Power BI Premium Deployment - What We Tell Clients Before They Buy Capacity
Most Australian organisations we talk to arrive at the capacity conversation the same way. They started with a handful of Pro licences in the finance team, Power BI spread across the business, and one day somebody in procurement notices the licence bill has crept past the cost of a dedicated capacity. Then the question becomes "should we buy Premium?", and usually it gets answered on licence maths alone.
That is the wrong way to decide, or at least an incomplete one. Buying capacity changes how your reports perform, who can see them, where your data lives and who gets the phone call when everything slows down at 9am on the first business day of the month. Microsoft's Power BI Premium deployment whitepaper covers the mechanics. This post is the version I would give you across the desk, based on what we have seen go well and badly with clients.
First, a note on naming
The guidance still says "Premium", but the world has moved. Microsoft has been retiring the old P SKUs in favour of Fabric F SKUs, and most new purchases we see now are Fabric capacities. An F64 is roughly the equivalent of a P1, and it is the threshold where report viewers without a Pro licence can consume content. Below F64 you get the capacity compute, but every viewer still needs a Pro or Premium Per User licence.
That F64 line matters more than anything else in the buying decision. I have watched organisations buy an F32 because it looked like a sensible starting point, then discover their 400 frontline viewers all still need licences. Do the viewer maths before you do anything else.
The deployment principles in the whitepaper still hold regardless of the SKU name. Capacity is a pool of compute that you assign workspaces to. Everything in those workspaces shares that pool. Plan around that fact and most of the other decisions follow.
Sizing - start with evidence, not guesses
The single most common mistake is sizing capacity on headcount. Headcount tells you very little. What actually consumes capacity is:
- Dataset refreshes, especially large import models refreshing during business hours
- Heavy DAX queries from poorly designed models
- Paginated reports that export thousands of pages
- Dataflows doing transformation work that should probably live elsewhere
We had a client in professional services with 1,200 staff and an estimate from their previous provider that they needed a P2. After two weeks of looking at their actual usage, the real problem was four import models refreshing every hour during the day, each one pulling full tables from an on-premises SQL Server through a gateway. We moved those to incremental refresh, shifted the schedule, and an F64 handled the load with headroom.
If you are already on Pro, look at your refresh history and query patterns before you buy. If you are on a trial or a smaller capacity, install the Fabric Capacity Metrics app on day one and leave it running. It is not the friendliest tool Microsoft has ever shipped, and the smoothing and throttling concepts take a while to get your head around, but it is the only honest view of what is happening.
A word on throttling. Fabric capacities smooth background operations like refreshes over 24 hours, and interactive operations over shorter windows. You can burst above your capacity for a while, and then you pay it back. When you overspend for long enough, interactive queries get delayed and eventually rejected. Users experience this as "Power BI is slow today", and nobody connects it to the 40GB model somebody published last Tuesday. That is why monitoring is not optional.
Workspace layout decides your future pain
The whitepaper talks about assigning workspaces to capacity, and it makes it sound like an administrative step. In practice your workspace layout determines how easily you can isolate problems later.
What we recommend for most mid-sized organisations:
Separate capacities for production and development where budget allows. Developers testing a model with a cross join will happily consume half a capacity. If that capacity also serves your executive dashboards, you will hear about it. For smaller budgets, a single F64 for production plus a small F SKU (or Pro workspaces) for development is a reasonable compromise.
Keep shared semantic models in their own workspaces. A certified sales model used by 30 reports should not live in the same workspace as somebody's experiment. This makes deployment pipelines cleaner and makes it obvious which items are important.
Don't put every workspace on capacity. Small team workspaces with five users who all have Pro licences do not need capacity. Putting them there just adds noise to your metrics and risk to your shared pool.
Use deployment pipelines or Git integration from the start. Retrofitting dev, test and prod stages onto 60 workspaces is a miserable job. We have done it. It takes weeks, and you discover all the reports that point at the wrong dataset along the way.
Data residency for Australian organisations
Your capacity region is chosen when you create it, and it determines where the data in assigned workspaces is stored. For most Australian clients that means Australia East or Australia Southeast. Your tenant home region may be different from your capacity region, and some metadata stays in the home region, so if you are in government, health or financial services, check what your compliance team actually needs before you assume a capacity in Sydney solves everything.
We have had clients with a tenant created years ago that landed somewhere unexpected. Multi-geo capacities can help, but it is a conversation to have early, not after a security review flags it.
Governance - who is allowed to break things
Capacity admins have real power. They can assign workspaces, change settings and, in the worst case, take down reporting for the whole organisation. Keep that group small, and keep it to people who understand what they are looking at.
A few settings we look at in every deployment:
- Workload settings like maximum memory per query and query timeouts. The defaults are generous. Tightening them stops one bad report from starving everyone else.
- Autoscale or surge protection. These features exist to stop you falling over during spikes. Surge protection in particular is worth turning on for capacities that run Fabric workloads alongside Power BI, because a runaway Spark job can otherwise hurt your report users.
- Who can create Fabric items. If you buy an F SKU, every workspace on it can suddenly run lakehouses, notebooks and pipelines. That is great if you planned for it. If you didn't, your Power BI capacity becomes a data engineering playground overnight. Use tenant settings to limit item creation to the groups that should have it.
What is still rough
I'll be honest about the parts that frustrate us.
The Capacity Metrics app is powerful but hard to read. Explaining CU seconds and smoothing to a finance manager who just wants to know if they bought the right size is a challenge every time.
Pricing comparisons are messy now that Power BI capacity and Fabric compute share the same SKU. A capacity that was sized for reports gets slower when the data team starts running notebooks on it. You need an agreement inside the organisation about who pays for what, or you will have arguments.
Paginated reports remain a heavy workload. If you are migrating a large SSRS estate onto capacity, test the worst reports first. Some of those 200-page monthly exports will cost more compute than you expect.
A sensible rollout order
When we run these projects, the sequence looks something like this:
- Audit current usage: users, viewers, refreshes, model sizes, gateway load.
- Decide on SKU and region based on viewer counts and the audit, not headcount.
- Design workspace layout and naming before assigning anything.
- Set tenant and capacity settings, including who can create Fabric items.
- Move a pilot group of workspaces, watch the metrics for two weeks.
- Fix the models that misbehave (there are always a few).
- Move the rest in waves, with communication to report owners.
Step six is where most of the value is. A capacity does not fix a badly designed model. It just gives the bad model more room to be bad.
Where we fit
We help organisations plan and run these migrations, from licence analysis through to model tuning and governance. If you are weighing up capacity, our Power BI consultants can run a usage audit before you sign anything. If you are thinking further ahead to lakehouses and pipelines on the same capacity, our Microsoft Fabric consulting team works on exactly that transition. And if the end goal is better decision-making rather than just faster dashboards, have a look at what we do with AI for business intelligence.
My advice in one line: measure first, size second, and design your workspaces like you'll be living with them for five years, because you probably will.