One platform, one copy of the data
Contents
About Microsoft
Microsoft, founded in 1975 by Bill Gates and Paul Allen, headquartered in Redmond, Washington, and led by chairman and chief executive Satya Nadella since 2014: is one of the world's most valuable companies (NASDAQ: MSFT). The Microsoft Cloud spans Azure, Microsoft 365, Dynamics 365, Microsoft Security, Microsoft Fabric and GitHub. Volge delivers across that estate through six alliance offerings.
What Volge delivers on Microsoft: six offerings
The offering this book covers: marked ◆: is Applied AI, delivered on Microsoft Fabric.
| Volge offering | What it covers | Anchored on |
|---|---|---|
| Modern Workspace & E7 | Secure modern work on the E7 bundle; human-agent collaboration | Microsoft 365 · Copilot |
| Applied AI ◆ | Building, governing and running enterprise AI agents and apps | Azure AI Foundry |
| Zero Trust Security | Identity, threat protection and data governance | Microsoft Entra · Defender · Purview |
| Business Applications | ERP, CRM and low-code, with in-process agents | Dynamics 365 · Power Platform |
| Digital Product Engineering | Autonomous software engineering and CI/CD | GitHub · GitHub Copilot |
| Cloud Infrastructure & Operations | Cloud build, run and cost governance (FinOps) | Microsoft Azure |
Sources: Microsoft corporate and product documentation, 2025–26.
This book: Applied AI on Microsoft Fabric
Microsoft Fabric is an end-to-end SaaS analytics platform that unifies data engineering, warehousing, data science, real-time intelligence, integration and BI on OneLake, one open, tenant-wide lake holding a single copy of data in Delta Parquet (GA November 2023).
In the field in 2026
- OneLake, a single tenant-wide lake in open Delta Parquet, with shortcuts and near-real-time mirroring (no ETL)
- Direct Lake semantic models in Power BI that query OneLake directly, on any Fabric capacity (F2+)
- Real-Time Intelligence: Eventstream, Eventhouse (KQL) and Activator for streaming and event-driven action
- Copilot in Fabric and Fabric Data Agents, plus SQL database in Fabric, governed by Microsoft Purview
Executive summary
The analytics estate fragmented one procurement decision at a time. A warehouse was bought for structured reporting, a lake for the data science team, a streaming stack for operational telemetry, a separate business-intelligence tier for the dashboards, a governance tool to catalogue what the others held. Each was licensed on its own meter, operated by its own team, and connected to the rest by pipelines that someone has to keep running. The estate works. It also leaks, cost, latency, and trust, and it leaks at the seams between the services, not inside any one of them.
That leakage has a name: the integration tax. It is the engineering paid to move and reconcile the same data across stores that do not share one, the compute paid to maintain redundant copies, the latency added by every hop, and the trust lost each time two services disagree about the same customer because each holds its own copy. The tax is rarely a line item, which is exactly why it persists. It is distributed across pipeline maintenance, duplicated storage, idle reserved capacity on five separate meters, and the slow erosion of confidence in numbers that should match and do not.
Microsoft Fabric is a bet that the tax is large enough to justify collapsing the stack. Fabric is an end-to-end software-as-a-service analytics platform that unifies the workloads a data estate runs, data engineering, data warehousing, data science, real-time intelligence, data integration, business intelligence, and event-driven activation, over one shared lake, OneLake, under one governance model and one capacity model (Microsoft, 2026). It consolidates what were separately licensed Azure services, Azure Synapse Analytics, Azure Data Lake Storage, Power BI Premium, and Azure Data Factory, onto a single platform over one copy of the data (Microsoft, 2026). The warehouse and the lakehouse read the same files. Power BI queries those files in place. The streaming engine, the data-science notebooks, and the pipelines all operate on one logical store rather than on copies of it.
The adoption signal is not ambiguous. Microsoft reports roughly 35,000 paid Fabric customers, up about 60 per cent year over year, with the volume of data held in OneLake up nearly four times year over year and an annual revenue run rate above two billion United States dollars (Microsoft earnings, FY2026). Consolidation onto one platform is the fastest-growing pattern in enterprise analytics.
The central finding. The decision a data leader actually faces is not which engine is fastest. It is what collapsing the analytics stack onto one platform saves, what it costs, and where it does not fit. Fabric's value is the integration tax it removes, one copy of the data, one governance plane, one capacity meter, one operational surface, set against the specialised depth a best-of-breed estate retains. Where the estate is fragmented, Microsoft-aligned, and dominated by integration toil, consolidation pays. Where a workload demands the engineering or warehousing ceiling of a specialist platform, it does not, and the honest architecture is hybrid.
This is a paper about consolidation economics, not about artificial intelligence. AI-readiness is one outcome of a consolidated, governed estate, and a real one, but it is a downstream benefit treated briefly here. The companion Volge paper, The data estate is the accuracy budget (Topic 7), carries the agent-grounding argument in full; this paper points to it once and stays on the data leader's own drivers, estate consolidation, migration off legacy services, cost and capacity, real-time analytics, governance, and business intelligence.
Five conclusions, developed in the chapters that follow.
First, the integration tax is the lever, not engine performance. Cost, latency, and trust leak between services, not within them. The saving from consolidation comes from removing copies, pipelines, and redundant capacity, the connective tissue, rather than from any single engine running faster.
Second, OneLake is the architectural decision that makes the rest follow. One logical lake, storing data once in an open format and exposing it to every workload without duplication, is what lets the warehouse, the lakehouse, and the BI tier operate on one copy. Remove copy sprawl at the storage layer and the contradictions it manufactures disappear upstream of every consumer.
Third, the capacity model is the commercial mechanism, and F64 is its inflection. Fabric replaces per-service licensing with one pooled compute meter measured in capacity units. The F64 tier is the point at which Power BI Premium-equivalent features unlock, chiefly free report consumption for viewers, which is why it anchors most enterprise deployments and why the total-cost comparison turns on capacity sizing rather than on headline list prices (industry pricing analyses, 2026).
Fourth, the migration is a programme, not a switch. Moving off Synapse, standalone Power BI Premium, and Azure Data Factory is staged work with real cutover risk. Microsoft has not declared Synapse end-of-life wholesale, but the strategic direction and targeted deprecations are unambiguous (Microsoft, 2025–2026). The migration is underwritten by the integration tax it retires, not by novelty.
Fifth, consolidation is not the answer for every workload. Databricks remains stronger for large-scale data engineering and machine learning on Apache Spark; Snowflake remains stronger for high-concurrency SQL warehousing and frictionless data sharing. Fabric wins on unification and on integration with the Microsoft and Power BI ecosystem. A mature estate is frequently hybrid, and Fabric's zero-copy interoperability with both platforms is what makes hybrid viable rather than a contradiction.
Fabric is the primary lens throughout. Databricks and Snowflake appear where the contrast sharpens the decision, and Fabric's interoperability with both is treated as a design feature rather than a concession. Volge's position is plain: the consolidation decision is an economics decision before it is a technology one, and it does not survive being made on a feature sheet.
Chapter one: The fragmented analytics estate and the integration tax
The takeaway
The analytics estate fragmented by accumulation, and the cost of that fragmentation sits in the integration between services, not inside any one of them. Cost leaks through duplicated storage and redundant reserved capacity; latency leaks through every pipeline hop; trust leaks every time two services disagree about the same entity. Consolidation is the lever because it attacks the seams, the copies, the pipelines, and the separate meters, rather than the engines. Read the tax first, because it is the saving that underwrites everything that follows.
2.1 How the estate fragmented
No one designed the fragmented estate. It accreted. A data warehouse arrived first, sized for structured financial and operational reporting. The data science team needed cheap, schema-flexible storage the warehouse could not give them economically, so a data lake followed. Operational use cases needed sub-minute latency the batch warehouse could not serve, so a streaming stack was bought. The business needed dashboards, so a business-intelligence platform was licensed on its own terms. Someone then needed to know what all of this held and who could see it, so a cataloguing and governance tool was added on top.
Each decision was locally rational. The sum is an estate of five or six separately licensed, separately operated services, each with its own storage, its own compute, its own security model, and its own operational team. The data does not live in one place; it lives in all of them, in copies, and the copies are kept roughly aligned by pipelines that run on a schedule and break on their own one.
2.2 The integration tax, itemised
The cost of fragmentation is not the cost of the services. It is the cost of holding them together. That cost is distributed, which is why it rarely appears as a number a finance team can point at, and four components carry most of it.
Duplicated storage. The same dataset lands in the lake, is copied to the warehouse, is re-exported to the BI tier's import model, and is staged again for the streaming joins. One logical dataset becomes four physical copies, each metered, each drifting from the others between refreshes.
Pipeline toil. Every copy is a pipeline, and every pipeline is code that someone writes, monitors, debugs at the wrong hour, and rewrites when an upstream schema changes. Integration debt, brittle interfaces, undocumented transformations, inconsistent access controls, accumulates fastest here, and it is the line of the business case most often underestimated.
Redundant capacity. Each service is sized and reserved on its own meter, against its own peak. Because the peaks do not coincide, the estate carries idle reserved capacity across several services at once, paid for, provisioned, and unused, because compute cannot be pooled across products that do not share one.
Trust erosion. This is the tax that does not appear on any invoice and costs the most. When the warehouse and the BI extract disagree about a revenue figure because each holds its own copy refreshed at a different hour, the data team spends its credibility reconciling them. Numbers that should match and do not are the slow poison of a fragmented estate, and they are manufactured by copy sprawl, not by any engine being wrong.
By the numbers
- USD 5.5 trillion: estimated global losses carried by data silos (industry survey, 2024)
- 68 per cent: surveyed respondents ranking data silos as their single biggest data challenge (industry survey, 2024)
The figure is indicative, but the direction is not in dispute: the seams, not the services, are where the value leaks.
2.3 The diagram of the problem
The fragmented estate and the consolidated alternative are easiest to read side by side. The left is the accumulated estate; the right is the same workloads over one shared lake.
The left-hand estate spends its effort on the dotted lines: the copies and the catalogue reaching across products that do not share storage. The right-hand estate removes the dotted lines by removing the copies. That is the whole consolidation thesis in one picture.
By the numbers: Fabric adoption (Microsoft earnings, FY2026)
- ~35,000: paid Fabric customers, up ~60 per cent year over year
- ~4×: growth in OneLake data volume year over year
- > USD 2 billion: annual revenue run rate
2.4 Why consolidation is the lever
The reflex fix for a slow or expensive estate is to tune an engine or renegotiate a service. The reflex is misdirected, because the cost does not sit in the engines. It sits in the integration between them, and no amount of engine tuning reaches it.
Consolidation attacks the cost where it actually lives. Store the data once and the duplicated storage disappears. Read it in place from every workload and the copy pipelines disappear with it. Pool the compute under one meter and the redundant reserved capacity collapses to a single sizing decision. Govern one copy and the trust erosion stops at the source, because there is one authoritative record rather than a population of disagreeing ones. None of this requires any single engine to be the fastest in its class. It requires the seams to be removed, and removing the seams is what a single platform over one lake is for.
The trade is real and is stated honestly here. Consolidation buys operational simplicity and removes the integration tax; it costs the specialised depth that a dedicated, best-of-breed engine retains. The rest of this paper is the working-through of that trade: the architecture that delivers the consolidation in chapter two, the economics and the migration in chapter three, and the honest limit of the approach in the same chapter, where best-of-breed still wins.
Chapter two: The Fabric architecture: OneLake, the workloads, Direct Lake, governance
The takeaway
Fabric is one software-as-a-service platform: seven workloads over one shared lake, governed and metered as a unit. OneLake is the decision that makes the rest follow, one logical store, one open copy of the data, read in place by every workload. Direct Lake lets Power BI query that copy without importing or duplicating it, which closes the largest remaining seam in the classic BI architecture. The medallion pattern organises the lake into rising tiers of quality, and Purview governs the whole estate as a layer rather than a bolt-on. Understand OneLake and the rest of the platform is a consequence of it.
3.1 OneLake: the floor the platform stands on
OneLake is one unified logical data lake for the entire organisation, provisioned automatically with every Fabric tenant, which Microsoft frames as OneDrive for data (Microsoft, 2025–2026). It is the single most consequential design choice in the platform, because every other workload reads from it.
The architectural payoff is the elimination of copies. OneLake stores data once, in the open Delta Parquet format, and exposes that one copy to every Fabric engine, the warehouse, the lakehouse, the data-science runtime, the real-time engine, and Power BI, without duplication. A shortcut lets data that physically lives elsewhere, in Amazon Simple Storage Service, in Azure Data Lake Storage, or in another organisation's estate, surface inside OneLake without being copied across. The estate becomes one logical surface over data that may remain physically distributed.
This is the structural answer to the copy sprawl of chapter one. In the fragmented estate, each service held its own copy and the pipelines kept them roughly aligned. In OneLake, there is one copy, and every workload reads it in place. The pipelines that existed only to move data between services have nothing left to do, and the storage those copies consumed is reclaimed. The open format matters here for the same reason: because tables live in Delta Parquet rather than a proprietary warehouse format, every engine reads the same files without conversion, which forecloses the lock-in and the re-export proliferation that proprietary formats impose.
3.2 The workloads over one lake
Fabric unifies the workloads a complete analytics estate runs, each presented as a role-tailored experience over the same OneLake storage (Microsoft, 2026). The point is not that each workload is individually best in class; it is that they share one copy of the data, one governance plane, and one capacity meter.
The table below maps each workload to the separately licensed service it consolidates and to the estate problem it addresses.
| Fabric workload | What it does | Consolidates | Estate problem it addresses |
|---|---|---|---|
| Data Engineering | Spark notebooks and lakehouse over OneLake | Synapse Spark; lake compute | Lake and engineering on the shared copy, no separate store |
| Data Warehouse | T-SQL warehouse on open Delta tables | Synapse dedicated SQL pools | Warehouse reads the same files as the lake; no copy between them |
| Data Science | Notebooks, model training, MLflow | Synapse / Azure ML notebooks | Models train on the curated lake, not on a re-exported slice |
| Real-Time Intelligence | Streaming and event analytics via Eventhouse and KQL | Synapse Data Explorer; separate streaming stack | Closes the recency gap without a standalone meter |
| Data Factory | Pipelines, mirroring, 200+ connectors | Azure Data Factory | Ingestion and integration on-platform; mirroring replaces copy pipelines |
| Power BI | Semantic models and reports, Direct Lake | Power BI Premium | BI queries OneLake in place; no import refresh |
| Activator | Triggers and actions on data conditions | Bespoke alerting glue | Event-driven action without a custom service |
Read the consolidation column down the table. Each row retires a separately licensed, separately operated service, and the retirement is what removes the integration tax, not because any single Fabric workload outperforms the specialist it replaces, but because all seven now share one copy and one meter.
3.3 Direct Lake: closing the BI seam
The classic business-intelligence architecture carries one of the most expensive seams in the estate: the import. To make a dashboard fast, the BI tier copies data out of the warehouse into its own in-memory model and refreshes that copy on a schedule. The copy is another store to size, the refresh is another pipeline to run and to fail, and the model is stale between refreshes by construction.
Direct Lake removes the seam. It is a Power BI semantic-model storage mode, available in Fabric, that lets the engine read Delta tables in OneLake directly, loading only the columns a query needs into memory, without importing the data and without a scheduled refresh (Microsoft, 2026). It targets near-import query speed with near-real-time freshness, which is the combination the import-versus-direct-query choice never offered.
For the consolidation thesis this is the closing move. The warehouse, the lakehouse, and now the BI tier all read one copy of the data in OneLake. The import copy is gone, the refresh pipeline is gone, and the dashboard reflects the curated data as it stands rather than as it stood at the last refresh. Free report consumption for viewers is one of the capabilities that unlocks at the F64 capacity tier, which is part of why F64 anchors the economics in chapter three; Direct Lake itself runs on any Fabric capacity.
3.4 The medallion pattern on one lake
A single lake is not the same as an organised one. OneLake removes the copies; the medallion architecture organises what remains into rising tiers of quality, and it is the recommended pattern for structuring a Fabric lakehouse (Microsoft Learn, 2025–2026).
Bronze holds data as it arrived, an unedited record of source. Silver standardises formats, removes duplicates, and resolves each entity to one consistent representation. Gold organises the validated data into curated models fit for reporting and analytics. The tiers operate independently, so a new source can be exercised in bronze without disturbing the gold models the business consumes. Because all three tiers live in OneLake as one copy, the staging happens in place: the data is refined where it sits rather than copied between systems to be cleaned, which is precisely the copy-and-pipeline pattern the consolidated estate exists to remove.
3.5 Governance as a layer, not a bolt-on
In the fragmented estate, governance was the tool added last and on top: a catalogue reaching across products that did not share a security model, perpetually behind the data it was meant to govern. Fabric inverts that. Microsoft Purview is the governance control plane across the estate, supplying lineage, data-quality scoring, classification, and policy that travel with the data through ingestion, transformation, modelling, and consumption (Microsoft, 2025; industry analysis, 2025–2026).
Two properties make this consolidation rather than relabelling. Lineage is continuous and automatic across lakehouses, pipelines, and analytics, so the path from a report back through the gold model, the silver transformations, and the bronze source is recorded rather than reconstructed after an incident. Policy follows the data across one lake rather than being re-implemented per service, so a sensitivity label or access rule set once is honoured by every workload that reads the copy. OneLake security and Purview together govern one estate as a layer; in the fragmented world, the same control had to be configured separately in the warehouse, the lake, the BI tier, and the streaming stack, and the gaps between those configurations were where exposure lived.
This is also the bridge to the AI-readiness benefit, which this paper treats briefly and by reference. A consolidated, governed, current estate is the same estate an enterprise needs to ground AI agents trustworthily, one authoritative copy, measured quality, continuous lineage, and policy that follows the data. The full argument for why data quality, not the model, caps agent accuracy is the subject of the companion Volge paper, The data estate is the accuracy budget (Topic 7), and the reader who needs the agent-grounding case is directed there. The point for this paper is narrower and stands on the data leader's own terms: the governance that makes consolidation safe is the same governance that makes the estate AI-ready, so AI-readiness arrives as a by-product of the consolidation rather than as a separate programme.
Chapter three: The economics and the migration, and where Fabric does not fit
The takeaway
Fabric's commercial mechanism is one pooled capacity meter that replaces several per-service meters, and the total-cost comparison turns on capacity sizing rather than on headline prices. F64 is the inflection that unlocks Power BI Premium-equivalent features (chiefly free report viewing); Direct Lake itself runs on any Fabric capacity, which is why F64 anchors most enterprise deployments. The migration off Synapse, standalone Power BI Premium, and Azure Data Factory is a staged programme with real cutover risk, underwritten by the integration tax it retires. And consolidation is not the answer everywhere: Databricks and Snowflake retain genuine, specialised advantages, and the honest architecture for many estates is hybrid.
4.1 The capacity model
Fabric replaces per-service licensing with a single pooled compute meter. Capacity is purchased as an F-SKU, measured in capacity units, and that one pool of compute is shared across every workload, the warehouse, the lakehouse, the pipelines, the real-time engine, and Power BI all draw from it (Microsoft, 2026). This is the commercial expression of the architecture: one platform, one lake, one meter.
The SKU range is wide. Capacity runs from F2, at roughly 263 United States dollars per month, to F2048, at roughly 269,000 United States dollars per month, on pay-as-you-go monthly billing; reserved one-year commitments are materially cheaper, and enterprise-agreement rates are typically well below list (industry pricing analyses, 2026). Pricing changes and is region- and contract-specific, and the figures here should be reconfirmed against current Microsoft pricing before any commitment.
| F-SKU | Approx. capacity units | Approx. monthly list (PAYG) | Typical fit |
|---|---|---|---|
| F2 | 2 | ~USD 263 | Development, small workloads, evaluation |
| F8 | 8 | ~USD 1,050 | Departmental analytics |
| F64 | 64 | ~USD 5,257 | Inflection: Power BI Premium-equivalent features (free report viewers); common enterprise start |
| F128 | 128 | ~USD 10,515 | Large enterprise; common Fortune 500 starting band with F64 |
| F256 | 256 | ~USD 21,030 | Heavy, broad estate |
| F2048 | 2048 | ~USD 269,000 | Very large, multi-domain estate |
Figures are indicative pay-as-you-go list prices from industry pricing analyses, 2026; reserved and enterprise-agreement rates are lower. Pricing is region- and contract-specific and changes; reconfirm against current Microsoft pricing.
4.2 Why F64 is the inflection
The F64 tier, at roughly 5,257 United States dollars per month on pay-as-you-go, is where the economics turn, and for a specific reason. F64 is the point at which Fabric includes Power BI Premium-equivalent capabilities, notably free report consumption for viewers, since Direct Lake itself runs on any Fabric capacity (industry pricing analyses, 2026). Two consequences follow.
First, report consumers no longer need individual Power BI Pro licences. At F64 and above, viewers consume reports on free licences, which for an enterprise with thousands of report consumers can dominate the total-cost comparison on its own, the per-seat Pro licensing that a standalone Power BI deployment carries collapses into the capacity. Second, Direct Lake, which is available on any Fabric capacity, not gated at F64: removes the import-and-refresh seam described in chapter two and, with it, the duplicated import storage and the refresh pipelines.
By the numbers: Fabric capacity pricing (industry pricing analyses, 2026)
- ~USD 263/month: F2, the entry tier
- ~USD 269,000/month: F2048, the top tier
- ~USD 5,257/month: F64, the inflection: includes Power BI Premium-equivalent features (free report viewers); Direct Lake runs on any Fabric capacity
- Free licences: report viewers consume at F64 and above
Pricing is region- and contract-specific and changes; reconfirm against current Microsoft pricing.
This is why F64 to F128 is the common starting band for large enterprises and why the total-cost question is a capacity-sizing question rather than a headline-price comparison. The list price of an engine is the wrong unit of analysis; the right one is the total of capacity, retired per-seat licensing, reclaimed duplicate storage, and removed pipeline operation, set against the several meters the consolidated estate replaces.
4.3 Total cost of ownership, before and after
The total-cost case for consolidation is not that any one Fabric workload is cheaper than its specialist counterpart. It is that the consolidated estate removes whole categories of cost that the fragmented estate carries structurally. The table sets the two side by side.
| Cost component | Fragmented multi-service estate | Consolidated on Fabric |
|---|---|---|
| Compute licensing | Separate meter per service, each sized to its own peak | One pooled capacity meter shared across workloads |
| Storage | Duplicated copies per service | One copy in OneLake; copies reclaimed |
| BI licensing | Per-seat Pro for every consumer | Free viewers at F64+; licensing in the capacity |
| Pipeline operation | Copy pipelines between every service | Copy pipelines removed; data read in place |
| Idle reserved capacity | Redundant across non-coincident peaks | Single pooled sizing decision |
| Governance tooling | Separate catalogue and policy per service | Purview as a layer across one estate |
| Operational headcount | Specialists per service and per integration | One operational surface and skill base |
The saving concentrates in the rows that describe the seams, duplicated storage, copy pipelines, redundant capacity, per-seat licensing, and the headcount that maintains the integration. Those are precisely the integration-tax components of chapter one, and they are the components consolidation retires. The compute-licensing row is the one that requires care: a poorly sized capacity that throttles under load, or an oversized one that idles, can erase the saving, which is why capacity sizing and the governance of capacity consumption are where consolidation programmes are won or lost in practice.
4.4 The migration
The estate Fabric consolidates is, for most enterprises, an Azure estate already: Azure Synapse Analytics, Azure Data Lake Storage, Power BI Premium, and Azure Data Factory. The migration is the work of moving those onto the one platform, and it is a programme rather than a switch.
Microsoft has not declared Azure Synapse Analytics end-of-life wholesale, and that nuance matters for planning. What is clear is the strategic direction and a set of targeted deprecations: Synapse dedicated SQL pools receive security updates but no new features, Synapse Data Explorer is being retired in favour of the Fabric Eventhouse, and migration tooling, assessment-first pipeline and warehouse migration assistants, has shipped to support the move (Microsoft, 2025–2026). Fabric is positioned as the long-term platform, and the deprecations point one way.
Stage one: assess and land. Stand up the OneLake foundation and a right-sized capacity, inventory the Synapse, ADLS, Power BI Premium, and ADF estate, and run an assessment-first compatibility review. The objective is a landed platform and an honest map of what moves cleanly and what does not.
Stage two: migrate the warehouse. Move Synapse dedicated SQL workloads to the Fabric Data Warehouse, validating query parity and performance against the open Delta tables. This is the stage with the most cutover risk, because reporting depends on it; run it in parallel before cutting over.
Stage three: move pipelines and BI. Migrate Azure Data Factory pipelines to Fabric Data Factory, and move Power BI Premium semantic models to Direct Lake on OneLake, retiring the import copies and refresh schedules. The objective is the BI tier reading one copy in place.
Stage four: make current and decommission. Migrate streaming and event workloads to Real-Time Intelligence on the Eventhouse, then decommission the legacy services and their separate meters. The decommissioning is where the integration-tax saving is actually realised; until the old meters are switched off, the enterprise pays for both estates.
The cutover risk is real and is the reason for the parallel-run discipline. The dominant failure mode is decommissioning a legacy service before its Fabric replacement has been validated at production load, then discovering a capacity-sizing or feature-parity gap with no fallback. Migrate, validate in parallel, then decommission, in that order, never the reverse.
4.5 Where Fabric does not fit: the honest best-of-breed contrast
Consolidation is a trade, and the trade does not favour Fabric everywhere. A data leader is owed the honest version of where best-of-breed still wins.
Databricks is the stronger platform for large-scale data engineering and machine learning. It was founded by the creators of Apache Spark and built Delta Lake into a de facto open-format standard; its Spark performance, its granular Unity Catalogue governance, its mature ML and model-serving stack, and its multi-cloud reach are the reasons engineering-heavy and ML-heavy estates choose it (industry analysis, 2026). Its scale reflects that position: roughly 5.4 billion United States dollars in annual recurring revenue, growing about 65 per cent, with more than 650 customers above one million dollars a year (industry and equity-research analysis, 2026). Where the workload is large-scale, code-first data engineering or production ML at the frontier, Databricks retains a genuine ceiling Fabric does not match.
Snowflake is the stronger platform for high-concurrency SQL warehousing and frictionless, secure data sharing. Its separation of storage and compute, its elastic per-warehouse scaling, and its data-sharing and marketplace capabilities are mature and hard to beat for that profile (industry analysis, 2026). Its scale is substantial: more than 13,300 customers, with 733 customers above one million dollars in trailing-twelve-month product revenue as of the end of fiscal 2026 (Snowflake earnings, FY2026). Where the requirement is heavy concurrent SQL analytics across many tenants, or data sharing across organisational boundaries, Snowflake retains advantages Fabric does not fully close.
| Decision axis | Choose Fabric | Choose Databricks | Choose Snowflake |
|---|---|---|---|
| Primary goal | Consolidate a fragmented estate; cut integration tax | Large-scale data engineering and ML | High-concurrency SQL; data sharing |
| Ecosystem | Microsoft, Azure, Power BI estate | Multi-cloud, open, code-first | Multi-cloud, SQL-first |
| Strength | Unification; BI integration; SaaS simplicity | Spark depth; ML maturity; open governance | Concurrency; elastic scaling; sharing |
| Team profile | Mixed, BI-led, low-to-pro-code | Engineering- and ML-heavy | SQL- and analyst-heavy |
| Honest limit | Specialist engineering and warehousing ceiling | Less unified; heavier to operate | Narrower than a full unified estate |
The same trade reads as a selection decision-tree, mirroring the axes of the table above.
The decisive practical point is that this is not a single choice for the whole estate. Fabric's bi-directional, zero-copy interoperability with both Databricks and Snowflake means an enterprise can consolidate the fragmented majority of its estate onto Fabric while leaving a specialised engineering or warehousing workload on the platform that serves it best, sharing data between them without copying it (Microsoft, 2025–2026). The honest architecture for a large estate is frequently hybrid, and the interoperability is what makes hybrid coherent rather than a return to the copy sprawl consolidation set out to remove.
Chapter four: Reference architecture, the adoption roadmap, and a maturity model
The takeaway
A durable consolidated estate is layered, open, and governed end to end: sources unified into OneLake, refined through the medallion tiers, governed through Purview, made current through Real-Time Intelligence, and consumed by Power BI through Direct Lake and by every other workload on one copy. The roadmap that delivers it is staged on purpose, because decommissioning legacy services before their replacements are validated is the dominant way these programmes fail. The maturity model gives leadership a way to locate the estate and name the next increment honestly.
5.1 A reference architecture
The architecture below is anchored to Fabric, in line with the paper's lens, but the layering is portable. Each layer can be sourced, governed, and evolved on its own, which is the structural hedge against both lock-in and instability.
Read the architecture as one governed pipeline feeding several consumers from a single copy. The unification choices on the left, mirroring for operational systems that change often, shortcuts for large data already sitting in a governed lake, connectors for SaaS, and direct ingestion for streams: all converge on one OneLake copy. Every consumer on the right reads that copy rather than its own re-export. The governance plane spans both. That is the consolidated estate; everything else is sizing and sequencing.
5.2 The adoption and migration roadmap
The roadmap is the migration of chapter three set against the maturity model below. It is disciplined on validation and conservative on decommissioning by design, because the failure mode is always the same: a legacy meter switched off before its replacement has earned the trust.
Phase one, foundation. Land OneLake, size an initial capacity at F64 or F128, and connect the first sources through mirroring and shortcuts. Stand up Purview from the start rather than after. The objective is a governed platform and a baseline.
Phase two, consolidate the core. Migrate the warehouse and the primary pipelines, build the medallion tiers, and move the highest-value Power BI models to Direct Lake. Run the legacy and Fabric estates in parallel. The objective is the core reporting estate reading one copy.
Phase three, extend and make current. Add Real-Time Intelligence where recency determines accuracy, extend to the data-science workloads, and bring adjacent domains onto the platform. The objective is the breadth of the estate consolidated, not just the core.
Phase four, decommission and optimise. Retire the validated legacy services and their meters, tune capacity against observed consumption, and govern capacity to prevent throttling and idle waste. The objective is the realised saving, which arrives only when the old meters are switched off.
5.3 A data-platform maturity model
The model locates the enterprise and defines its next increment. It is the diagnostic counterpart to the roadmap.
| Level | Description | Estate and storage | Governance and capacity | Typical outcome |
|---|---|---|---|---|
| 0: Fragmented | Separately licensed services; copies everywhere | Many stores; duplicated copies | Per-service; bolted-on catalogue | Full integration tax; trust erosion |
| 1: Landed | OneLake stood up; first sources connected | One lake emerging beside legacy | Purview started; capacity sized | Foundation set; saving not yet realised |
| 2: Core consolidated | Warehouse and core BI on Fabric; Direct Lake live | One copy for core; legacy in parallel | Lineage and policy active; capacity tuned | Copies and import refreshes retired for the core |
| 3: Estate consolidated | Pipelines, science, real-time on platform | One copy across the estate | Policy follows data; capacity governed | Most integration tax removed |
| 4: Optimised unified estate | Legacy decommissioned; hybrid only by design | One copy; best-of-breed by deliberate choice | Estate-wide governance; capacity optimised | Saving realised; AI-ready by-product |
Most enterprises sit between Level 0 and Level 1 as of mid-2026, with consolidation under way but legacy meters still running in parallel. The saving is realised at Level 4, when the legacy services are decommissioned, not at Level 1, when the platform is merely landed. Any residual best-of-breed workload at Level 4 is hybrid by deliberate design, connected through zero-copy interoperability rather than by reintroduced copies.
5.4 The before-and-after, summarised
The consolidation reads most clearly as a shift in the conditions under which the estate operates.
| Condition | Fragmented estate | Consolidated on Fabric |
|---|---|---|
| Data location | Copies across many services | One copy in OneLake |
| Storage format | Proprietary per service | Open Delta Parquet, read by all |
| BI delivery | Import copies and scheduled refresh | Direct Lake on one copy |
| Compute meter | One per service | One pooled capacity |
| Governance | Bolted-on, per service | Purview as a layer across the estate |
| Real-time | Separate streaming stack | Real-Time Intelligence on-platform |
| Operational surface | Several teams and skill sets | One platform, one skill base |
The architecture names the target, the roadmap sequences the journey, and the maturity model measures the distance still to travel. Follow the sequence and the estate consolidates onto one governed copy and the integration tax is retired; skip the validation discipline and the programme either throttles in production or pays for two estates at once.
Chapter five: The data-leadership agenda
Consolidating the analytics estate onto one platform is the defining data-platform decision of the present cycle, and the evidence in this paper resolves to a short list of conclusions for the chief data officer and the data-and-analytics leadership.
The cost, latency, and trust that leak from a fragmented estate leak at the seams between services, not inside any one engine. Consolidation is the lever because it removes the seams, the copies, the pipelines, the redundant meters, rather than tuning the engines. Fabric delivers that consolidation through OneLake, the shared workloads, Direct Lake, and Purview, and the saving it offers is the integration tax it retires, set against the specialised depth a best-of-breed estate keeps.
For most enterprises the platform decision is shaped by the existing Azure and Power BI estate as much as by a feature comparison, and the real question is sequencing and sizing rather than selection. The migration off Synapse, Power BI Premium, and Azure Data Factory is a staged programme underwritten by the saving it realises, and the discipline that determines its success is validation before decommissioning, never the reverse.
Five questions belong on the leadership agenda. First, is the integration tax, duplicated storage, copy pipelines, redundant capacity, trust erosion, actually quantified, or merely assumed. Second, is the estate consolidating onto one copy in OneLake, or are copies still proliferating across services. Third, is the capacity right-sized and governed, given that a mis-sized capacity erases the saving the consolidation was meant to deliver. Fourth, does the migration validate each replacement in parallel before the legacy meter is switched off. Fifth, where best-of-breed genuinely wins, large-scale engineering on Databricks, high-concurrency warehousing on Snowflake, is the hybrid deliberate and connected by zero-copy interoperability, rather than an accident of unfinished consolidation.
Volge's work is in the execution. The platform decision is, for most enterprises, shaped by existing Microsoft commitments; the disciplines that convert that commitment into a realised saving, the integration-tax case, the capacity model, the staged migration, the honest best-of-breed boundary, are where data programmes succeed or stall, and where Volge concentrates. We would welcome the chance to apply this framework to a specific estate in detail.
Appendices
Appendix A: Platform capability comparison: Fabric, Databricks, Snowflake
| Capability | Microsoft Fabric (OneLake) | Databricks | Snowflake |
|---|---|---|---|
| Delivery model | End-to-end SaaS, unified workloads | Lakehouse platform, multi-cloud | Cloud data warehouse, multi-cloud |
| Unified logical lake | OneLake, provisioned per tenant | Unity Catalogue over lakehouse | Single platform; lake plus warehouse |
| Open storage format | Delta Parquet, open by default | Delta Lake and Iceberg | Iceberg and proprietary |
| Greatest strength | Unification; BI and Microsoft-ecosystem integration | Large-scale data engineering and ML on Spark | High-concurrency SQL; data sharing |
| BI integration | Native Power BI with Direct Lake | Connects to external BI | Connects to external BI |
| Real-time | Real-Time Intelligence, Eventhouse, KQL | Structured streaming | Snowpipe streaming |
| Governance | Microsoft Purview, integrated | Unity Catalogue | Horizon governance |
| Capacity / pricing | Pooled F-SKU capacity; one meter | Compute per workload; consumption | Per-warehouse compute; consumption |
| Principal constraint | Specialist engineering/warehousing ceiling | Less unified; heavier to operate | Narrower than a full unified estate |
| Interoperability | Bi-directional zero-copy with both | Reads OneLake via Unity Catalogue | Interoperability with Fabric |
Fabric is the primary lens of this paper. Interoperability is treated as a design feature, not an exception; a consolidated estate need not be single-vendor. Scale figures: Databricks ~USD 5.4bn ARR, 650+ customers above USD 1M (industry/equity-research analysis, 2026); Snowflake 13,300+ customers, 733 above USD 1M trailing-twelve-month product revenue (Snowflake earnings, FY2026). Sources: Microsoft and vendor disclosures and industry analysis, 2025–2026.
Appendix B: Migration and readiness checklist
A minimum set of conditions for a defensible consolidation programme, organised by phase.
Assess and land
- The integration tax is quantified: duplicated storage, copy pipelines, redundant reserved capacity, and per-seat BI licensing are itemised against current spend.
- The OneLake foundation is provisioned and an initial capacity is sized, with F64 or F128 as the typical enterprise start.
- The existing estate, Synapse, ADLS, Power BI Premium, ADF, is inventoried, with an assessment-first compatibility review completed.
- Microsoft Purview is stood up from the outset, not deferred to a later phase.
Consolidate
- Operational systems that change frequently are mirrored into OneLake; large data already in a governed lake is exposed through shortcuts rather than copied.
- The medallion tiers are built, with deduplication and entity resolution in silver and curated models in gold.
- High-value Power BI models are moved to Direct Lake, retiring import copies and refresh schedules.
- The legacy and Fabric estates run in parallel; nothing is decommissioned before its replacement is validated at production load.
Govern and optimise
- Continuous lineage traces every report to its source; classification and access policy follow the data across the lake.
- Capacity consumption is monitored and governed to prevent both throttling and idle waste.
- Legacy services and their separate meters are decommissioned only after parallel validation, and the realised saving is measured.
- Any retained best-of-breed workload is connected through zero-copy interoperability, not reintroduced copies.
Appendix C: Glossary
- Capacity unit (CU). The unit of pooled compute a Fabric F-SKU provides, shared across all workloads.
- Activator (formerly Data Activator). The Fabric workload that triggers actions when data meets defined conditions.
- Delta Parquet / Delta Lake. The open table format in which OneLake stores data, readable by every engine without conversion.
- Direct Lake. The Power BI storage mode that reads Delta tables in OneLake directly, without import or scheduled refresh.
- Eventhouse. The Real-Time Intelligence engine for high-volume time-series and event data, queried with KQL.
- F-SKU. A Fabric capacity tier (F2 to F2048); F64 is the inflection that unlocks Power BI Premium-equivalent features such as free report viewing; Direct Lake runs on any Fabric capacity.
- Integration tax. The distributed cost of holding a fragmented estate together: duplicated storage, copy pipelines, redundant capacity, and trust erosion.
- Lakehouse. A construct combining lake-scale storage with warehouse structure and performance, sitting on OneLake.
- Medallion architecture. The staged design pattern refining raw data through bronze, silver, and gold tiers.
- Mirroring. Near-real-time replication of an operational database into OneLake.
- OneLake. Fabric's single, unified, logical data lake for the entire organisation, provisioned per tenant.
- Purview. Microsoft's governance control plane for lineage, data quality, classification, and policy across the estate.
- Real-Time Intelligence. The Fabric workload for ingesting and analysing streaming and event data.
- Shortcut. A reference that exposes externally stored data in OneLake without copying it.
- Workload. One of Fabric's role-tailored experiences, data engineering, warehousing, science, real-time intelligence, data factory, Power BI, and Activator, over shared OneLake storage.
Appendix D: Sources
This paper draws on the following, accessed June 2026. Figures are attributed in-text to their primary source. Where a statistic appears across multiple secondary outlets, the most authoritative primary attribution is given.
- Microsoft: Fabric platform documentation and announcements (2025–2026): the seven workloads over OneLake; consolidation of Azure Synapse, Azure Data Lake Storage, Power BI Premium, and Azure Data Factory; OneLake as one logical lake; Direct Lake mode; mirroring, shortcuts, and Databricks/Snowflake interoperability.
- Microsoft: FY2026 earnings disclosures on Fabric: roughly 35,000 paid customers, up about 60 per cent year over year; OneLake data volume up nearly four times year over year; annual revenue run rate above two billion United States dollars.
- Microsoft earnings (FY2026), approximately 35,000 paid Fabric customers, up about 60 per cent year over year; OneLake data volume up nearly four times year over year; annual revenue run rate above two billion United States dollars. (The adoption and paid-customer figures use different definitions and are attributed to their distinct sources.)
- Microsoft Learn, medallion lakehouse architecture; Direct Lake overview; Real-Time Intelligence and Eventhouse guidance (2025–2026).
- Microsoft: Synapse and Azure Data Factory to Fabric migration guidance; targeted deprecations (Synapse dedicated SQL pools, Synapse Data Explorer); assessment-first migration assistants (2025–2026).
- Microsoft Security and Purview blogs, continuous lineage, data observability, and AI-ready governance across Fabric (2025).
- Industry pricing analyses (2026): Fabric F-SKU pricing from F2 (~USD 263/month) to F2048 (~USD 269,000/month); F64 (~USD 5,257/month) as the inflection including Power BI Premium-equivalent features (free report viewing), Direct Lake runs on any Fabric capacity; F64–F128 as the common enterprise starting band. Pricing is region- and contract-specific and changes; reconfirm against current Microsoft pricing.
- Industry and equity-research analysis (2026): Databricks at roughly USD 5.4 billion annual recurring revenue, ~65 per cent growth, 650+ customers above USD 1 million; Apache Spark and Delta Lake heritage and engineering/ML positioning.
- Snowflake earnings (FY2026), more than 13,300 customers, with 733 above USD 1 million in trailing-twelve-month product revenue; cloud-data-warehouse positioning and data-sharing strength.
- Industry survey (2024): data-silo losses estimated at 5.5 trillion United States dollars; 68 per cent ranking silos as their single biggest data challenge.
- Volge: The data estate is the accuracy budget (Topic 7), the companion paper carrying the agent-grounding argument referenced once here.
Addendum: How Volge's forward-deployed engineers shall help: Data & Analytics
The Volge engagement team
Volge delivers through one accountable team that runs from the boardroom to the production branch. Four roles carry the engagement, and they hold their remit from the first workshop to operational handover rather than handing off between phases.
| Role | What they own |
|---|---|
| Principal | The client relationship and the account. Sets the engagement's business outcomes, the governance cadence, and the value tracking, and stays the single point of accountability throughout. |
| Architect | The target architecture and the transformation roadmap. Sequences the work, sets the standards, and makes the platform, build-versus-buy, and control-design calls. |
| Principals | The senior engineering leads. Each owns a workstream, makes the hard technical calls, sets the quality bar, and mentors the embedded engineers. |
| Forward-deployed engineers | The embedded build. They sit inside the client's team, build and ship in the client's own environment, and own the last mile to production. |
The Principal holds the outcome, the Architect holds the design, the Principals hold the workstreams, and the forward-deployed engineers hold the keyboard, one team, accountable from strategy to shipped software.
This paper makes the case that the saving sits in the seams, and that consolidating the analytics estate onto Microsoft Fabric and OneLake, one platform, one copy of the data, one governance plane, one capacity meter, is what retires the integration tax. The distance between that case and a realised saving is execution: the staged migration off Synapse, standalone Power BI Premium, and Azure Data Factory, the medallion structuring of OneLake, the certified semantic models, and the Purview governance that makes the single copy trustworthy. Volge closes that distance with forward-deployed engineers, and that delivery model is the subject of this addendum.
A forward-deployed engineer is a senior Volge data engineer embedded directly in the client's team, building and migrating inside the client's own Fabric and OneLake environment rather than handing over a deck and a plan. The FDE works on the real estate, the client's Synapse workloads, the client's Power BI Premium models, the client's Azure Data Factory pipelines, and stays through to a governed, operating estate, not to a pilot. The point of the model is ownership of the last mile: the FDE carries the consolidation from assessment to a decommissioned legacy meter and a single governed copy in production, which is where the integration tax is actually retired and where most programmes stall when execution is left to a hand-off. It is in that execution that Volge concentrates its Data & Analytics practice, building the consolidated, governed estate this paper describes rather than presenting another feature comparison.
What Volge delivers
The capability set below maps the Data & Analytics offering to the client outcome each capability is intended to produce. Each is drawn from work Volge does on the Microsoft data platform, and each addresses one of the seams this paper identifies.
| Capability | Client outcome |
|---|---|
| Fabric and OneLake implementation | A single tenant-wide lake holding one copy of the data in open Delta Parquet, reachable by every workload, so the duplicated storage and copy pipelines of the fragmented estate have nothing left to do. |
| Lakehouse and medallion architecture | Raw, cleaned, and curated data staged in place through bronze, silver, and gold, with deduplication and entity resolution in silver, so analytics and AI consume business-ready data rather than a re-exported slice. |
| Synapse, Power BI Premium, and ADF migration | A supported, assessment-first move from the legacy Azure estate onto Fabric, validated in parallel before any legacy meter is switched off, so the migration realises its saving rather than paying for two estates. |
| Power BI semantic models and BI governance | Certified semantic models with centralised DAX measures and row-level security, so one endorsed definition of each metric serves many consumers and two executives no longer arrive with different numbers. |
| Purview governance, lineage, and classification | A catalogued, classified estate with continuous lineage from report back to source, so sensitivity labels and access policy follow the data across the single copy and a regulator's question is answerable. |
| Data pipelines and ETL/ELT | Idempotent, monitored Fabric Data Factory pipelines with quality gates between the medallion layers, so every figure traces through an auditable transformation to source and the estate stays current. |
The modernisation engagement
Volge's data-estate modernisation runs as a staged engagement, disciplined on validation and conservative on decommissioning, because the dominant failure mode is a legacy meter switched off before its replacement has earned trust.
The saving is realised at the final stage, when the validated legacy services and their separate meters are decommissioned, not when the platform is merely landed.
The Volge difference
Four disciplines separate a consolidation that realises its saving from one that throttles in production or pays for two estates at once.
Migration with minimal disruption. Volge runs the move as a programme, not a switch: the legacy and Fabric estates run in parallel, each replacement is validated at production load before its predecessor is retired, and the warehouse cutover, the stage carrying the most reporting risk, is parallel-run before it is trusted. Migrate, validate, then decommission, in that order, never the reverse.
Single-copy and governance discipline. Operational systems that change often are mirrored into OneLake and large data already in a governed lake is exposed through shortcuts rather than copied, so the single copy is held rather than quietly reintroduced. Purview is stood up from the outset, so classification and lineage travel with the data rather than being reconstructed after an incident.
Capacity and cost optimisation. Capacity is right-sized against observed consumption and governed to prevent both throttling and idle waste, because a mis-sized capacity is the one row in the total-cost case that can erase the saving the consolidation was meant to deliver. Organisations migrating to Fabric report a 30 to 40 per cent reduction in total cost of ownership, and Fabric capacity right-sizing and tuning typically recover a further 25 to 45 per cent (Microsoft and Fabric FinOps guidance, 2026; illustrative, and to be validated against the specific estate).
Accelerators and the feedback loop. Because the FDE builds inside the client's environment, every engagement feeds the next: reusable migration accelerators, medallion and semantic-model reference architectures, and a maturity diagnostic that locates the estate and names its next increment honestly are carried from one estate to the following one, so each engagement starts from a known position rather than a blank sheet. The relationship is outcomes-based, the FDE is measured against the decommissioned legacy meter and the realised saving, not against hours billed, which aligns the firm and the client on the same target and is what makes the embedded model work for both.
Capability to outcome
The map below reads the offering as a single governed pipeline: the Volge capabilities on the left converge on one OneLake copy, and the client outcomes on the right are read from that copy rather than from a re-export.
A bounded estate assessment: the first step. The concrete entry point is a fixed-scope estate assessment: an inventory of the Synapse, Azure Data Lake Storage, Power BI Premium, and Azure Data Factory estate, an itemised quantification of the integration tax it carries, duplicated storage, copy pipelines, redundant reserved capacity, and per-seat BI licensing, and a right-sized capacity recommendation with a staged, validation-first migration plan. It is bounded, it commits the enterprise to nothing beyond the assessment itself, and it converts the consolidation case in this paper into a costed plan for one specific estate. We would welcome the opportunity to run it.
Prepared by Volge · June 2026. This document is a point of view intended to inform data and analytics leadership decision-making. It is not financial, licensing, or investment advice. Microsoft Fabric capacities, feature availability, and pricing change and are contract- and region-specific; the capacity and pricing figures here are indicative as of June 2026 and should be reconfirmed against current Microsoft pricing before they are relied upon for budgeting or architectural decisions. Statistics are attributed to their sources and were accurate to the best of available knowledge as of June 2026; the Microsoft Fabric platform and the wider data and analytics field are moving rapidly, and figures should be reconfirmed before they are relied upon.
Voice of Volge: the firm's point of view on the technology it delivers. Sourced, year-stamped, and built to be handed to a board.
Volge is an independent deep-tech enterprise service provider and premier Microsoft delivery partner. That interest is disclosed; the analysis in this book is independently sourced.