Microsoft Fabric is everywhere. Every Microsoft partner talks about it. Microsoft calls it “an analytics platform that supports end‑to‑end data workflows”, from loading data to the reports.
I prefer on-premises myself. My data warehouses run on SQL Server, with SSIS, SSAS and SSRS. So the question I ask myself is: what do you really gain if you trade that for Fabric? Below is the comparison, without the sales pitch, with the costs included.
What Fabric is, in one paragraph
Fabric is software as a service. You install nothing. You buy a capacity, from F2 to F8192, and all components run on it: Data Factory, Data Engineering, Data Warehouse, Data Science, Real-Time Intelligence, databases and Power BI. All data sits in OneLake. Every tenant has exactly one, built on Azure Data Lake Storage, and the tables are stored in it as Delta Parquet. With shortcuts you point to data elsewhere without copying it.
What is left of your SQL Server stack
| On-premises | In Fabric |
|---|---|
| SQL Server database | Fabric Warehouse, or SQL database in Fabric |
| SSIS packages | Pipelines, Dataflow Gen2, copy job |
| SQL Server Agent | Schedules on pipelines, or an Apache Airflow job |
| SSAS Tabular | Semantic model, with Direct Lake |
| SSRS | Paginated reports in the Power BI service |
| Source on your own network | On-premises data gateway, or mirroring |
| Master Data Services | Nothing from Microsoft |
On paper almost everything has a successor. In practice the difference is in the details.
The T-SQL is not the same T-SQL
The Fabric Warehouse speaks T-SQL, but not all of it. According to Microsoft’s documentation:
- Triggers are not supported.
ALTER TABLEcan only add and drop columns, and add constraints withNOT ENFORCED. So a foreign key checks nothing.ALTER COLUMNis in preview.- You cannot create a table with
datetime,smalldatetime,nvarchar,nchar,money,tinyint,xmlorgeography. You usedatetime2andvarchar. MERGE, identity columns, temp tables andTRUNCATE TABLEdo work.
Count how many nvarchar and datetime columns are in your own data warehouse. Microsoft has a Migration Assistant that moves tables, views and stored procedures, but it admits as much itself: “Currently, there isn’t full T-SQL compatibility.”
A lakehouse also has a SQL endpoint, but it is read-only. Insert, update and delete happen there with Spark.
If you really want SQL Server back, there is SQL database in Fabric. It runs on the same engine as Azure SQL Database and puts the data in OneLake in near real time. Microsoft called it generally available in November 2025. It does have limits: at most 32 vCores and 4 TB, and no change data capture.
Are Fabric pipelines better than SSIS?
I built SSIS packages for years. Are pipelines better? On a few points, yes.
- You can put pipelines, Dataflow Gen2 and copy jobs in Git (Azure DevOps or GitHub) and move them from test to production with deployment pipelines. An SSIS project can go in Git too, but there you arrange the deployment yourself.
- A copy job loads incrementally or with CDC, without you building the logic.
- You no longer have a server to patch.
On other points, no.
- If your sources are on-premises, you need an on-premises data gateway. That is another Windows server on your own network.
- Every run costs capacity. Pipelines count on the Data Orchestration and Data Movement meters, Dataflow Gen2 on Standard Compute. An SSIS package that runs every night costs nothing extra on-premises.
- Your existing packages only run in Fabric through the Invoke SSIS Package activity, and that is in preview. The packages must sit in OneLake, on-premises sources and destinations do not work yet, and custom or third-party components are not supported. An SSIS Integration Runtime as in Azure Data Factory does not exist in Fabric.
- For Script Tasks I found no successor at Microsoft. For custom code, Microsoft points to Spark notebooks. Then you write Python or Spark SQL instead of C#.
- The monitoring screen for pipeline runs, the Job runs page in the Monitor hub, is still in preview.
My conclusion: if your data warehouse loads every night from sources on your own network, SSIS is not worse. If your sources are already in the cloud, pipelines are more convenient.
Power BI or an HTML report?
In Fabric, Power BI is the report layer. Anyone who creates reports needs a Pro or Premium Per User licence. Anyone who views them does too, unless your capacity is F64 or larger. Only from F64 can someone with a free licence and the viewer role view the reports.
On Reddit I read more and more often that teams trade Power BI for plain HTML reports. I understand why. You host an HTML page that gets its data from a T-SQL query on your own web server, and it costs no licence. Two examples on this site are made exactly that way: the Netherlands by construction year on a map and branch service areas on a postcode map.
You pay the price in time. You build every report yourself, and you handle access yourself. Power BI gives you row-level security and self-service for users who want to click around. For a fixed dashboard that shows the same thing every morning, you do not need that.
What does it cost?
You pay for Fabric per capacity unit (CU) per hour. In West Europe that is €0.1975 per CU per hour, according to Microsoft’s price list on 17 November 2026. For a one-year reservation, the same price list shows €1,028.77 per CU per year. That is about 41% less, and Microsoft gives that percentage on the pricing page too. I calculate with 730 hours per month:
| Capacity | Per month, pay as you go | Per month, with a one-year reservation |
|---|---|---|
| F2 | € 288 | € 171 |
| F8 | € 1,153 | € 686 |
| F64 | € 9,227 | € 5,487 |
On top of that:
- Storage in OneLake: €0.0215 per GB per month.
- Below F64, a Power BI licence per viewer. Microsoft lists $14 per user per month for Pro and $24 for Premium Per User.
- You can pause a capacity. Then the bill stops, but nothing runs either.
On-premises you pay the SQL Server 2025 licence. Microsoft’s list prices, in dollars:
| Licence | Price |
|---|---|
| Standard, per 2 cores | $3,945 one-time |
| Enterprise, per 2 cores | $15,123 one-time |
| Standard, subscription per 2 cores | $1,418 per year |
| Enterprise, subscription per 2 cores | $5,434 per year |
Hardware, Windows and your own admin work come on top. You know those amounts better than I do. A CU is also not a core, so you cannot compare the two tables one to one.
What does stand out: an F64 with a reservation costs about €65,800 per year. Only at a few hundred report viewers is that cheaper than giving everyone a Pro licence. A smaller capacity is affordable, but then you pay per viewer. And the bill runs every month, also when you build nothing new.
Last week’s outage
On Thursday 12 November, around 19:00 Dutch time, Microsoft 365 went down worldwide. Exchange Online, Teams, Copilot, Intune and Defender did not work, or only partly, for hours. Microsoft named a recent change in the network configuration that coincided with the start of the outage. Only on Saturday morning 14 November was everything back to normal.
Fabric and Power BI were not in the reports. So I will not pretend your reports were down that day. But the outage does show how it works: when things go wrong at Microsoft, all you can do is wait. You cannot restart anything and you cannot investigate anything.
Your own server goes down sometimes too. Then you can look for the cause yourself, and decide yourself what to do about it.
A middle way: mirroring
You do not have to choose between all or nothing. With mirroring, Fabric copies the tables of your own SQL Server to OneLake in near real time. That works for SQL Server 2016 up to and including 2025. Your source stays on-premises, and in Fabric you can report on it.
- For every CU you buy, you get 1 TB of mirroring storage for free. So an F64 gives 64 TB.
- The copying itself costs no capacity. Querying the data does.
- Behind a firewall you need a data gateway.
- SQL Server 2025 must be connected to Azure Arc.
- If the capacity is paused, the replication stops.
Master data: Fabric has nothing for it
Microsoft removed Master Data Services from SQL Server 2025 and names no successor. The list of Fabric components has no master data management either. For MDM, Microsoft Purview points to partners: CluedIn, Profisee, Reltio and Semarchy.
So if you run MDS, you cannot move to Fabric in one step. You have three ways:
- You keep a SQL Server 2022 with MDS. Microsoft still supports MDS there, and SQL Server 2022 has extended support until 11 January 2033.
- You put your master data in an MDM tool on your own SQL Server, such as Primentra. If that database runs on Standard or Enterprise, you bring it into Fabric with mirroring.
- You take an MDM package that runs inside Fabric itself. Profisee offers one and has called it generally available since March 2026. That is an enterprise package.
More about the options after MDS is in my article on MDS end of life.
My conclusion
Fabric is not a bad product. For an organisation whose sources are already in Azure or Dynamics, and that wants to show reports to hundreds of people, it fits well.
For a data warehouse that loads every night from its own systems, with SSIS and a fixed group of users, I see little gain. You trade a fixed licence for a monthly bill. You trade T-SQL you know for a T-SQL with gaps. Your master data has to stay somewhere else. And when things go wrong, you wait for Microsoft.
That is why I stay on-premises. If you want to try Fabric, Microsoft gives you a 60-day trial, at no cost.
I wrote earlier about lakehouse, warehouse and datamart in general in this article.
Sources
- Microsoft Fabric overview
- OneLake overview
- Fabric licenses
- Pause and resume a capacity
- Fabric trial
- T-SQL surface area in Fabric Data Warehouse
- Data types in Fabric Data Warehouse
- Lakehouse or warehouse: decision guide
- Fabric Migration Assistant
- SQL database in Fabric and its limits
- Microsoft databases and Microsoft Fabric (Azure blog, November 2025)
- Discontinued database engine functionality (MDS and DQS)
- Master data management in Microsoft Purview
- Profisee on MDM in Microsoft Fabric (March 2026)
- Invoke SSIS Package activity
- Data Factory: upgrade best practices
- Data Factory: on-premises data
- Data Factory: pricing
- Git integration
- Monitor hub: job runs
- Mirroring overview and mirroring for SQL Server
- Azure price list for Microsoft Fabric (Retail Prices API, West Europe, EUR)
- Power BI pricing
- SQL Server 2025 pricing (pdf)
- The Register on the outage of 12 November
- SecurityWeek on the outage of 12 November
- Wikipedia: 2026 Microsoft outage
Frequently asked questions
What is the difference between SQL Server on-premises and Microsoft Fabric?
SQL Server runs on your own server and you pay licences per core. Microsoft Fabric is a SaaS platform from Microsoft: you rent capacity per hour, and all data sits in OneLake in Delta Parquet. You no longer manage a server, but you also have less under your own control.
Can I run my SSIS packages in Microsoft Fabric?
Only with the Invoke SSIS Package activity, and that is still in preview. The packages must sit in OneLake, on-premises sources and destinations do not work yet, and custom or third-party components are not supported. An SSIS Integration Runtime as in Azure Data Factory does not exist in Fabric.
Does my T-SQL work in a Fabric Warehouse?
Not all of it. Triggers are not supported, ALTER TABLE can do only a few things, and constraints only work as NOT ENFORCED. Data types such as datetime, nvarchar, money, tinyint and xml cannot be used in a table. MERGE, identity columns and temp tables do work.
Do viewers of Power BI reports in Fabric need a Pro licence?
On a capacity smaller than F64 they do: every viewer then needs Pro, Premium Per User or a trial. From F64, someone with a free licence and the viewer role can see the reports. Anyone who creates reports always needs Pro or Premium Per User.
Can I connect SQL Server on-premises to Fabric without migrating?
Yes, with mirroring. Fabric copies the tables of SQL Server 2016 up to and including 2025 to OneLake in near real time. Behind a firewall you need a data gateway, and SQL Server 2025 must be connected to Azure Arc. The capacity must be running, or the replication stops.
Comments
Related blogs
Upgrading to SQL Server 2025: what you gain and what it costs
Microsoft MDS end of life: your options before 2028
Faster AS400 loads into SQL Server: one column per row