need only a free licence on F64 and larger. And a capacity can now share compute with Fabric workloads such as Spark and warehouses, which can raise CU use for report users. Our Premium to Fabric case study shows one move, and our Fabric SKU guide for executives helps pick the starting size.
Confirm Power BI Capacity Throttling in the Capacity Metrics App
Start with evidence. When a capacity rejects requests you see specific errors: a CapacityLimitExceeded status, a message that your organization's compute capacity has exceeded its limits, or a message that a model cannot be loaded because of capacity limits. Slow reports without those errors are often a design problem, not throttling.
In the Fabric Capacity Metrics app, use the Throttling chart and the System events table on the Compute page. Utilization above 100% is an overage, but it is not automatically throttling, so judge by the Throttling chart. Drill through to the timepoint page to see which interactive and background operations contributed most, with the item, operation and user. The Overages tab shows carryforward and burndown, which tells you how long the capacity needs to recover.
Capacity admins can also set up email alerts for capacity thresholds, so you hear about a problem before users do. Our guide to Microsoft Fabric capacity utilization shows how to read the app.
Root Causes of Capacity Spikes
Once throttling is confirmed, rank the operations by CU use and look for patterns. The same few causes appear again and again.
| Cause | What to look for | Typical fix | Effort |
|---|---|---|---|
| Many large refreshes at once | Background operations are smoothed over 24 hours, so look at total daily CU use | Incremental refresh; trim refresh frequency; review which models need refresh | Low |
| Inefficient DAX or model design | Interactive spikes on a few semantic models | Review the heaviest measures and iterators; simplify relationships and columns | Medium |
| Power Query that does not fold | High background CU on refresh | Fold to the source or move transformation to dataflows, pipelines or a warehouse | Medium |
| High concurrency on large models | Interactive CU peaks during business hours | Aggregations, composite models or Direct Lake on the heaviest models | High |
| Unchecked report complexity | Heavy visuals and many interactions per page | Author guardrails, query reduction and a publishing review | Medium |
| Other Fabric workloads on the same capacity | Spark, warehouse or pipelines raise CU use for BI users | Separate capacities for BI and engineering | Medium |
| Genuinely undersized capacity | Sustained high throttling after optimisation | Scale the F SKU or add a reservation for steady load | High |
Inefficient DAX, such as nested iterators over large tables, drives interactive CPU load. Power Query that does not fold to the source runs inside the capacity during refresh. High concurrency against a large model without aggregations multiplies the cost of every user interaction. Refresh design still matters, but because Fabric smooths background work over 24 hours, the question is total daily consumption, not just the start time.
- Confirm throttling in the Throttling chart before you change anything.
- Interactive work is smoothed over minutes and background work over 24 hours.
- Rank operations by CU use and fix the heaviest few items first.
- Separate BI and engineering workloads if they compete for one capacity.
- Resize only after optimisation is documented.
Fixes That Reduce Power BI Capacity Throttling
Work through these in order, because the cheapest fixes come first.
- Tune the heaviest semantic models. Review the measures and iterators behind the top interactive operations and simplify the model before touching the SKU.
- Reduce refresh load. Use incremental refresh so each refresh processes less data, and remove refreshes nobody uses.
- Move transformation upstream. Fold Power Query steps to the source, or move heavy shaping into dataflows, pipelines or a warehouse.
- Reduce query cost on large models. Consider aggregations, composite models or Direct Lake for the models with the highest concurrency.
- Govern authors. Set query reduction defaults, limit visual complexity and review heavy reports before publishing. Our Power BI governance platform supports this.
Then use the capacity levers. Throttling is segregated by capacity, so putting BI and data engineering on separate capacities stops one from throttling the other. On F SKUs you can temporarily scale up, pause and resume, or enable capacity overage, which Microsoft says costs three times the normal rate and is meant as a bridge. For steady load, compare pay-as-you-go with a reservation.
When to Resize Instead of Optimise
Resize when throttling stays high after the heavy items are tuned, when growth in users or models is documented, or when you need workload isolation. Each F SKU step doubles the CUs, and pay-as-you-go cost scales with it, so a documented optimisation pass is cheaper than a blind step up. Size the target with our Microsoft Fabric capacity estimator, and compare options in the Fabric capacity overage guide.
Executive checklist: review the top five CU-consuming items before approving any SKU change; make incremental refresh a standard for large models; require a model review for items that dominate interactive use; set publishing guardrails; and treat migration or resizing as the last step, after optimisation is documented. Our data maturity scorecard can show where governance gaps sit.
Power BI Capacity Throttling FAQ
Does 100% utilization mean my Power BI capacity is throttling?
No. Even above 100% utilization, Fabric does not throttle immediately. Overage protection lets a capacity use the next 10 minutes of capacity without throttling, and the Throttling chart in the Capacity Metrics app, not the Utilization chart, shows whether a limit was crossed.
Why do my reports fail with a capacity limit error?
Interactive rejection starts when smoothed usage exceeds the next 60 minutes of capacity, and Fabric then rejects new interactive requests. Users see messages such as the organization's compute capacity has exceeded its limits, or that a model cannot be loaded because of capacity limits.
Do I need to stagger refresh schedules on Fabric?
It matters less than on a fixed-capacity model. Fabric smooths background operations such as refreshes over 24 hours, so simultaneous jobs are less likely to spike on their own. Total daily consumption still counts, so heavy refreshes still need to be efficient.
Is it slow because of throttling or because of the model?
Check before you resize. Microsoft notes that slow performance is often due to the design of an item and only sometimes due to throttling. Confirm throttling in the Capacity Metrics app, and tune the heaviest items first.
How do I stop throttling quickly?
On F SKUs you can temporarily scale the SKU, pause and resume the capacity, or enable capacity overage, which Microsoft says costs three times the normal capacity rate. Pausing can make content unavailable, so check usage first. On P SKUs, the documented option is Autoscale.
Are Power BI Premium capacities going away?
Microsoft is retiring Power BI Premium per-capacity P SKUs at the end of each agreement term and no longer sells new ones. Per-user Pro and Premium Per User licences are not affected. Fabric F SKUs are the replacement, and F64 and larger give free-licence viewers access to Power BI content.
If you want a second pair of eyes on your capacity, see our Power BI consulting service or request a free capacity review.