Fabric F2048 capacity
F2048 is for large, consolidated estates where many heavy models share one capacity. Note that the per-model memory limit stops growing at 400 GB from F1024 upward — beyond that, scale comes from parallelism, not bigger single models.
- Pay-as-you-go
- $269,107/mo
- 1-yr reserved
- $160,085/mo
- Memory per model
- 400 GB
- Free viewers
- Yes
F2048 limits at a glance
| Limit | Value |
|---|---|
| Capacity units (CU) | 2048 |
| Equivalent Power BI SKU | None |
| Max memory per semantic model | 400 GB |
| Max memory per query | 40 GB |
| Model refresh parallelism | 1280 |
| Max concurrent DirectQuery connections (per model) | 200 |
| Direct Lake rows per table | 24 billion |
| Direct Lake Parquet files per table | 10,000 |
| Direct Lake max model size on OneLake | Unlimited |
| Viewers with free licence | Allowed (Viewer role) |
| Copilot | Supported (not on trial capacity) |
What F2048 costs
Fabric bills capacity units by the hour. At US East list prices, F2048 costs $368.64 per hour, or about $269,107.20 for a full month. A 1-year reservation brings that to about $160,085.33 per month — roughly 40.5% less — in exchange for committing to the size for a year.
Because free-licence viewers are allowed on F2048, the comparison that matters is against per-user licensing. A reserved F2048 costs the same as about 11,435 Pro viewers (19,222 at pay-as-you-go). Below that, Pro licences are usually cheaper unless you also need Fabric workloads or larger models.
Prices are Microsoft list prices for US East, verified 2026-10-10, and vary by region and agreement.
Is F2048 the right size?
Good fit for
- Organisation-wide distribution to viewers on free licences
- Import models up to roughly 200 GB each, or larger Direct Lake models
- Direct Lake tables up to 24 billion rows with no model-size cap on OneLake
Watch out for
- Throttling from smoothed background usage — watch the Capacity Metrics app before scaling up
- Paying for headroom that model optimisation could remove
- Per-model memory plateaus at 400 GB from F1024 upward
The usual sizing mistake is choosing a SKU from the average model when the largest model and the morning refresh window decide whether a capacity throttles. The free Fabric capacity estimator checks both.
Compare neighbouring sizes
Fabric F2048: common questions
- How much does Fabric F2048 cost per month?
- At US East list prices, F2048 costs about $269,107.20 per month pay-as-you-go ($368.64 per hour, 2048 CUs × $0.18), or about $160,085.33 per month with a 1-year reservation — roughly 40.5% less. Prices vary by region and pay-as-you-go capacity can be paused when idle.
- Do report viewers need a Power BI Pro licence on F2048?
- No. On F64 and larger, users with a free Fabric licence and the Viewer role can consume Power BI content. Report authors and publishers still need Pro or PPU. At $14 per Pro licence, F2048 costs the same as about 11,435 Pro viewers (reserved) or 19,222 (pay-as-you-go).
- How large can a semantic model be on F2048?
- Each semantic model can use up to 400 GB of memory, including refresh overhead. Because a full Import refresh keeps the old copy available while the new one loads, a practical rule of thumb is to keep Import models below about 200 GB, or use incremental refresh. Each query is limited to 40 GB.
- What are the Direct Lake guardrails on F2048?
- Up to 24 billion rows and 10,000 Parquet files (and row groups) per table, with no model-size cap on OneLake. If a table exceeds them, Direct Lake on SQL endpoints falls back to DirectQuery (when fallback is enabled), while Direct Lake on OneLake fails to refresh until the table is optimised back under the limit.
- Can I use Copilot on F2048?
- Yes. Copilot in Fabric and Power BI requires an F2 or larger capacity (or a P SKU) in a supported region; it is not available on trial capacities. Copilot usage consumes capacity units, so it competes with refreshes and queries on the same capacity.
Sources: Microsoft Learn — capacity limits, Direct Lake guardrails, Power BI pricing, Premium to Fabric migration, Copilot in Fabric. Verified 2026-10-10.