Subscribe
Share
Share
Embed
Copilot in Fabric has no per-user license fee. But every request draws from the same capacity pool as your pipelines and reports. We dig into what 'included' actually costs — and when it throttles production.
Episode 24 • 2026-06-12 Duration: 9:11
Copilot in Fabric has no per-user license fee. But every request draws from the same capacity pool as your pipelines and reports. We dig into what 'included' actually costs — and when it throttles production.
AI-generated voices. Matthias — cloned voice. Fabia — designed AI co-host. See Matthias live on YouTube (Fabric Friday), at his meetups, and at conferences like FabCon.
Hosted by Matthias Falland — Microsoft Data Platform MVP and community architect behind the Fabric Periodic Table. New episodes every Friday.
Have an architecture decision you are wrestling with? DM Matthias on LinkedIn — find him as Matthias Falland. Three to five sentences about the decision, your team size, and your current stack. We anonymize before airing.
This podcast was generated by AI.
Brand design based on fabricperiodictable.com.
Architecture decisions for Microsoft Fabric. Anonymized real customer scenarios, cost realism, counter-arguments included. Weekly episodes aligned with Fabric Friday recordings.
Speaker 1
No additional license required. You'll hear that in every copilot in fabric conversation.
Speaker 2
And it's accurate. No $30 a month like M365 Copilot. No separate subscription. It's included in your fabric capacity.
Speaker 1
So why are teams finding out the hard way that it's expensive? Architecture decisions for Microsoft Fabric. This is the Fabric Architecture Podcast.
Speaker 2
That's what we're digging into. What included actually means when every copilot request draws from the same capacity pool your pipelines and reports depend on. Copilot in Fabric isn't one product. It's a collection of workload-specific assistance. There's a different one in your notebook, then in your warehouse, then in your Power BI report. They all call Azure OpenAI through a shared pipeline, but the grounding, the context added to every prompt, is different depending on where you are.
Speaker 1
Different how?
Speaker 2
In a warehouse, Copilot gets table names, column names, primary and foreign keys. Structural metadata. The actual data values in your tables don't travel to Azure OpenAI In a notebook, it's broader. Workspace context, attached lakehouse schema, code structure, even runtime state. In Power BI, it's your semantic model metadata, table names, measure definitions, descriptions.
Speaker 1
So the quality of what Copilot gives you is completely capped by how well you've documented your model.
Speaker 2
Microsoft flags this explicitly. A DAX measure that works in the query view can produce incorrect results in a different filter context in your report. Copilot doesn't know where you'll use that measure. That's a grounding limitation you handed it.
Speaker 1
Can I tell you what the actual number one copilot problem is though? It's not output quality. It's copilot is grayed out. I keep running into it. Someone enables fabric, expects copilot to be there and the buttons dead. Every single time it's governance. Tenant hoggles off, capacity's too small, or the workspace is sitting on the wrong skew.
Speaker 2
The diagnostic is straightforward. Tenant Setting First. There's a master toggle in the admin portal under Copilot and Azure OpenAI Service. Then SKU. You need paid F2 or higher. Trial capacities are genuinely ambiguous. Microsoft's own docs contradict each other on whether trial supports Copilot. Most pages say no. The notebook known limitations table says yes. Don't rely on trial.
Speaker 1
Classic.
Speaker 2
Then region. If your capacity's outside the US or EU, Copilot's disabled by default. You need the Cross Geo processing toggle, and that one's tenant admin only. It doesn't delegate to capacity admins. By design, it's a data sovereignty decision, not an IT checkbox.
Speaker 1
And most teams I've talked to don't realize that's what they're signing when they flip.
Speaker 2
A listener wrote in with this. Our pipeline started queuing and reports got slow. We traced it to copilot usage on our F2. How can a free AI feature throttle our whole capacity?
Speaker 1
Because nobody reads past no additional license. They hear free and stop.
Speaker 2
Every copilot request consumes capacity units. The same CUs that run your pipelines, your queries, your report refreshes. The rates are published. Input tokens, your prompt plus grounding data, cost 100 CU seconds per thousand tokens. Output tokens cost 400 CU seconds per thousand. Output is where the bill grows because generated code and report pages are token heavy.
Speaker 1
400 versus 100 on output?
Speaker 2
Microsoft phrases it as output being three times more expensive. The absolute rate ratio is four to one. Same numbers, two ways of describing them. Either way, a typical request, say 2,000 input tokens and 500 output, works out to about 400 cu seconds. Call it 6. 5 CU minutes per request.
Speaker 1
6. 5 Cu minutes for one question. How many does an F2 get in a day?
Speaker 2
Copilots classified as a background job, smoothed over 24 hours. An F64, 64 CUs times 24 hours, gives you roughly 1500 CU hours a day. Microsoft's figure is about 13,800 requests before exhaustion on F64. Scale that to F2. Two CUs, 48 CU hours, and you're at roughly 400 requests a day.
Speaker 1
A team of 10, each asking 50 questions, burns through that by lunch.
Speaker 2
And when you exhaust the budget, you hit carry forward. See you debt borrowed against the next 24-hour window. That's when interactive workloads start queuing. Your report viewers don't know anyone's been using Copilot. They just know Power BI got slow on a Tuesday afternoon.
Speaker 1
If you're on anything below an F64 and you've got more than a handful of people using copilot, I'd designate a separate fabric copilot capacity Route all copilot CU to a dedicated F2. Yes, it's a second capacity to manage. But the cost of that extra F2 is trivial compared to your pipelines missing their SLA window because someone asked Copilot to refactor a notebook. That's the trade-off I take every time.
Speaker 2
Let's steel man the other side though. If you're a small team, three or four people, running on an F-64, shared capacity is genuinely fine. 13,000 requests a day is more than you'll generate, and adding a second capacity means a second thing to monitor, a second billing line, management overhead that doesn't pay for itself at that scale.
Speaker 1
Fair. And if Copilot's only running in deaf workspaces, the stakes change entirely. A throttled development environment is annoying. A throttle production capacity is an incident.
Speaker 2
Which makes where you run Copilot the real architecture decision, more than whether you run it at all.
Speaker 1
What about the output quality side? Even with capacity sorted out, how bad does it get with a messy schema?
Speaker 2
Microsoft's DAC stocks publish hard limits. Copilot may use a reduced model if your semantic model exceeds 500 tables or a thousand columns. Descriptions get truncated at 200 characters. Below those ceilings, quality still scales directly with how well you've named and documented things In a team of eight under cost pressure, nobody's writing column descriptions. And then they wonder why Copilot generates garbage.
Speaker 1
The risk I'd actually worry about isn't cost or quality. It's the privacy nuance that gets skipped in rollout checklists. Microsoft's core commitment is solid Azure OpenAI boundary, no training on your data, scoped to the current user's permissions. But notebook copilot and data agents store conversation history for 28 days unless you manually clear the chat. For a regulated shop running notebooks over patient records or financial data, that's the detail your DPO needs to hear. And most enablement guides don't mention it.
Speaker 2
Pattern dictates platform, and with Copilot, the pattern is this. It's a productivity tool that's already wired into your cost structure, whether you planned for it or not. The teams that get value from it are the ones who treat enablement as a governance project, monitor capacity impact from day one, and fix their schema before they blame the AI.
Speaker 1
The most expensive support ticket you'll ever file is the one where the answer is rename your columns.
Speaker 2
Somewhere right now there's an F2 running copilot for a whole team. The pipelines are queuing, and nobody's checked the capacity metrics app. Because the AI assistant they didn't budget for doesn't show up on any invoice.
Speaker 1
If you've got a copilot governance story or a capacity question, DM me on LinkedIn. Find me as Matthias Falland.
Speaker 2
See you next Friday.
Speaker 1
This was the Fabric Architecture Podcast. This podcast was generated by AI. New episode every Friday.