The requirement: a capacity, not a license
Start with the sentence that resets most budget conversations, verbatim from Microsoft Learn: "Your organization needs a paid Fabric capacity (F2 or higher) or Power BI Premium capacity (P1 or higher)... A Power BI Pro or Premium Per User (PPU) license alone isn't sufficient - Copilot requires organizational capacity." Trial SKUs are not supported either.
Two dated facts frame the current state. In April 2025, Microsoft removed the old F64-only gate - Copilot opened to all paid SKUs from F2 up, which made the headline entry price 30× cheaper and is why older "$5K/month minimum" articles are stale. And in April 2025, per-user license prices also rose: Pro to $14/user/month, PPU to $24. The entry door got cheaper; the seats and the meters did not.
The SKU ladder
From Azure's published US pricing (12 June 2026):
| Capacity | Pay-as-you-go | Reserved (1-yr) | What it means for Copilot |
|---|---|---|---|
| F2 | $262.80/mo | $156.33/mo | The legal minimum - and, per community measurement, exhausted after ~20 questions/day |
| F32 | $4,204.80/mo | ~$2,501/mo | Mid-rung; viewers still need Pro seats |
| F64 | $8,409.60/mo | $5,002.67/mo | The practical threshold: viewer Pro licenses no longer required; Microsoft's own consumption examples are sized here |
| F128 | $16,819.20/mo | ~$10,005/mo | What practitioners recommend for real concurrency ("Even a F64 can be brought to its knees by a handful of concurrent Copilot users") |
The license interaction hiding in that table is the F64 threshold, from Azure's own pricing page: "with an F64 - a Premium P1 equivalent - or larger capacities, Power BI report consumers do not require a Power BI Pro license." Below F64, every report consumer needs Pro at $14/user/month - so for viewer-heavy organizations, the per-seat line can rival the capacity line, and crossing to F64 partially pays for itself in license relief.
The second meter: tokens against your capacity
Owning capacity is admission; using Copilot then draws it down per token: 100 capacity-unit seconds per 1,000 input tokens, 400 per 1,000 output tokens. Microsoft's worked example prices a modest request (2,000 input + 500 output tokens) at 400 CU-seconds and concludes an F64 supports "over 13,824 Copilot requests per day." The doc also states the failure mode without euphemism: "Once the capacity is exhausted, all operations will shut down" - not just Copilot; the capacity's workloads generally.
Now the field data, from Microsoft's own community forums. One user measured ~10,000 CU-seconds per question - 25× the doc example - because real questions carry grounding context (schema, report metadata) and multi-call chains; his F2 died "after roughly 20 questions" and paused for 24 hours. A practitioner reply elsewhere: "In my experience you need at least a F128 for a meaningful Copilot experience." Microsoft's consumption doc itself warns: "Deploying broadly without validating consumption patterns can quickly exhaust your capacity," and notes consumption rates "are subject to change at any time." Plan with the field numbers, not the example.
One more mechanism worth knowing: Fabric Copilot capacity lets Copilot usage by Pro/PPU users in non-capacity workspaces bill to one designated F2+/P1+ capacity. It centralizes the bill; it does not remove it.
Worked examples (all figures Microsoft's)
| Organization | Configuration | Monthly | ~Annual | The catch |
|---|---|---|---|---|
| Pilot team (10 users) | F2 PAYG + 10 Pro seats | $262.80 + $140 = $402.80 | ~$4,800 | ~20 questions/day before the capacity pauses, per field measurement |
| Mid-size (100 users) | F2 PAYG + 100 Pro seats | $262.80 + $1,400 = $1,662.80 | ~$20,000 | Same ~20-question ceiling shared by 100 people - functionally a demo |
| Real deployment (100+ users) | F64 PAYG, viewers license-free | $8,409.60 | ~$100,900 | Sized for sustained use; concurrency may still push toward F128 (~$202K PAYG) |
And the line no table captures: the preparation labour. Microsoft's accuracy guidance - star-schema remodeling, naming, descriptions, linguistic modeling, AI instructions, verified answers, iterative testing - is the price of usable output, documented in Microsoft's own words ("If you don't prepare these elements, Copilot mainly produces low-quality and inaccurate outputs"). We covered it fully in Why Power BI Copilot Gives Confidently Wrong Answers; budget it as a project, not a setting.
Five cost-control levers if you proceed
- Reserve if you're committed: 1-year reservations cut capacity cost ~41% (F64: $8,409.60 → $5,002.67/month).
- Measure before you roll out: run the Capacity Metrics app against a pilot's real questions - your CU-per-question number, not the 400 CU-s doc example, is your planning constant.
- Mind the 24-hour cache: identical prompts within 24 hours answer from cache - good for cost, but it masks consumption (and accuracy) variance during testing.
- Model the F64 crossover: at $14/seat, ~600 Pro-licensed viewers cost what the F2→F64 jump does; viewer-heavy orgs may find F64 cheaper than staying small.
- Cap the blast radius: Copilot shares the capacity with your other Fabric workloads - "all operations will shut down" means an enthusiastic Copilot rollout can pause your pipelines. Isolate or surge-protect accordingly.
Where Colrows changes the math (our product)
Disclosure: Colrows competes for the question-answering workload (the head-to-head is at Colrows vs Power BI Copilot). The pricing-relevant difference is the cost shape: no per-seat BI license multiplying against headcount, no capacity that pauses mid-quarter, and no preparation project - the semantic graph builds and maintains itself. Compute is metered, but against deterministic compiled queries rather than token-hungry generation loops, and the free tier (unlimited datasources, users, and access policies) makes the pilot phase $0. If the workload you are pricing is "trustworthy answers for many users," compare the architectures before comparing the SKUs.
Frequently asked questions
How much does Power BI Copilot cost?
Capacity from $262.80/month (F2 PAYG) plus Pro seats ($14/user/month) below F64, plus token-metered consumption. Realistic sustained use prices at F64 ($8,409.60/month PAYG; $5,002.67 reserved) or above - roughly $100K/year before preparation labour.
Does Copilot work with just a Pro license?
No - Microsoft is explicit that Pro or PPU alone "isn't sufficient"; a paid F2+ or P1+ capacity is required. The April 2025 change lowered the capacity floor from F64 to F2; it did not remove the capacity requirement.
How is usage metered?
100 CU-seconds per 1,000 input tokens, 400 per 1,000 output, drawn against your capacity. Doc example: ~400 CU-s/request; field measurements: 5,000-10,000 CU-s once grounding is included. Exhausted capacity shuts down all operations on it.
What does it cost for 100 users?
About $1,663/month at the F2-plus-Pro-seats floor (with a ~20-question/day shared ceiling), or $8,409.60/month on an F64 sized for actual use - viewers license-free at that tier.
A note on the claims
All prices are Azure/Microsoft published US figures and Microsoft Learn consumption rates as of 12 June 2026; community burn rates are linked to the Fabric Community threads where users posted their measurements. Microsoft notes consumption rates may change at any time, and capacity pricing varies by region - verify against the live pages before budgeting. We are a competitor; the sources are linked so you can audit the framing. This page is reviewed quarterly.
