>
volgeMicrosoft · 2027
DATA & ANALYTICS · VOICE OF VOLGE

One platform, one copy of the data

OneLake, the lakehouse, real-time intelligence, and the economics of one platform

Voice of Volge is a series of books, and the firm's thought leadership on the solutions and the domains Volge specialises in. Each edition takes one of them and sets out what the technology makes possible, what it makes mandatory, and what an operator has to put in place before either is true of them.

OneLake, the lakehouse, real-time intelligence, and the economics of one platform

Figures carry their basis and their date. Where published estimates disagree, the range is given and the basis named rather than averaged away, and where a widely quoted number does not survive checking it is corrected here rather than repeated. Every chart plots a figure stated in the text, and nothing is projected beyond what a named source projects.

Contents

  1. 02Executive summary
  2. 03, The fragmented analytics landscape and the integration tax
  3. 05, The Fabric architecture
  4. 08, The economics and the migration, and where Fabric does not fit
  5. 11, Reference architecture, controls and operations
  6. 15, The data platform in the Gulf and India
  7. 20, The data-leadership agenda
  8. 21Appendices
Voice of Volge is a series of books, and the firm’s thought leadership on the solutions and the domains Volge specialises in.
Microsoft · 2027
Data & AnalyticsVoice of Volge01 •

Executive summary

The ana­lyt­ics land­scape 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 land­scape 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 land­scape runs. Mi­crosoft Learn lists nine of them: Power BI, Data­bases, Data Fact­ory, In­dustry Solu­tions, Real-Time In­tel­li­gence, Data En­gin­eer­ing, Data Sci­ence, Data Ware­house, and Fab­ric IQ, which is in pre­view. They run over one shared lake, OneLake, under one gov­ernance model and one ca­pa­city model (Mi­crosoft Learn, 2026). It con­sol­id­ates what were sep­ar­ately li­censed Azure ser­vices, among them 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 and 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 44,000 paid Fab­ric cus­tom­ers as at the first quarter of fin­an­cial year 2027, up more than 50 per cent year over year and 4,000 more than a quarter earlier (Mi­crosoft earn­ings, Q1 FY2027, 28 Oc­to­ber 2026). The same call put more than 19,000 cus­tom­ers on Foundry and Fab­ric to­gether, which is the clearest pub­lished meas­ure of en­ter­prises wir­ing agents to gov­erned op­er­a­tional data, and Foundry it­self at 120,000 cus­tom­ers against 100,000 a quarter earlier. Earlier dis­clos­ures fill in the tra­ject­ory. The volume of data held in OneLake was up nearly

four times year over year (Mi­crosoft earn­ings, Q3 FY2026, 29 April 2026), and the an­nual rev­enue run rate passed two bil­lion United States dol­lars along­side more than 31,000 cus­tom­ers (Mi­crosoft earn­ings, Q2 FY2026, 28 Janu­ary 2026). 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, set against the spe­cial­ised depth a best-of-breed land­scape re­tains. What it re­moves is one copy of the data, one gov­ernance plane and one ca­pa­city meter, and one op­er­a­tional sur­face. Where the land­scape 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.

A second find­ing sits in front of the first, and most con­sol­id­a­tion cases omit it. The plat­form is not avail­able every­where its in­fra­struc­ture is. Qatar Cent­ral, a live Azure re­gion in Doha, car­ries Power BI and not the Fab­ric work­loads, on Mi­crosoft's own re­gion avail­ab­il­ity page (Mi­crosoft Learn, page up­dated 5 Novem­ber 2026). Saudi Ar­a­bia has no Azure re­gion at all, and no Mi­crosoft 365 res­id­ency com­mit­ment of any kind (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). For a Gulf or In­dian en­ter­prise the con­sol­id­a­tion ques­tion there­fore has a prior ques­tion in front of it. Where may each data do­main law­fully rest, and where may it law­fully be pro­cessed. Mi­crosoft com­mits to stor­age at rest in the ten­ant geo­graphy. It does not com­mit to pro­cessing in re­gion. The European Union Data Bound­ary is the only pro­cessing com­mit­ment it pub­lishes any­where.

This is a paper about con­sol­id­a­tion eco­nom­ics, not about ar­ti­fi­cial in­tel­li­gence. AI-read­i­ness is one out­come of a con­sol­id­ated, gov­erned land­scape, 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, namely con­sol­id­a­tion of the land­scape and mi­gra­tion off leg­acy ser­vices, cost and ca­pa­city, real-time ana­lyt­ics and gov­ernance, and busi­ness in­tel­li­gence.

Six con­clu­sions, de­veloped in the chapters that fol­low.

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

Data & AnalyticsVoice of Volge02 •

op­er­ate on one copy. Re­move copy sprawl at the stor­age layer and the con­tra­dic­tions it man­u­fac­tures dis­ap­pear up­stream of every con­sumer.

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.

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 and Dir­ect Lake be­come avail­able. That is why it an­chors most en­ter­prise de­ploy­ments, and why the total-cost com­par­ison turns on ca­pa­city siz­ing rather than on head­line list prices (Mi­crosoft Fab­ric pri­cing, veri­fied 8 Novem­ber 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 land­scape is fre­quently hy­brid, and Fab­ric's zero-copy in­ter­op­er­ab­il­ity with both plat­forms is what makes hy­brid vi­able rather than a con­tra­dic­tion.

Sixth, the re­gion de­cides the shape of the land­scape be­fore the eco­nom­ics do. India, Qatar and the United Arab Emir­ates hold dur­able Mi­crosoft 365 com­mit­ments on data loc­a­tion. Four Gulf states hold none. Fab­ric's full work­load set runs in UAE North, Cent­ral India and South India (Mi­crosoft Learn, page up­dated 5 Novem­ber 2026). It does not run in Qatar Cent­ral, UAE Cent­ral or India West. In­dian res­id­ency rules bite per class of data rather than per ten­ant, so pay­ment data, in­sur­ance re­cords and tele­com­mu­nic­a­tions net­work data each carry a rule of their own. A con­sol­id­a­tion de­signed on one global ca­pa­city meets sev­eral of those rules at once. The se­quence that works puts clas­si­fic­a­tion be­fore place­ment, and place­ment be­fore pro­vi­sion­ing. Chapter five sets out the in­stru­ments and the vendor foot­print be­hind each of those state­ments.

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. On Volge's read­ing the con­sol­id­a­tion de­cision is an eco­nom­ics de­cision be­fore it is a tech­no­logy one, and it does not sur­vive being made on a fea­ture sheet.

, The fragmented analytics landscape and the integration tax

The analytics landscape 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.

How the landscape fragmented

No one de­signed the frag­men­ted land­scape. 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 land­scape 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.

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

Data & AnalyticsVoice of Volge03 •

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 land­scape car­ries idle re­served ca­pa­city across sev­eral ser­vices at once. That ca­pa­city is paid for and pro­vi­sioned, and it sits 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 land­scape, and they are man­u­fac­tured by copy sprawl, not by any en­gine being wrong.

The seams, not the services. The integration tax is distributed by construction, which is why no invoice carries it and why it survives the scrutiny a line item would not. It is spread across pipeline maintenance, duplicated storage, idle reserved capacity on several meters at once, and the reconciliation work that follows every disagreement between two copies of the same record.

Quantifying the tax before arguing about it

A con­sol­id­a­tion case stands or falls on a baseline taken be­fore the mi­gra­tion, and most pro­grammes never take one. The meas­ure­ment is not dif­fi­cult. It is four counts, each already held some­where in the land­scape, and each in a unit a fin­ance func­tion will ac­cept. Du­plic­ated stor­age is meas­ured in bytes. Take the gross foot­print of every land­ing, ware­house and ex­tract in­stance of a data­set, over the foot­print of the au­thor­it­at­ive copy. A ratio above two says the land­scape is pay­ing three times to hold one re­cord. Pipeline toil is meas­ured in en­gin­eer­ing days a month, against the count of

pipelines whose only pur­pose is to move data between two ser­vices. Take the days from the change log and the on-call re­cord, never from an es­tim­ate.

The other two counts are the ones that move the num­ber. Re­served ca­pa­city util­isa­tion is read per meter, as con­sumed units over re­served units across a full busi­ness cycle, be­cause the waste sits in the peak each ser­vice re­serves for it­self and never shares. Re­con­cili­ation ef­fort is meas­ured in hours per close, from the fin­ance and data teams' own time re­cords. It should then be split between the re­con­cili­ation a reg­u­lator re­quires and the re­con­cili­ation two cop­ies of one re­cord re­quire. Only the second is the tax. The first sur­vives con­sol­id­a­tion and should never be claimed as a sav­ing.

A de­fens­ible baseline has four prop­er­ties. The meas­ure­ment win­dow is stated and cov­ers at least one full quarter, each fig­ure has a named owner, the raw ex­tract is re­tained, and the method is writ­ten down be­fore the res­ult is known. A baseline is in­ad­miss­ible where it rests on a vendor's cost cal­cu­lator, where it counts head­count the en­ter­prise would carry in any case, or where it prices trust erosion at a num­ber nobody can trace to a source. Trust erosion is real and should be de­scribed rather than priced. The quan­ti­fied tax is the floor of the case, not the whole of it.

The diagram of the problem

The frag­men­ted land­scape and the con­sol­id­ated al­tern­at­ive are easi­est to read side by side. The left is the ac­cu­mu­lated land­scape; the right is the same work­loads over one shared lake.

The left-hand land­scape spends its ef­fort on the dot­ted lines , the cop­ies and the cata­logue reach­ing across products that do not share stor­age. The right-hand land­scape re­moves the dot­ted lines by re­mov­ing the cop­ies. That is the whole con­sol­id­a­tion thesis in one pic­ture.

Data & AnalyticsVoice of Volge04 •
By the numbers: Fabric adoption (Microsoft earnings, Q1 FY2027, 28 October 2026; Q3 FY2026, 29 April 2026; Q2 FY2026, 28 January 2026) - 44,000 paid Fabric customers, up more than 50 per cent year over year (Q1 FY2027; 40,000 in Q4 FY2026) - > 19,000 customers running Foundry and Fabric together (Q1 FY2027) - 120,000 Foundry customers, against 100,000 a quarter earlier (Q1 FY2027) - ~4x growth in OneLake data volume year over year (Q3 FY2026) - > USD 2 billion annual revenue run rate, with more than 31,000 customers (Q2 FY2026)

Why consolidation is the lever

The re­flex fix for a slow or ex­pens­ive land­scape 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.

, The Fabric architecture

OneLake, the workloads, Direct Lake, governance

The takeaway

Fab­ric is one soft­ware-as-a-ser­vice plat­form, nine 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 gov­ernance ar­rives as a layer rather than a bolt-on, through the OneLake Cata­log for dis­cov­ery and item gov­ernance and through Mi­crosoft Pur­view for pro­tec­tion and com­pli­ance. Un­der­stand OneLake and the rest of the plat­form is a con­sequence of it.

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 own store, sur­face in­side OneLake without being copied across. The land­scape 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 land­scape, 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. Tables live in Delta Par­quet rather than a pro­pri­et­ary ware­house format, so every en­gine reads the same files without con­ver­sion, which fore­closes the lock-in and the re-ex­port pro­lif­er­a­tion that pro­pri­et­ary formats im­pose.

Data & AnalyticsVoice of Volge05 •

The workloads over one lake

Fab­ric uni­fies the work­loads a com­plete ana­lyt­ics land­scape 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 land­scape prob­lem it ad­dresses.

Fabric workloadWhat it doesConsolidatesLandscape problem it addresses
Power BISemantic models and reports, Direct LakePower BI PremiumBI queries OneLake in place; no import refresh
DatabasesSQL database in Fabric for operational dataA separately licensed operational database tierOperational data lands in OneLake without an export
Data FactoryPipelines, mirroring, 200+ connectorsAzure Data FactoryIngestion and integration on-platform; mirroring replaces copy pipelines
Industry SolutionsIndustry-specific data models and solutionsBespoke industry build-outDomain models arrive on the platform rather than as a project
Real-Time IntelligenceStreaming and event analytics via Eventhouse and KQL, with Activator for triggers and actionsSynapse Data Explorer; separate streaming stack; bespoke alerting glueCloses the recency gap, and acts on conditions, without a standalone meter
Data EngineeringSpark notebooks and lakehouse over OneLakeSynapse Spark; lake computeLake and engineering on the shared copy, no separate store
Data ScienceNotebooks, model training, MLflowSynapse / Azure ML notebooksModels train on the curated lake, not on a re-exported slice
Data WarehouseT-SQL warehouse on open Delta tablesSynapse dedicated SQL poolsWarehouse reads the same files as the lake; no copy between them
Fabric IQ (preview)Ontology, Fabric Graph, semantic models, data and operations agentsBespoke semantic layerOne set of business semantics over the shared copy

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 nine now share one copy and one meter.

Direct Lake , closing the BI seam

The clas­sic busi­ness-in­tel­li­gence ar­chi­tec­ture car­ries one of the most ex­pens­ive seams in the land­scape, 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. So is the re­fresh pipeline, and the dash­board re­flects the cur­ated data as it stands rather than as it stood at the last re­fresh. Dir­ect Lake is one of the cap­ab­il­it­ies that be­comes avail­able at the F64 ca­pa­city tier, which is part of why F64 an­chors the eco­nom­ics in chapter three.

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 land­scape ex­ists to re­move.

Two prop­er­ties make this con­sol­id­a­tion rather than re­la­belling. Lin­eage is con­tinu­ous and auto­matic across lake­houses and pipelines, and across ana­lyt­ics, so the path from a

Data & AnalyticsVoice of Volge06 •

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. The OneLake Cata­log, OneLake se­cur­ity, and Pur­view to­gether gov­ern one land­scape 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.

Governance as a layer, not a bolt-on

In the frag­men­ted land­scape, 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, and it does so across two sur­faces rather than one. The OneLake Cata­log is the cent­ral­ised ex­per­i­ence for dis­cov­er­ing, ex­plor­ing, se­cur­ing, and gov­ern­ing Fab­ric items, and for as­sess­ing gov­ernance state and act­ing on re­com­mend­a­tions. Mi­crosoft Pur­view sup­plies in­form­a­tion pro­tec­tion, data loss pre­ven­tion, audit, and risk and com­pli­ance for sup­por­ted Fab­ric data, with lin­eage and clas­si­fic­a­tion that travel with the data from in­ges­tion and trans­form­a­tion through mod­el­ling to con­sump­tion (Mi­crosoft Learn, 2026).

This is also the bridge to the AI-read­i­ness be­ne­fit, which this paper treats briefly and by ref­er­ence. A con­sol­id­ated, gov­erned, cur­rent land­scape is the same land­scape an en­ter­prise needs to ground AI agents trust­wor­thily. That means one au­thor­it­at­ive copy with 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). 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 land­scape AI-ready, so AI-read­i­ness ar­rives as a by-product of the con­sol­id­a­tion rather than as a sep­ar­ate pro­gramme.

What has to be decided before anything lands

Four de­cisions gov­ern a Fab­ric land­scape, and all four are taken be­fore the first data­set lands. They are the ca­pa­city and ten­ant to­po­logy, the do­main and work­space bound­ary, the ac­cess pat­tern chosen for each source, and the own­er­ship of each medal­lion layer. Each is cheap to de­cide and ex­pens­ive

to re­visit, be­cause each be­comes a de­pend­ency of everything built on it. A pro­gramme that be­gins by pro­vi­sion­ing a ca­pa­city and cre­at­ing work­spaces on re­quest has taken all four de­cisions by de­fault and doc­u­mented none of them.

Ca­pa­city to­po­logy an­swers who pays and who can throttle whom. One ca­pa­city for the whole or­gan­isa­tion gives the best pool­ing and the worst blast ra­dius, be­cause a single run­away note­book con­sumes the units the fin­ance close needs. One ca­pa­city for each busi­ness unit in­verts both prop­er­ties. The work­able middle is a small num­ber of ca­pa­cit­ies aligned to fund­ing bound­ar­ies, with a sep­ar­ate ca­pa­city for any­thing that must not be starved and an­other for de­vel­op­ment. Ca­pa­city place­ment also fixes the re­gion. The re­gion fixes the law that ap­plies, which is why chapter five be­longs be­fore the pro­vi­sion­ing de­cision rather than after it.

Do­main and work­space bound­ar­ies de­cide whether gov­ernance is en­force­able. A do­main holds the data of a busi­ness area and car­ries its own stew­ards and policy; a work­space holds the items a team builds. Where work­spaces are cre­ated per pro­ject, the lake ac­quires a shadow struc­ture nobody owns within a year. The rule that sur­vives is one work­space per team and li­fe­cycle stage, with the medal­lion lay­ers sep­ar­ated by work­space rather than by folder, so that per­mis­sions are gran­ted on a bound­ary the plat­form it­self re­cog­nises.

The ac­cess pat­tern for each source is the de­cision with the most money in it. A short­cut leaves the data where it is and reads it in place, which suits large volumes already held in a gov­erned lake. Mir­ror­ing brings an op­er­a­tional data­base into OneLake con­tinu­ously, which suits sys­tems that change fre­quently and quer­ies that can­not tol­er­ate a source out­age. A copy is jus­ti­fied on four con­di­tions only: the source can­not be reached dir­ectly, the con­tract for­bids dir­ect ac­cess, the source can­not carry the read load, or the trans­form­a­tion is heavy enough that re-read­ing the source is the greater cost. Ask of every pro­posed copy which con­di­tion it meets. A copy that meets none is the in­teg­ra­tion tax being re­built in­side the new plat­form.

Medal­lion own­er­ship de­cides whether the lay­ers mean any­thing. Bronze is owned by the in­ges­tion team and is never ed­ited. Sil­ver is owned by the do­main and car­ries a pub­lished con­tract, namely a schema, a stated re­fresh fre­quency and a qual­ity test that fails loudly. Gold is owned by the con­sumer and may be re­built at will. Pro­mo­tion between lay­ers needs a rule writ­ten in ad­vance, nam­ing who may pro­mote, which test must pass, and what a con­sumer is owed when a con­tract changes. Without that rule sil­ver be­comes a sta­ging area and gold be­comes the new ware­house.

Data & AnalyticsVoice of Volge07 •

Where these decisions fail in production

Each of those de­cisions fails in a re­cog­nis­able way, and the fail­ure shows up in pro­duc­tion rather than in re­view. A single shared ca­pa­city fails as in­ter­mit­tent re­port timeouts dur­ing the close, re­por­ted by the busi­ness as a re­port­ing prob­lem and dia­gnosed weeks later as a note­book. Work­space sprawl fails as a re­quest for a re­port nobody can trace to a source. A copy taken for con­veni­ence fails as two num­bers in two decks, which is the trust erosion of this chapter ar­riv­ing in­side the con­sol­id­ated plat­form.

Two fur­ther fail­ures are worth nam­ing be­cause they are quiet. An un­owned sil­ver layer fails slowly. Con­sumers build dir­ectly on bronze to avoid wait­ing for the do­main, and within two quar­ters the lake has a second, un­doc­u­mented serving layer. A ca­pa­city sized from a vendor table rather than from ob­served con­sump­tion fails at the first month-end, when a heavy batch load and an in­ter­act­ive peak ar­rive in the same hour. Neither fail­ure pro­duces an in­cid­ent. Both pro­duce a land­scape that looks con­sol­id­ated and be­haves frag­men­ted.

, The economics and the migration, and where Fabric does not fit

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 and Direct Lake, which is why it 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 landscapes is hybrid.

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 and the lake­house, the pipelines and 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 and one meter.

The SKU range is wide. Mi­crosoft lists thir­teen F-SKUs, run­ning from F2 to F8192. On pay-as-you-go billing, at about 0.18 United States dol­lars per ca­pa­city unit per hour for ca­pa­city

left run­ning con­tinu­ously, that is roughly 263 dol­lars a month at F2 and roughly 1,076,000 dol­lars a month at F8192. A one-year or three-year re­ser­va­tion saves about 41 per cent against pay-as-you-go, and en­ter­prise-agree­ment rates are typ­ic­ally below list (Mi­crosoft Fab­ric pri­cing, veri­fied 8 Novem­ber 2026). Pri­cing changes and is re­gion- and con­tract-spe­cific, and the fig­ures here should be re­con­firmed against cur­rent Mi­crosoft pri­cing be­fore any com­mit­ment.

F-SKUApprox. capacity unitsApprox. monthly list, pay-as-you-goTypical fit
F22~USD 263Development, small workloads, evaluation
F88~USD 1,051Departmental analytics
F6464~USD 8,410Inflection: Power BI Premium P1 equivalent + Direct Lake; common enterprise start
F128128~USD 16,819Large enterprise; common Fortune 500 starting band with F64
F256256~USD 33,638Heavy, broad landscape
F20482048~USD 269,107Very large, multi-domain landscape
F81928192~USD 1,076,429Top of the published range

Every row is a pay-as-you-go list price for ca­pa­city run­ning con­tinu­ously, at about USD 0.18 per ca­pa­city unit per hour over a 730-hour month (Mi­crosoft Fab­ric pri­cing, veri­fied 8 Novem­ber 2026). A one-year or three-year re­ser­va­tion saves about 41 per cent against these fig­ures, which puts F64 re­served near USD 4,960 a month. Pri­cing is re­gion- and con­tract-spe­cific and changes; re­con­firm against cur­rent Mi­crosoft pri­cing.

Why F64 is the inflection

The F64 tier, at roughly 8,410 United States dol­lars per month on pay-as-you-go and nearer 4,960 on a one-year re­ser­va­tion, is where the eco­nom­ics turn. The reason is spe­cific. F64 is the Power BI Premium P1 equi­val­ent, which is the tier at which Fab­ric in­cludes Power BI Premium-equi­val­ent cap­ab­il­it­ies and makes Dir­ect Lake avail­able across the work­loads (Mi­crosoft Fab­ric pri­cing, veri­fied 8 Novem­ber 2026). Two con­sequences fol­low.

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 land­scape re­places.

Data & AnalyticsVoice of Volge08 •

First, re­port con­sumers no longer need in­di­vidual Power BI Pro li­cences. At F64 and above, view­ers con­sume re­ports on free li­cences. For an en­ter­prise with thou­sands of re­port con­sumers that can dom­in­ate the total-cost com­par­ison on its own, be­cause the per-seat Pro li­cens­ing that a stan­dalone Power BI de­ploy­ment car­ries col­lapses into the ca­pa­city. Second, Dir­ect Lake be­comes avail­able, which 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 on pay-as-you-go (Microsoft Fabric pricing, verified 8 November 2026) - ~USD 263/month, F2, the entry tier - ~USD 8,410/month, F64, the inflection: the Power BI Premium P1 equivalent, which unlocks Direct Lake - ~USD 1,076,000/month, F8192, the top of the published range - ~41 per cent, the saving a one-year or three-year reservation makes against pay-as-you-go - Free licences, report viewers consume at F64 and above Pricing is region- and contract-specific and changes; reconfirm against current Microsoft pricing.

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 land­scape re­moves whole cat­egor­ies of cost that the frag­men­ted land­scape car­ries struc­tur­ally. The table sets the two side by side.

Cost componentFragmented multi-service landscapeConsolidated on Fabric
Compute licensingSeparate meter per service, each sized to its own peakOne pooled capacity meter shared across workloads
StorageDuplicated copies per serviceOne copy in OneLake; copies reclaimed
BI licensingPer-seat Pro for every consumerFree viewers at F64+; licensing in the capacity
Pipeline operationCopy pipelines between every serviceCopy pipelines removed; data read in place
Idle reserved capacityRedundant across non-coincident peaksSingle pooled sizing decision
Governance toolingSeparate catalogue and policy per serviceOneLake Catalog and Purview over one landscape
Operational headcountSpecialists per service and per integrationOne operational surface and skill base

The sav­ing con­cen­trates in the rows that de­scribe the seams. The seams are du­plic­ated stor­age and copy pipelines, re­dund­ant ca­pa­city and 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. That is why ca­pa­city siz­ing and the gov­ernance of ca­pa­city con­sump­tion are where con­sol­id­a­tion pro­grammes are won or lost in prac­tice.

The migration

The land­scape Fab­ric con­sol­id­ates is, for most en­ter­prises, an Azure land­scape 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 was re­tired in fa­vour of the Fab­ric Event­house on 7 Oc­to­ber 2025, and mi­gra­tion tool­ing , as­sess­ment-first pipeline and ware­house mi­gra­tion as­sist­ants , has shipped to sup­port the move (Mi­crosoft Learn, 2025-2026). No re­tire­ment date ex­ists for Azure Data Fact­ory either, and the only dated item there is the re­tire­ment of its Leg­acy Trus­ted Ser­vice mode in Au­gust 2027. 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 land­scape, 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.

Data & AnalyticsVoice of Volge09 •

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 land­scapes.

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.

Running a capacity, and what to do when it throttles

Ca­pa­city is sized from ob­served con­sump­tion, not from a table of typ­ical fits. The in­puts are the con­cur­rency at the peak hour rather than the mean, the dur­a­tion of the heav­iest sched­uled job, the re­fresh pro­file of the se­mantic mod­els the busi­ness ac­tu­ally opens, and the head­room a close or a quarter-end needs. Each is meas­ur­able in the land­scape being re­placed. A siz­ing that can­not show those four in­puts is a guess wear­ing a SKU num­ber, and it errs in the vendor's dir­ec­tion. The dis­cip­line that works is simple. Size for the ob­served peak, run a full busi­ness cycle, then res­ize on evid­ence rather than on com­plaint.

When a ca­pa­city runs out of units the symp­tom reaches the busi­ness be­fore it reaches the plat­form team. In­ter­act­ive re­ports slow and then queue. Fail­ures fol­low, and the first re­port is a com­plaint about a dash­board. The re­sponse has an order. Identify the con­sumer that moved, not the re­port that broke. Move the of­fend­ing work­load to its own ca­pa­city or its own sched­ule be­fore buy­ing a lar­ger SKU. An over­sized ca­pa­city hides the be­ha­viour that caused the prob­lem and then charges for it every month. Raise the SKU where the load is genu­ine, sus­tained and at­trib­ut­able to a fun­ded out­come.

The par­al­lel run is the largest single cost of the mi­gra­tion and the one most often left out of the busi­ness case. For the length of each stage the en­ter­prise pays the leg­acy meter and the Fab­ric ca­pa­city at once, plus the en­gin­eer­ing to keep two pipelines pro­du­cing the same an­swer. That cost is bounded only by the length of the stage, which is why a stage with no exit test runs in­def­in­itely and con­sumes the sav­ing it was meant to prove. Budget the over­lap ex­pli­citly, stage by stage, and treat an over­run as a pro­gramme issue rather than as an op­er­at­ing cost.

A roll­back path be­longs to each stage and ex­pires only when the stage closes. In prac­tice the leg­acy ser­vice stays run­ning

and reach­able, its data con­tin­ues to re­fresh, and the switch back is tested once rather than as­sumed. The test is cheap while both land­scapes exist. It is im­possible af­ter­wards, which is why the order of op­er­a­tions in this chapter never changes. Mi­grate, val­id­ate at pro­duc­tion load, then de­com­mis­sion.

Each stage closes on a test rather than on a date. The tests below are the min­imum, and each is passed against pro­duc­tion load rather than against a sample.

  • As­sess and land. The baseline of chapter one is re­cor­ded with its raw ex­tracts, the OneLake found­a­tion is pro­vi­sioned in a re­gion that matches the res­id­ency po­s­i­tion for each data do­main, and one non-crit­ical data­set is serving a real re­port through the medal­lion lay­ers.
  • Mi­grate the ware­house. The mi­grated ware­house an­swers the ten heav­iest pro­duc­tion quer­ies within the latency the busi­ness already has, and a full re­con­cili­ation against the leg­acy ware­house ties out for two con­sec­ut­ive closes.
  • Move pipelines and busi­ness in­tel­li­gence. Every mi­grated pipeline has run un­at­ten­ded for a full cycle, the se­mantic mod­els carry the same totals as their pre­de­cessors, and re­port con­sumers have been moved rather than mirrored.
  • Make cur­rent and de­com­mis­sion. The leg­acy meter is switched off, the sav­ing ap­pears in an in­voice rather than in a pro­jec­tion, and the roll­back path has been re­tired on a re­cor­ded date.

Re­con­cili­ation dur­ing the par­al­lel run is the part teams un­der­es­tim­ate, be­cause prov­ing that two ware­houses agree is harder than mi­grat­ing one. The test is not a row count. It is a set of named meas­ures, each at the grain the busi­ness re­ports it, com­pared across both land­scapes for the same period, with a stated tol­er­ance and a doc­u­mented ex­plan­a­tion for every dif­fer­ence in­side it. Dif­fer­ences will ap­pear, and most will be le­git­im­ate: a late-ar­riv­ing re­cord, a dif­fer­ent time zone at the bound­ary of a day, a round­ing con­ven­tion car­ried in a leg­acy view. Each has to be iden­ti­fied once and then sup­pressed de­lib­er­ately, rather than re-in­vest­ig­ated every cycle. Two con­sec­ut­ive closes that tie out on that basis is the evid­ence a fin­ance func­tion will ac­cept. One close that ties out is a co­in­cid­ence until the second con­firms it. Budget the work ex­pli­citly, as en­gin­eer­ing days per meas­ure rather than as a con­tin­gency, and name the per­son in fin­ance who signs it off.

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.

Data & AnalyticsVoice of Volge10 •

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 land­scapes choose it (in­dustry ana­lysis, 2026). Its scale re­flects that po­s­i­tion: a rev­enue run-rate above 7 bil­lion United States dol­lars, grow­ing more than 80 per cent, with more than 1,000 cus­tom­ers con­sum­ing above one mil­lion dol­lars a year (Dat­ab­ricks, Au­gust 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: 14,554 cus­tom­ers, with 828 of them above one mil­lion dol­lars in trail­ing-twelve-month product rev­enue, in the quarter ended 31 July 2026 (Snow­flake earn­ings, Q2 FY2027, 2 Septem­ber 2026). 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 axisChoose FabricChoose DatabricksChoose Snowflake
Primary goalConsolidate a fragmented landscape; cut integration taxLarge-scale data engineering and MLHigh-concurrency SQL; data sharing
EcosystemMicrosoft, Azure, Power BI landscapeMulti-cloud, open, code-firstMulti-cloud, SQL-first
StrengthUnification; BI integration; SaaS simplicitySpark depth; ML maturity; open governanceConcurrency; elastic scaling; sharing
Team profileMixed, BI-led, low-to-pro-codeEngineering- and ML-heavySQL- and analyst-heavy
Honest limitSpecialist engineering and warehousing ceilingLess unified; heavier to operateNarrower than a full unified landscape

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 land­scape. Fab­ric's bi-dir­ec­tional, zero-copy in­ter­op­er­ab­il­ity with both Dat­ab­ricks and Snow­flake means an en­ter­prise can con­sol­id­ate the frag­men­ted ma­jor­ity of its land­scape 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 land­scape is fre­quently hy­brid, and the in­ter­op­er­ab­il­ity is what makes hy­brid co­her­ent rather than a re­turn to the copy sprawl con­sol­id­a­tion set out to re­move.

, Reference architecture, controls and operations

A durable consolidated landscape 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 architecture is the easy part. What makes it defensible is a control with a test behind it, a named failure mode with a detection, a measure taken from a baseline, and an exit criterion a steering committee can refuse to waive.

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 land­scape; everything else is siz­ing and se­quen­cing.

Prerequisites, and what fails without them

The ar­chi­tec­ture above is not build­able until five things exist, and none of them is a Fab­ric fea­ture. They are a clas­si­fic­a­tion scheme with an owner, an ac­cess model ex­pressed in groups rather than in in­di­vidu­als, a named fun­ded owner for each data do­main, a source-sys­tem con­tract for every sys­tem in scope, and a ca­pa­city budget some­body con­trols. A pro­gramme that starts without them does not stall. It pro­ceeds, and ab­sorbs each miss­ing de­cision as a de­fault nobody chose.

The order mat­ters as much as the list. Clas­si­fic­a­tion and the ac­cess model gate everything that fol­lows, be­cause they de­cide what may be placed where and who may read it. The source con­tracts gate the pipelines. The ca­pa­city budget gates the siz­ing. A pro­gramme that can name an owner

Data & AnalyticsVoice of Volge11 •

against each of the five is ready to build. One that can­not is ready to write a design doc­u­ment.

  • A clas­si­fic­a­tion scheme with an owner. Classes named, with a defin­i­tion for each and a named owner who de­cides bor­der­line cases within days rather than at the next com­mit­tee. Without it, clas­si­fic­a­tion be­comes a free-text field and the res­id­ency re­gister of chapter five has noth­ing to at­tach to.
  • An ac­cess model in groups. En­ti­tle­ments ex­pressed as groups mapped to roles, pro­vi­sioned from the iden­tity land­scape and re­viewed on a cycle. Without it, per­mis­sions ac­cu­mu­late item by item and per­son by per­son, and the ac­cess con­trol the ap­plic­able rules re­quire can­not be evid­enced at all.
  • A fun­ded owner for each data do­main. One ac­count­able name, with a budget, who can re­fuse a con­sumer. Without it the do­main layer has no stew­ard, and the lake grows a serving layer nobody pub­lished.
  • A source-sys­tem con­tract. For every source in scope, the agreed ex­trac­tion pat­tern, the re­fresh win­dow, the no­tice period for a schema change, and who is called when the source changes. Without it, pipeline drift is dis­covered by a con­sumer.
  • A ca­pa­city budget with a con­trol­ler. A stated monthly ceil­ing, an owner who ap­proves a change of SKU, and vis­ib­il­ity of con­sump­tion by con­sumer. Without it, the con­sol­id­a­tion sav­ing is spent on a lar­ger ca­pa­city to hide a be­ha­viour.

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 land­scapes in par­al­lel. The ob­ject­ive is the core re­port­ing land­scape read­ing one copy.

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.

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 land­scape con­sol­id­ated, not just the core.

A data-platform maturity model

The model loc­ates the en­ter­prise and defines its next in­cre­ment. It is the dia­gnostic coun­ter­part to the roadmap.

LevelDescriptionLandscape and storageGovernance and capacityTypical outcome
0 , FragmentedSeparately licensed services; copies everywhereMany stores; duplicated copiesPer-service; bolted-on catalogueFull integration tax; trust erosion
1 , LandedOneLake stood up; first sources connectedOne lake emerging beside legacyPurview started; capacity sizedFoundation set; saving not yet realised
2 , Core consolidatedWarehouse and core BI on Fabric; Direct Lake liveOne copy for core; legacy in parallelLineage and policy active; capacity tunedCopies and import refreshes retired for the core
3 , Landscape consolidatedPipelines, science, real-time on platformOne copy across the landscapePolicy follows data; capacity governedMost integration tax removed
4 , Optimised unified landscapeLegacy decommissioned; hybrid only by designOne copy; best-of-breed by deliberate choiceLandscape-wide governance; capacity optimisedSaving realised; AI-ready by-product

On Volge's own as­sess­ment, rather than on any pub­lished sur­vey, most en­ter­prises sit between Level 0 and Level 1 as at late 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.

The before-and-after, summarised

The con­sol­id­a­tion reads most clearly as a shift in the con­di­tions under which the land­scape op­er­ates.

Data & AnalyticsVoice of Volge12 •
ConditionFragmented landscapeConsolidated on Fabric
Data locationCopies across many servicesOne copy in OneLake
Storage formatProprietary per serviceOpen Delta Parquet, read by all
BI deliveryImport copies and scheduled refreshDirect Lake on one copy
Compute meterOne per serviceOne pooled capacity
GovernanceBolted-on, per serviceOneLake Catalog and Purview as a layer across the landscape
Real-timeSeparate streaming stackReal-Time Intelligence on-platform
Operational surfaceSeveral teams and skill setsOne platform, one skill base

The controls that make a single copy defensible

A single copy of the data is only de­fens­ible if some­body can prove it. The con­trols below are the min­imum set for a con­sol­id­ated land­scape, and each is writ­ten as some­thing test­able rather than as a policy state­ment. The test mat­ters more than the con­trol. A clas­si­fic­a­tion scheme with no sampling test is a spread­sheet. A lin­eage claim nobody has traced end to end is a dia­gram. Each con­trol also pro­duces evid­ence, which is what an aud­itor, a reg­u­lator or an in­com­ing chief data of­ficer asks for. Evid­ence is what sep­ar­ates a gov­erned land­scape from a tidy one.

ControlWhat it enforcesHow you test itEvidence it produces
Single authoritative copyOne owning workspace per dataset, with every other instance declared a shortcut, a mirror or a retired copyReconcile the copy register against the lake's own item inventory each quarter, and query two named consumers for the same totalA copy register with an owner against every entry, and a dated reconciliation
Data domain residencyEach domain placed in a region permitted for its class, with processing stated separately from storageSample five datasets per domain and trace the region of the capacity, the region of the source and any support access from outside the geographyA versioned residency register, one line per data domain
Classification at ingestionEvery dataset carries a classification before it reaches silverAttempt to promote an unclassified dataset and confirm that the promotion failsA failed-promotion log
Access on a boundaryPermissions granted on the workspace and domain boundary rather than item by itemPull the item-level permission exceptions and require a named justification for eachAn exception list, shrinking between reviews
ControlWhat it enforcesHow you test itEvidence it produces
Continuous lineageEvery published report traces to its source without manual annotationTake the two reports the board uses most and walk the lineage back to the source systemTwo traced lineage paths, dated
Capacity governanceConsumption attributable to a funded owner, with no unbounded workloadReplay the heaviest week and confirm that every consumer is attributableA consumption report by owner
Retention and erasureRetention per data class, including any statutory floorRun an erasure request end to end on a test subject and confirm that each medallion layer and the backups honour itA completed erasure drill record
Log integrityAccess and change logs retained for the period the applicable rules require, in a permitted locationRequest a log extract for a named dataset and a named hour the way a supervisor would, and time itA dated log extract and the elapsed time to produce it

Two of those con­trols are com­monly skipped, and both are the ones that fail an ex­am­in­a­tion. The eras­ure drill is skipped be­cause it is awk­ward, since an eras­ure that reaches bronze, sil­ver, gold and the backup cop­ies touches every layer the medal­lion pat­tern cre­ated. The log ex­tract is skipped be­cause it looks trivial until it is timed. Where the ap­plic­able rules re­quire pro­duc­tion within a stated period, the con­trol is the elapsed time and not the ex­ist­ence of the log. Source: Volge ana­lysis, 2026, writ­ten against the in­stru­ments set out in chapter five.

Failure modes, and what each looks like in production

A con­trol list is only half the work, be­cause the fail­ures that mat­ter an­nounce them­selves in the busi­ness rather than on the plat­form. The table below names the ones seen most often in con­sol­id­a­tion pro­grammes, with the de­tec­tion that catches each one early enough to be cheap. Read the second column first. It is what some­body will re­port, and the point of the ex­er­cise is to re­cog­nise the cause from the symp­tom without a month of in­vest­ig­a­tion.

One fail­ure de­serves sep­ar­ate treat­ment be­cause it is in­vis­ible until it is ex­pens­ive. Res­id­ency drift hap­pens through or­din­ary pro­vi­sion­ing. A team needs a ca­pa­city, the portal of­fers a re­gion, and a reg­u­lated do­main ac­quires a new loc­a­tion without any­body de­cid­ing. The con­trol is to re­strict pro­vi­sion­ing to the re­gions already ap­proved for that do­main, and to make the ex­cep­tion route amend the re­gister rather than

Data & AnalyticsVoice of Volge13 •
Failure modeWhat it looks like in productionHow it is detectedWhat closes it
Copy sprawl inside the platformTwo Fabric items report different totals for the same measure, and each has a defenderScheduled comparison of named measures across items, run as a job rather than as an auditA copy register, and deletion of the loser once its consumers have been moved
Capacity starvation at the closeInteractive reports queue or fail in the last three days of the monthPeak-hour consumption by consumer, held against the close calendarWorkload separation or rescheduling, taken before a larger SKU is bought
Unowned silverConsumers build on bronze to avoid waiting for the domainCount of gold items reading directly from bronzeA published contract for every silver table, with a named steward
Semantic models sized on a sampleA model performs well against test volumes and slowly against production onesQuery duration distribution per model, compared with its predecessorModel or table redesign, taken before the report is published
Residency driftA capacity or workspace is created in a convenient region and a regulated domain follows itRegion of every capacity and workspace reviewed against the residency registerProvisioning restricted to approved regions, with an exception route that changes the register
Decommissioning before validationA legacy service is switched off and a gap appears at the next close, with no fallbackThe stage exit tests, which is their only purposeReinstatement where it is still possible, and a written stage gate where it is not
Pipeline driftA migrated pipeline succeeds while producing a stale or partial resultFreshness and row-count assertions on the target, not pipeline statusAssertions promoted to blocking, with a named owner paged

by­pass it. Chapter five sets out why the re­gister is kept per data do­main and not per ten­ant.

The pat­tern across that list is that the symp­tom and the cause sit in dif­fer­ent teams. A busi­ness user re­ports a slow re­port and the cause is a note­book in an­other work­space on the same ca­pa­city. A fin­ance ana­lyst re­ports a wrong num­ber and the cause is a copy taken for con­veni­ence eight­een months earlier. De­tec­tion there­fore can­not sit with the team that feels the symp­tom. It has to be in­stru­mented on the plat­form, against named meas­ures and named con­sumers, and re­viewed on a ca­dence some­body owns.

What to measure

Meas­ure the con­di­tions the con­sol­id­a­tion was meant to change, against a baseline taken be­fore it star­ted. Six

meas­ures carry the case. Each has a defin­i­tion that fits in a sen­tence, a source in­side the land­scape, and a dis­cip­line about the baseline that mat­ters more than the tar­get does.

  • Copy ratio. The gross stored foot­print of a data­set's in­stances over the foot­print of its au­thor­it­at­ive copy, read monthly from stor­age meter­ing. The baseline is taken be­fore mi­gra­tion and held fixed. A ratio that falls be­cause a source was re­tired, rather than be­cause a copy was re­moved, is re­por­ted sep­ar­ately.
  • Pipelines in ser­vice. The count of pipelines whose only pur­pose is to move data between two ser­vices, read from the or­ches­tra­tion in­vent­ory. The dis­cip­line is that a pipeline re­built in­side Fab­ric still counts. Con­sol­id­a­tion that moves the toil rather than re­mov­ing it shows up here and nowhere else.
  • Ca­pa­city util­isa­tion at the peak hour. Con­sumed units over re­served units in the busiest hour of the busi­ness cycle, not the mean. Two baselines are needed, one for the close and one for an or­din­ary week, be­cause an av­er­age across both hides the only hour that mat­ters.
  • Time to a trus­ted num­ber. Elapsed days from a re­quest for a new re­por­ted meas­ure to its pub­lic­a­tion with lin­eage and clas­si­fic­a­tion com­plete. Meas­ured from the re­quest and not from the point the work was sched­uled, which is where this fig­ure is usu­ally flattered.
  • Re­con­cili­ation hours per close. Hours spent re­con­cil­ing two in­ternal cop­ies of the same re­cord, taken from time re­cords and sep­ar­ated from the re­con­cili­ation a reg­u­lator re­quires. Only the first is ex­pec­ted to fall.
  • Res­id­ency ex­cep­tions open. The count of data do­mains whose ac­tual loc­a­tion dif­fers from the po­s­i­tion re­cor­ded in the re­gister, with the age of each ex­cep­tion. The tar­get is zero and the num­ber that mat­ters is the age.

What is de­lib­er­ately ab­sent here is a single ma­tur­ity score. A score ag­greg­ates these six into a num­ber nobody can act on, and it im­proves when the easi­est of them im­proves. The ma­tur­ity model above does a dif­fer­ent job, which is to loc­ate an en­ter­prise and name its next in­cre­ment. Keep the six meas­ures sep­ar­ate, dated and owned by name. Source: Volge ana­lysis, 2026.

What done looks like

Done is a set of con­di­tions, and a pro­gramme that can­not state them is not se­quen­cing a mi­gra­tion but run­ning a plat­form pro­ject. The con­di­tions below are ex­pressed per stage so that a steer­ing com­mit­tee can re­fuse to close a stage.

Data & AnalyticsVoice of Volge14 •

Each is ob­serv­able by some­body out­side the pro­gramme, which is the only prop­erty that makes an exit cri­terion use­ful.

  • Found­a­tion done. The baseline of chapter one is re­cor­ded with its raw ex­tracts and a named owner against each fig­ure; the res­id­ency re­gister ex­ists per data do­main, with the pro­cessing po­s­i­tion stated sep­ar­ately from the stor­age po­s­i­tion; ca­pa­city is pro­vi­sioned only in ap­proved re­gions; and one non-crit­ical data­set serves a pro­duc­tion re­port through the medal­lion lay­ers.
  • Core con­sol­id­ated. The mi­grated ware­house an­swers the heav­iest pro­duc­tion quer­ies within the latency the busi­ness already had; two con­sec­ut­ive closes re­con­cile against the leg­acy ware­house; clas­si­fic­a­tion blocks the pro­mo­tion of an un­clas­si­fied data­set; and the copy re­gister names an owner for every in­stance of every data­set in scope.
  • Cur­rent and gov­erned. Real-time work­loads serve the de­cisions that need re­cency, with the re­cency re­quire­ment writ­ten down per de­cision; lin­eage traces the board's own re­ports to source without manual an­nota­tion; the eras­ure drill has been run end to end; and a log ex­tract has been pro­duced in­side the period the ap­plic­able rules re­quire.
  • Sav­ing real­ised. The leg­acy meters are switched off, the re­duc­tion ap­pears in an in­voice rather than in a pro­jec­tion, the roll­back path has been re­tired on a re­cor­ded date, and ca­pa­city siz­ing has been re­vised once against ob­served con­sump­tion rather than against the ori­ginal es­tim­ate.

The cadence that keeps it true

A gov­erned land­scape de­cays un­less some­thing de­cides on a rhythm, and the rhythm costs less than the re­medi­ation. Four for­ums carry the work, and each ex­ists to take a named de­cision rather than to re­view a re­port. Keep them short and keep the de­cision rights ex­pli­cit, be­cause a forum with no de­cision right be­comes a status meet­ing and is then can­celled for being one.

  • Weekly, the plat­form op­er­a­tions re­view. Ca­pa­city con­sump­tion by con­sumer against the budget, failed and drif­ted pipelines, and the ex­cep­tions opened in the week. De­cision right: res­ched­ule or isol­ate a work­load, and raise a change where the budget is threatened.
  • Monthly, the do­main own­ers' forum. New data­sets and their clas­si­fic­a­tion, pro­mo­tion re­quests between medal­lion lay­ers, and con­tract changes a con­sumer will feel. De­cision right: ap­prove or re­fuse a pro­mo­tion, and set the no­tice a con­sumer is owed.
  • Quarterly, the con­trol test cycle. The copy re­gister re­con­cili­ation, the sampled res­id­ency trace, the eras­ure drill and the timed log ex­tract. De­cision right: ac­cept the evid­ence, or re­quire re­medi­ation with a date against it.
  • At each stage bound­ary, the mi­gra­tion gate. The exit tests of chapter three, read against the stage's own defin­i­tion of done. De­cision right: close the stage, ex­tend it with a named reason, or re­verse it.

The evid­ence each con­trol pro­duces needs a home and a re­ten­tion period, and both are de­cisions rather than de­faults. Keep the re­gisters, the drill re­cords and the dated ex­tracts where an aud­itor can reach them without the plat­form team's help. Evid­ence only one team can pro­duce is evid­ence a su­per­visor will dis­count. Re­ten­tion fol­lows the longest ap­plic­able re­quire­ment rather than the most con­veni­ent one. In India it is a floor rather than a ceil­ing, since rules 6(1)(e) and 8(3) of the Di­gital Per­sonal Data Pro­tec­tion Rules, 2025 re­quire a min­imum of one year for per­sonal data, as­so­ci­ated traffic data and logs of pro­cessing, once that tranche com­mences on 13 May 2027.

The ar­chi­tec­ture names the tar­get, the roadmap se­quences the jour­ney, and the con­trols de­cide whether the res­ult can be de­fen­ded. Fol­low the se­quence and the land­scape 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 land­scapes at once. Skip the con­trol tests and it pays for one land­scape it can­not evid­ence.

, The data platform in the Gulf and India

The platform is not available everywhere its infrastructure is. Qatar has an Azure region, a Microsoft 365 residency commitment and a Foundry presence, and it cannot run the Fabric workloads this paper is about. Saudi Arabia has no region and no residency commitment of any kind, against a law that forbids exporting government agency data. India's rules attach to classes of data rather than to tenants, which makes residency a question about domains and not about a tenant. Classification comes before placement, and placement before provisioning.

The plat­form this paper ar­gues for is not avail­able in every coun­try that can run it. Qatar Cent­ral is the Azure re­gion at Doha. It has three avail­ab­il­ity zones and a Mi­crosoft 365 Local Re­gion Geo­graphy be­hind it, and for Fab­ric pur­poses it is a Power BI only re­gion (Mi­crosoft Learn, Fab­ric re­gion avail­ab­il­ity, page up­dated 5 Novem­ber 2026). The lake­house can­not run there. Neither can the ware­house, data en­gin­eer­ing or Real-Time In­tel­li­gence. Mi­crosoft's own rem­edy is stated

Data & AnalyticsVoice of Volge15 •

on the same page, which is to cre­ate a ca­pa­city in a re­gion where Fab­ric is avail­able. For a Qatari en­ter­prise that means leav­ing the coun­try.

That single fact is the right way into the re­gional pic­ture. It sep­ar­ates two ques­tions a con­sol­id­a­tion pro­gramme usu­ally merges. In­fra­struc­ture pres­ence is not plat­form avail­ab­il­ity, and plat­form avail­ab­il­ity is not a res­id­ency com­mit­ment. A coun­try may hold all three, two of the three, or none. Four Gulf states hold none. The order of work changes as a res­ult. Clas­si­fic­a­tion comes be­fore place­ment, and place­ment comes be­fore pro­vi­sion­ing. The con­sol­id­a­tion eco­nom­ics of chapter three are then cal­cu­lated in­side whatever those two allow.

One vendor, several footprints, one region

Read Mi­crosoft's own avail­ab­il­ity pages against each other and the Gulf pic­ture is un­even in a way the policy rhet­oric ob­scures. UAE North is the only Gulf re­gion car­ry­ing the full Fab­ric work­load set. UAE Cent­ral, at Abu Dhabi, is Power BI only and has no avail­ab­il­ity zones. It is also a re­stric­ted-ac­cess re­gion, to which a sub­scrip­tion must re­quest ac­cess through the Azure re­gion ac­cess re­quest pro­cess (Mi­crosoft Learn, Azure re­gions list, page up­dated 13 Oc­to­ber 2026). India runs the full set at Cent­ral India and South India. India West is Power BI only.

The sharpest item in the com­par­ison is an ab­sence. India South Cent­ral, the Hy­dera­bad re­gion, ap­pears in the Azure re­gions list with three avail­ab­il­ity zones and Cent­ral India as its pair. It does not ap­pear in the Fab­ric re­gion avail­ab­il­ity table at all (Mi­crosoft Learn, pages up­dated 13 Oc­to­ber and 5 Novem­ber 2026). A new Azure re­gion does not bring Fab­ric with it. An ar­chi­tec­ture that as­sumed oth­er­wise has planned a ca­pa­city where the plat­form is not doc­u­mented as avail­able. Settle that with the vendor in writ­ing, not in a design re­view.

Qatar Cent­ral car­ries a second con­straint, and a lake­house with a re­cov­ery ob­lig­a­tion feels it. It is an un­paired re­gion, as are Is­rael Cent­ral, Italy North, Po­land Cent­ral and Spain Cent­ral, so the Azure con­structs that de­pend on a re­gion pair do not apply to it (Mi­crosoft Learn, page up­dated 13 Oc­to­ber 2026). Re­cov­ery is there­fore de­signed ex­pli­citly, across re­gions the or­gan­isa­tion chooses for it­self. That choice re­opens the res­id­ency ques­tion. The in-coun­try re­gion was sup­posed to close it.

JurisdictionAzure regionFabric workloads in regionMicrosoft 365 durable commitmentCopilot Studio in region
Saudi ArabiaNoneNoneNoneNo
United Arab EmiratesTwo, UAE Central restrictedUAE North only; UAE Central is Power BI onlyProduct Terms, Multi-Geo and Advanced Data ResidencyYes, UAE North
QatarOne, unpairedNone; Power BI onlyProduct Terms, Multi-Geo and Advanced Data ResidencyNo; data stored in the United States
IndiaFour, two with availability zonesCentral India and South India; India West is Power BI onlyProduct Terms, Multi-Geo and Advanced Data ResidencyYes, two regions
Kuwait, Oman, BahrainNoneNoneNoneNo

Sources: Mi­crosoft Learn re­gion and data-loc­a­tion doc­u­ment­a­tion, pages up­dated between 8 Oc­to­ber and 5 Novem­ber 2026, veri­fied 8 Novem­ber 2026. Com­pil­a­tion by Volge. On the last column, Mi­crosoft states that where a ten­ant's loc­a­tion is not lis­ted in the Co­pi­lot Stu­dio data loc­a­tions table, data is stored in the United States (Mi­crosoft Learn, page up­dated 30 April 2026).

Saudi Arabia, where the law forbids export and the vendor commits to nothing

Saudi Ar­a­bia com­bines the re­gion's most elab­or­ate loc­al­isa­tion ap­par­atus with the thin­nest vendor foot­print of any large Gulf mar­ket. Mi­crosoft's pub­lished Azure re­gions list car­ries no Saudi entry under any name, live, re­stric­ted or in pre­view, the Middle East entries being Qatar Cent­ral, UAE Cent­ral, UAE North and Is­rael Cent­ral (Mi­crosoft Learn, page up­dated 13 Oc­to­ber 2026). The King­dom is ab­sent from Mi­crosoft 365 data res­id­ency as well. It is not a Local Re­gion Geo­graphy and is not eli­gible for the Ad­vanced Data Res­id­ency add-on. It does not ap­pear in the table of dur­able com­mit­ments on data loc­a­tion (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026).

What that ab­sence means is stated by Mi­crosoft rather than in­ferred. A Saudi ten­ant falls into the cat­egory Mi­crosoft calls All Other Geo­graph­ies. There, cus­tomer data "isn't com­mit­ted to reside in any par­tic­u­lar data­cen­ter", and Mi­crosoft "stores data in a loc­a­tion that en­ables ef­fect­ive ser­vice de­liv­ery, and that loc­a­tion is sub­ject to change without no­tice" (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). Kuwait, Oman and Bahrain sit in the same cat­egory. Qatar, the United Arab Emir­ates and India do not.

Set that against Saudi law and the col­li­sion is com­plete. Art­icle 3-3-5 of the Cloud Com­put­ing Ser­vices Pro­vi­sion­ing

Data & AnalyticsVoice of Volge16 •

Reg­u­la­tions ver­sion 4 pro­hib­its a pro­vider or a sub­scriber from trans­fer­ring con­tent com­pris­ing data of Saudi gov­ern­ment agen­cies out­side the King­dom. The pro­hib­i­tion cov­ers any pur­pose and any form, per­man­ently or tem­por­ar­ily, in­clud­ing tem­por­ary stor­age and backup, un­less an­other Saudi in­stru­ment ex­pressly per­mits it. The Reg­u­la­tions were is­sued by the Com­mu­nic­a­tions, Space and Tech­no­logy Com­mis­sion under De­cision No. 506/1445 and took ef­fect on 10 Oc­to­ber 2023. The art­icle num­ber and the word­ing rest on named law-firm com­ment­ary, Ald­habaan Law Firm writ­ing in the Cham­bers Cloud Com­put­ing 2026 guide, be­cause the CST file it­self could not be read. The law for­bids ex­port of that data. The vendor of­fers no com­mit­ment to avoid it.

Three fur­ther Saudi facts shape the ar­chi­tec­ture. A pro­vider serving any Saudi gov­ern­ment cus­tomer must hold CST re­gis­tra­tion at Class A or above. Class A cov­ers pub­lic gov­ern­ment data only. Con­fid­en­tial re­quires Class B, and secret or top secret re­quires Class C, per CST's own re­gister. For a fin­an­cial in­sti­tu­tion the bind­ing con­straint is an ap­proval rather than a geo­graphy. The Saudi Cent­ral Bank's Cyber Se­cur­ity Frame­work, No. 381000091275 of 24 May 2017 and in force, states at 3.4.3 that in prin­ciple only cloud ser­vices loc­ated in Saudi Ar­a­bia should be used, with ex­pli­cit SAMA ap­proval re­quired where they are not. It also re­quires SAMA ap­proval be­fore cloud ser­vices are used at all. SAMA's Rules on Out­sourcing, No. 41027017 of 15 Decem­ber 2019, add a writ­ten no-ob­jec­tion for over­seas out­sourcing, with a ra­tionale for why the work can­not be done within the King­dom.

One widely re­peated claim should be dropped. The Na­tional Data Man­age­ment Of­fice's Data Man­age­ment and Per­sonal Data Pro­tec­tion Stand­ards, ver­sion 1.5 of Janu­ary 2021, con­tain no data res­id­ency, loc­al­isa­tion or in-King­dom host­ing re­quire­ment, which was es­tab­lished by search­ing the whole in­stru­ment. The Na­tional Cy­ber­se­cur­ity Au­thor­ity's Cloud Cy­ber­se­cur­ity Con­trols, CCC-2:2024, de­leted the sub-con­trols of the 2020 edi­tion that had re­quired cloud ser­vices to be provided from in­side the King­dom. It re­cor­ded that loc­al­isa­tion con­trols had moved to the NDMO, and the re­ceiv­ing in­stru­ment has not been iden­ti­fied. Clas­si­fic­a­tion vocab­u­lary di­verges as well. The SDAIA Data Clas­si­fic­a­tion Policy, ver­sion 1 of 5 May 2020, uses Top Secret, Secret, Re­stric­ted and Pub­lic, and gives per­son­ally iden­ti­fi­able in­form­a­tion as an ex­ample of Re­stric­ted. CST's re­gis­tra­tion classes use a dif­fer­ent vocab­u­lary again. Quote the in­stru­ment being re­lied on, and name it.

The hon­est op­tions for a reg­u­lated Saudi work­load do not in­clude a sov­er­eign Azure re­gion. Mi­crosoft's own doc­u­ment­a­tion points else­where. The fea­tures that dis­tin­guish Sov­er­eign Pub­lic Cloud are doc­u­mented as European Union and EFTA fea­tures, namely the EU Data Bound­ary, Data Guard­ian and op­er­a­tions con­trolled by European res­id­ents. The only two Na­tional Part­ner Clouds Mi­crosoft names are Bleu in France and Delos Cloud in Ger­many (Mi­crosoft Learn, pages up­dated 31 March 2026). The branch that an­swers a Saudi or Kuwaiti ques­tion is Sov­er­eign Private Cloud. That means Azure Local and Mi­crosoft 365 Local, run­ning on in­fra­struc­ture the cus­tomer or a local part­ner op­er­ates. What those com­pon­ents in­clude against the cloud ser­vice is not es­tab­lished in the doc­u­ment­a­tion this paper re­lies on, and nor is whether Fab­ric runs on them in any form. Es­tab­lish both in writ­ing be­fore com­mit­ting. Out­side Europe, Mi­crosoft's sov­er­eignty story is en­cryp­tion and policy rather than jur­is­dic­tion.

Storage is committed, processing is not

The com­mit­ment a Gulf or In­dian en­ter­prise holds is about where data rests. It is not about where the data is worked on. India, Qatar and the United Arab Emir­ates are Local Re­gion Geo­graph­ies, so a ten­ant in any of the three holds dur­able com­mit­ments on data loc­a­tion for Ex­change On­line, SharePoint, OneDrive, Teams and Co­pi­lot (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). Those com­mit­ments come through Product Terms, Multi-Geo and the Ad­vanced Data Res­id­ency add-on. Every one of them is about stor­age. The European Union Data Bound­ary is the only pro­cessing com­mit­ment Mi­crosoft pub­lishes, and eli­gib­il­ity rests on a sign-up loc­a­tion in the European Union or EFTA. There is no Gulf or In­dian equi­val­ent.

Ad­vanced Data Res­id­ency is also sold the wrong way round. The core pro­ductiv­ity ser­vices already carry a Product Terms com­mit­ment for an eli­gible ten­ant. What the add-on buys, out­side Europe, is a dur­able in-coun­try com­mit­ment over the se­cur­ity and com­pli­ance tier: Mi­crosoft De­fender for Of­fice P1, Ex­change On­line Pro­tec­tion, Of­fice for the web, Viva Con­nec­tions and six named Pur­view ser­vices (Mi­crosoft Learn, page up­dated 27 Oc­to­ber 2026). The six are Audit Stand­ard, Audit Premium, Data Li­fe­cycle Man­age­ment, Data Loss Pre­ven­tion, In­form­a­tion Bar­ri­ers and In­form­a­tion Pro­tec­tion. The add-on also ad­mits no par­tial ad­op­tion. Every paid li­cence in the ten­ant must carry one.

Some work­loads leave the re­gion by de­fault, and no add-on brings them back. Viva In­sights Ad­vanced, Viva Learn­ing and Plan­ner de­fault to the European macro re­gion for every Gulf state, with Oman's Viva Learn­ing going to Asia Pa­cific in­stead

Data & AnalyticsVoice of Volge17 •

(Mi­crosoft Learn, page up­dated 8 Oc­to­ber 2026). Mi­crosoft Forms data for every Gulf and In­dian ten­ant sits in the United States. Ad­vanced Data Res­id­ency cov­ers none of those ser­vices, so there is no paid rem­edy. India has a ded­ic­ated Plan­ner geo­graphy, IND7, while its Viva data goes to Asia Pa­cific. For an ana­lyt­ics pro­gramme the con­sequence is nar­row and worth stat­ing. Tele­metry and sur­vey data that feed ad­op­tion re­port­ing may sit out­side the geo­graphy the lake­house is com­mit­ted to.

India, where residency is a question about data domains

India's res­id­ency rules do not at­tach to a ten­ant. They at­tach to classes of data, which is why an In­dian lake­house is de­signed per do­main. The Di­gital Per­sonal Data Pro­tec­tion Act, 2023 is phas­ing in. The Di­gital Per­sonal Data Pro­tec­tion Rules, 2025, no­ti­fied on 13 Novem­ber 2025 as G.S.R. 846(E), com­mence in tranches: rules 1, 2 and 17 to 21 on no­ti­fic­a­tion, rule 4 on con­sent man­agers from 13 Novem­ber 2026, and rules 3, 5 to 16, 22 and 23 from 13 May 2027. Everything of com­mer­cial con­sequence sits in the last tranche. Until then sec­tion 43A of the In­form­a­tion Tech­no­logy Act, 2000 and the SPDI Rules re­main the op­er­at­ive In­dian data se­cur­ity pro­vi­sions.

Two of those rules bear dir­ectly on a medal­lion design. Rule 6(1) re­quires en­cryp­tion, ob­fus­ca­tion, mask­ing or vir­tual tokens, ap­pro­pri­ate ac­cess con­trol for com­puter re­sources, and vis­ib­il­ity of ac­cess through logs, mon­it­or­ing and re­view. Rules 6(1)(e) and 8(3) im­pose a min­imum one-year re­ten­tion of per­sonal data, as­so­ci­ated traffic data and logs of pro­cessing, bind­ing on a pro­cessor en­gaged by a fi­du­ciary as well. That is a re­ten­tion floor. It cuts against de­le­tion by de­fault, and against ag­gress­ive ex­piry in the bronze layer. Rule 7 then re­quires breach in­tim­a­tion to each af­fected data prin­cipal without delay. The Data Pro­tec­tion Board is told without delay as well, and again in de­tail within sev­enty-two hours. There is no de min­imis threshold at all.

The loc­al­isa­tion lever in the DPDP in­stru­ments has not been pulled. Rule 13(4) per­mits the Cent­ral Gov­ern­ment to spe­cify per­sonal data, and the traffic data per­tain­ing to its flow, that may not be trans­ferred out­side India. The re­com­mend­a­tion comes from a com­mit­tee of of­fi­cials, and the rule reaches only a Sig­ni­fic­ant Data Fi­du­ciary once no­ti­fied, from 13 May 2027. No data has been spe­cified. No com­mit­tee has been an­nounced. Sec­tion 16 works as a neg­at­ive list rather than as an ad­equacy whitel­ist, so trans­fer is per­mit­ted by de­fault and re­stric­ted only to a coun­try the Gov­ern­ment names. No coun­try has been named.

The bind­ing In­dian res­id­ency rules are sec­toral, and each lands on a dif­fer­ent data do­main in­side the same lake­house. The Re­serve Bank of India's Stor­age of Pay­ment Sys­tem Data cir­cu­lar of 6 April 2018, RBI/2017-18/153, re­quires at clause 2(i) that the en­tire data re­lat­ing to pay­ment sys­tems be stored in a sys­tem only in India. The RBI's FAQs of 26 June 2019 per­mit pro­cessing abroad, on one con­di­tion. The data must be de­leted abroad and brought back to India within one busi­ness day or twenty-four hours, whichever is earlier. Re­ten­tion abroad pending dis­pute res­ol­u­tion is not per­mit­ted. SEBI's Frame­work for Ad­op­tion of Cloud Ser­vices by SEBI Reg­u­lated En­tit­ies, cir­cu­lar of 6 March 2023 and in force, re­quires MeitY-em­pan­elled pro­viders whose data centres hold valid STQC audit status. It also re­quires stor­age and pro­cessing of reg­u­lated-en­tity data within the legal bound­ar­ies of India, dis­aster re­cov­ery and near-dis­aster-re­cov­ery cop­ies in­cluded. The CSCRF's own loc­al­isa­tion con­trol, PR.DS.S2, has been in abey­ance since SEBI's cla­ri­fic­a­tions of 31 Decem­ber 2024, and re­mains so.

Three fur­ther rules com­plete the pic­ture. The IRDAI (Main­ten­ance of In­form­a­tion by the Reg­u­lated En­tit­ies and Shar­ing of In­form­a­tion by the Au­thor­ity) Reg­u­la­tions, 2025, no­ti­fied on 1 Janu­ary 2025, re­quire at reg­u­la­tion 9(iii) that re­cords of all policies is­sued and all claims made in India be held in data centres loc­ated and main­tained in India only. The Tele­com­mu­nic­a­tions (Au­thor­isa­tion for Tele­com­mu­nic­a­tion Net­work) Rules, 2026, G.S.R. 644(E), in force from 20 July 2026, re­quire at rule 25(3) that all sys­tems of a newly au­thor­ised tele­com­mu­nic­a­tion net­work be stored in India only, along with the data, logs and in­form­a­tion as­so­ci­ated with it. No cop­ies may be sent, shared or made avail­able out­side India. The Eng­lish ren­der­ing of that rule is MediaNama's, 24 July 2026. And the De­part­ment of Sci­ence and Tech­no­logy's geo­spa­tial guidelines of 15 Feb­ru­ary 2021 re­quire data finer than the stated threshold to be stored and pro­cessed only on serv­ers loc­ated in India.

Logs are the In­dian ques­tion answered wrongly most often in a lake­house design. CERT-In's Dir­ec­tions of 28 April 2022, No. 20(3)/2022-CERT-In, are in force and un­amended. Dir­ec­tion (iv) re­quires every ser­vice pro­vider and in­ter­me­di­ary, every data centre and body cor­por­ate, and every Gov­ern­ment or­gan­isa­tion to en­able logs of all their ICT sys­tems and main­tain them se­curely for a rolling period of 180 days, "main­tained within the In­dian jur­is­dic­tion". CERT-In's own FAQs of May 2022 say at Q35 that logs "may be stored out­side India also as long as the ob­lig­a­tion to pro­duce logs to CERT-In is ad­hered to by the en­tit­ies in a reas­on­able time". The bind­ing text says India. The cla­ri­fic­a­tion says any­where, provided pro­duc­tion is

Data & AnalyticsVoice of Volge18 •

timely. Cite both, de­cide de­lib­er­ately, and re­cord the de­cision. The six-hour in­cid­ent re­port­ing ob­lig­a­tion in dir­ec­tion (ii) is what the log loc­a­tion is ul­ti­mately for.

The gates sit on the supplier, not only on the data

Four ac­cred­it­a­tion gates in the re­gion bind the sup­plier rather than the data. A con­sol­id­a­tion plan that ig­nores them stalls at pro­cure­ment. CST re­gis­tra­tion by class is the Saudi gate. In Dubai, the Dubai Elec­tronic Se­cur­ity Cen­ter's Cloud Ser­vice Pro­vider Se­cur­ity Stand­ard is man­dat­ory, in DESC's own words, for all pro­viders wish­ing to offer cloud ser­vices for Dubai gov­ern­ment and semi-gov­ern­ment en­tit­ies (DESC, cer­ti­fic­a­tions page mod­i­fied 27 March 2026). It car­ries an­nual sur­veil­lance audits and re­cer­ti­fic­a­tion every three years. Kuwait re­quires a CITRA cloud li­cence. Clause 2.1 of the Cloud Com­put­ing Reg­u­lat­ory Frame­work ver­sion 2.4 per­mits only a CITRA-li­censed party with a data centre and in­fra­struc­ture in­side Kuwait to con­tract to provide cloud ser­vices to the pub­lic sec­tor, and the li­cence is not avail­able to for­eign op­er­at­ors dir­ectly. Oman re­quires a Tele­com­mu­nic­a­tions Reg­u­lat­ory Au­thor­ity au­thor­isa­tion be­fore con­tract­ing, under the Cloud First Policy pub­lished 1 July 2026.

Two of those re­gimes also con­strain the data it­self, in ways a lake­house feels im­me­di­ately. Kuwait's frame­work pro­hib­its host­ing Level 3 and Level 4 data out­side Kuwait, along with in­di­vidu­als' per­sonal data held by gov­ern­ment bod­ies, private-sec­tor com­pan­ies or ser­vice pro­viders. Level 3 data must sit in a private or hy­brid cloud (clauses 4.2.1.1 and 3.1.4.3 of the Ar­abic primary text). Clauses 4.3.3 and 4.3.4 go fur­ther. They for­bid a pro­vider from com­ply­ing with a for­eign jur­is­dic­tional de­mand for sub­scriber con­tent. Oman's Cloud Com­put­ing First Cir­cu­lar, MTCIT Sum­mary 10/2026 of 28 June 2026, re­quires gov­ern­ment en­tit­ies to clas­sify data be­fore stor­ing or pro­cessing it in a cloud en­vir­on­ment. Clas­si­fic­a­tion is there­fore a pre­con­di­tion of mi­gra­tion rather than a par­al­lel work­stream.

The United Arab Emir­ates runs two re­gimes that pull in op­pos­ite dir­ec­tions. The fed­eral Per­sonal Data Pro­tec­tion Law, Fed­eral De­cree-Law No. 45 of 2021, has been in force since 2 Janu­ary 2022 and still had no Ex­ec­ut­ive Reg­u­la­tions as at 8 Novem­ber 2026. It there­fore car­ries no com­pli­ance com­mence­ment date, no stat­utory breach no­ti­fic­a­tion dead­line and no pub­lished pen­alty sched­ule. Art­icle 2(2) carves out gov­ern­ment data, free zones with their own le­gis­la­tion, health data and bank­ing data. DIFC Law No. 5 of 2020 and the ADGM reg­u­la­tions sit sep­ar­ately, and the ADGM text could not be veri­fied. Against that, Fed­eral Law No. 2 of 2019 on the Use of In­form­a­tion and Com­mu­nic­a­tions Tech­no­logy in Health Fields is the coun­try's hard­est loc­al­isa­tion rule. Art­icle 13, in force since 14 May 2019, pro­hib­its stor­ing, pro­cessing, gen­er­at­ing or trans­fer­ring health data re­lat­ing to health ser­vices provided in­side the State out­side the State, free zones in­cluded. Art­icle 24 at­taches a fine of not less than AED 500,000 and not more than AED 700,000. An ana­lyt­ics land­scape that in­gests health ser­vice data in the UAE has a crim­inal pro­vi­sion to sat­isfy, not a policy.

What this constrains

The re­gional ground changes the se­quence rather than the des­tin­a­tion. Clas­sify the data do­mains be­fore pla­cing them, be­cause every bind­ing rule in this chapter at­taches to a class of data and not to a ten­ant. Place each do­main be­fore pro­vi­sion­ing a ca­pa­city. The ca­pa­city fixes the re­gion, and the re­gion fixes the law. Sep­ar­ate the stor­age po­s­i­tion from the pro­cessing po­s­i­tion in every doc­u­ment that reaches a reg­u­lator, be­cause the vendor does. And treat the out-of-re­gion de­faults as design facts to be dis­closed rather than dis­covered: Forms in the United States, Viva and Plan­ner in Europe for the Gulf, and the Co­pi­lot Stu­dio data loc­a­tion in the United States for a Qatari or Saudi ten­ant. A dis­clos­ure made in ad­vance is a design con­straint. The same fact found by a su­per­visor is a find­ing.

What to write down, and where

A re­gional po­s­i­tion is only use­ful once it sits on paper in three places, and the same facts serve all three. Keep the word­ing identical across them, be­cause a dif­fer­ence between them is what an ex­am­iner finds.

  • In the ar­chi­tec­ture de­cision re­cord. The re­gion of every ca­pa­city and work­space, the data do­mains each may hold, the pro­cessing po­s­i­tion stated sep­ar­ately from the stor­age po­s­i­tion, and the re­cov­ery design where an un­paired re­gion is in use.
Data & AnalyticsVoice of Volge19 •
  • In the con­tract. The res­id­ency com­mit­ment re­lied upon, by name, with the add-on it de­pends on and the ser­vices it ex­cludes. Mi­crosoft's own doc­u­ment­a­tion is the right ref­er­ence, be­cause it is spe­cific where a mar­ket­ing claim is not, and be­cause it car­ries a date.
  • In the su­per­vis­ory sub­mis­sion. The clas­si­fic­a­tion of each data do­main, the legal basis for any cross-bor­der stor­age or pro­cessing, the log loc­a­tion with the pro­duc­tion time the ap­plic­able dir­ec­tion re­quires, and the con­trol that would de­tect a breach of the stated po­s­i­tion rather than the policy that for­bids it.

One line be­longs in all three and is usu­ally miss­ing. Name what is not in re­gion. That in­cludes the iden­tity layer, any pre­view cap­ab­il­ity in use, and any sup­port ac­cess from out­side the geo­graphy. Mi­crosoft doc­u­ments that pre­view, beta and other pre-re­lease ser­vices typ­ic­ally store cus­tomer data in the United States and may store it glob­ally (Mi­crosoft Learn, page up­dated 7 Oc­to­ber 2026). That is the clause to read be­fore a res­id­ency-con­strained pro­gramme ad­opts an early-ac­cess cap­ab­il­ity.

, The data-leadership agenda

Con­sol­id­at­ing the ana­lyt­ics land­scape 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 land­scape 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 the OneLake Cata­log and Pur­view gov­ernance layer, and the sav­ing it of­fers is the in­teg­ra­tion tax it re­tires, set against the spe­cial­ised depth a best-of-breed land­scape keeps.

For most en­ter­prises the plat­form de­cision is shaped by the ex­ist­ing Azure and Power BI land­scape 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.

A sixth ques­tion now be­longs be­side those five, and in the Gulf and India it out­ranks sev­eral of them. Where may each data do­main law­fully rest, and where may it law­fully be

pro­cessed. The an­swer is not a ten­ant-level an­swer, and it is not a com­mit­ment the vendor makes about pro­cessing. Mi­crosoft com­mits to stor­age at rest in the ten­ant geo­graphy and pub­lishes a pro­cessing com­mit­ment only for the European Union, and the plat­form it­self is un­avail­able in a Gulf coun­try that has an Azure re­gion, a res­id­ency com­mit­ment and a Foundry pres­ence. Chapter five sets out what fol­lows for the se­quence of work.

Five ques­tions be­long on the lead­er­ship agenda. First, is the in­teg­ra­tion tax ac­tu­ally quan­ti­fied, or merely as­sumed. The tax is du­plic­ated stor­age and copy pipelines, re­dund­ant ca­pa­city and trust erosion. Second, is the land­scape 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, is the hy­brid de­lib­er­ate and con­nec­ted by zero-copy in­ter­op­er­ab­il­ity, rather than an ac­ci­dent of un­fin­ished con­sol­id­a­tion. Large-scale en­gin­eer­ing on Dat­ab­ricks and high-con­cur­rency ware­hous­ing on Snow­flake are two such cases.

The de­cision is not whether to con­sol­id­ate. It is what to con­sol­id­ate, in what order, in­side what the law of each jur­is­dic­tion al­lows, and against a baseline some­body has taken. A pro­gramme that can an­swer those four has a case. A pro­gramme that can­not is buy­ing a plat­form and hop­ing the sav­ing fol­lows.

Data & AnalyticsVoice of Volge20 •

Appendices

Appendix A , Platform capability comparison: Fabric, Databricks, Snowflake

CapabilityMicrosoft Fabric (OneLake)DatabricksSnowflake
Delivery modelEnd-to-end SaaS, unified workloadsLakehouse platform, multi-cloudCloud data warehouse, multi-cloud
Unified logical lakeOneLake, provisioned per tenantUnity Catalogue over lakehouseSingle platform; lake plus warehouse
Open storage formatDelta Parquet, open by defaultDelta Lake and IcebergIceberg and proprietary
Greatest strengthUnification; BI and Microsoft-ecosystem integrationLarge-scale data engineering and ML on SparkHigh-concurrency SQL; data sharing
BI integrationNative Power BI with Direct LakeConnects to external BIConnects to external BI
Real-timeReal-Time Intelligence, Eventhouse, KQLStructured streamingSnowpipe streaming
GovernanceOneLake Catalog and Microsoft Purview, integratedUnity CatalogueHorizon governance
Capacity / pricingPooled F-SKU capacity; one meterCompute per workload; consumptionPer-warehouse compute; consumption
Principal constraintSpecialist engineering/warehousing ceilingLess unified; heavier to operateNarrower than a full unified landscape
InteroperabilityBi-directional zero-copy with bothReads OneLake via Unity CatalogueInteroperability with Fabric

Fab­ric is the primary lens of this paper. In­ter­op­er­ab­il­ity is treated as a design fea­ture, not an ex­cep­tion. A con­sol­id­ated land­scape need not be single-vendor. On scale, Dat­ab­ricks re­ports >USD 7bn rev­enue run-rate and >1,000 cus­tom­ers above USD 1M (Dat­ab­ricks, Au­gust 2026), while Snow­flake re­ports 14,554 cus­tom­ers, 828 of them above USD 1M trail­ing-twelve-month product rev­enue in the quarter ended 31 July 2026 (Snow­flake earn­ings, Q2 FY2027). Sources: Mi­crosoft and vendor dis­clos­ures and in­dustry ana­lysis, 2025-2026.

As­sess and land

Con­sol­id­ate

Appendix B , Migration and readiness checklist

A min­imum set of con­di­tions for a de­fens­ible con­sol­id­a­tion pro­gramme, or­gan­ised by phase.

  • 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 land­scape , Syn­apse, ADLS, Power BI Premium, ADF , is in­vent­or­ied, with an as­sess­ment-first com­pat­ib­il­ity re­view com­pleted.
  • The OneLake Cata­log and Mi­crosoft Pur­view are stood up from the out­set, not de­ferred to a later phase.
  • 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 land­scapes 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.

Com­piled by Volge from Mi­crosoft Learn re­gion and data-loc­a­tion doc­u­ment­a­tion, pages up­dated between 8 Oc­to­ber and 5 Novem­ber 2026 and veri­fied 8 Novem­ber 2026.

Data & AnalyticsVoice of Volge21 •

Avail­ab­il­ity changes; re­con­firm against cur­rent Mi­crosoft doc­u­ment­a­tion be­fore an ar­chi­tec­tural de­cision.

Appendix E , Regional footprint and instruments, as at 8 November 2026

The two tables below carry the re­gional facts chapter five ar­gues from, so that an ar­chi­tec­ture can be checked against them without re­read­ing the chapter.

JurisdictionAzure regionMicrosoft 365 Local Region GeographyFabric workloads in regionFoundry project regionCopilot Studio in region
Saudi ArabiaNoneNoNoNoNo
United Arab EmiratesUAE North; UAE Central restrictedYesUAE North onlyUAE NorthUAE North
QatarQatar Central, unpairedYesNo, Power BI onlyQatar CentralNo, data stored in the United States
IndiaCentral India, India South Central, South India, West IndiaYesCentral India and South IndiaCentral India and South IndiaCentral India and South India
KuwaitNoneNoNoNoNo
OmanNoneNoNoNoNo
BahrainNoneNoNoNoNo
JurisdictionInstrumentWhat it doesStatus
Saudi ArabiaCST Cloud Computing Services Provisioning Regulations v4, Decision No. 506/1445, effective 10 October 2023Prohibits transfer of Saudi government agency data outside the Kingdom; gates providers by Class A, B or C registrationIn force
Saudi ArabiaSAMA Cyber Security Framework 3.4.3, No. 381000091275, 24 May 2017In-principle in-Kingdom cloud with explicit SAMA approval otherwise, and SAMA approval before cloud is used at allIn force
JurisdictionInstrumentWhat it doesStatus
Saudi ArabiaSDAIA Data Classification Policy v1, 5 May 2020Four levels, Top Secret, Secret, Restricted and Public, with personally identifiable information given as an example of RestrictedIn force
Saudi ArabiaNCA Cloud Cybersecurity Controls CCC-2:2024Deleted the 2020 edition's in-Kingdom provisioning sub-controls, recording transfer of localisation controls to the NDMOIn force; receiving instrument not identified
United Arab EmiratesFederal Decree-Law No. 45 of 2021, in force 2 January 2022Data protection law whose Executive Regulations have not been issued, leaving no commencement date, breach deadline or penalty scheduleIn force, machinery deferred
United Arab EmiratesFederal Law No. 2 of 2019 on ICT in Health Fields, Article 13, effective 14 May 2019Prohibits storing, processing, generating or transferring health data relating to health services provided in the State outside the State, free zones includedIn force
United Arab EmiratesDESC Cloud Service Provider Security StandardMandatory, in DESC's words, for cloud services to Dubai government and semi-government entities, with annual surveillance auditsIn force
QatarLaw No. 13 of 2016 on Personal Data Privacy ProtectionDiscourages measures that curb trans-border data flow; no adequacy list and no prior approval gate for transfersIn force
KuwaitCITRA Cloud Computing Regulatory Framework v2.4Requires a CITRA licence and an in-Kuwait data centre to serve the public sector; prohibits hosting Level 3 and Level 4 data outside KuwaitIn force
OmanMTCIT Cloud Computing First Circular, Summary 10/2026, 28 June 2026, with the Cloud First Policy published 1 July 2026Requires classification before cloud storage or processing, and contracting only with TRA-authorised providersIn force
Data & AnalyticsVoice of Volge22 •
JurisdictionInstrumentWhat it doesStatus
IndiaDPDP Rules, 2025, G.S.R. 846(E), 13 November 2025Security safeguards, a one-year retention floor on personal data and processing logs, and two-stage breach intimationPhasing in; principal tranche 13 May 2027
IndiaCERT-In Directions, No. 20(3)/2022-CERT-In, 28 April 2022Six-hour incident reporting, and 180-day ICT logs maintained within the Indian jurisdiction, qualified by FAQ Q35 of May 2022In force
IndiaRBI Storage of Payment System Data, RBI/2017-18/153, 6 April 2018Entire payment system data stored in a system only in India, with processing abroad requiring return within 24 hoursIn force
IndiaSEBI Framework for Adoption of Cloud Services, circular of 6 March 2023MeitY-empanelled providers with STQC-audited data centres, and data including disaster-recovery copies within IndiaIn force; CSCRF control PR.DS.S2 in abeyance
IndiaIRDAI (Maintenance of Information by the Regulated Entities and Sharing of Information by the Authority) Regulations, 2025, 1 January 2025Records of policies issued and claims made in India held in data centres located and maintained in India onlyIn force
IndiaTelecommunications (Authorisation for Telecommunication Network) Rules, 2026, G.S.R. 644(E), 20 July 2026Systems, data and logs of a newly authorised network stored in India only, with no copies made available outside IndiaIn force

Com­piled by Volge from primary in­stru­ments and, where the primary text could not be read, from named law-firm com­ment­ary iden­ti­fied at the point of use. This table is in­dic­at­ive and not legal ad­vice; Volge is not a law firm, and qual­i­fied coun­sel should be en­gaged for com­pli­ance de­cisions.

Pre­pared by Volge · June 2026, re­vised Novem­ber 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 at 8 Novem­ber 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 at 8 Novem­ber 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.

Data & AnalyticsVoice of Volge23 •

Reference

Glossary
  • Activator. The Fabric capability that triggers actions when data meets defined conditions. Formerly Data Activator, it was renamed and folded into Real-Time Intelligence; Reflex items are now Activator items.
  • Capacity unit (CU). The unit of pooled compute a Fabric F-SKU provides, shared across all workloads.
  • 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 F8192); F64 is the inflection that unlocks Power BI Premium-equivalent features and Direct Lake.
  • Integration tax. The distributed cost of holding a fragmented landscape 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.
  • OneLake Catalog. The centralised Fabric experience for discovering, exploring, securing, and governing items, and for assessing governance state.
  • Purview. Microsoft's control plane for information protection, data loss prevention, audit, and risk and compliance over supported Fabric data.
  • 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 nine role-tailored experiences over shared OneLake storage: Power BI, Databases, Data Factory, Industry Solutions, Real-Time Intelligence, Data Engineering, Data Science, Data Warehouse, and Fabric IQ, which is in preview.
Data & AnalyticsVoice of Volge24 •
Sources

This paper draws on the fol­low­ing, ac­cessed June 2026 and re-veri­fied 8 Novem­ber 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 nine 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 Learn , What is Microsoft Fabric (verified 8 November 2026): the nine Fabric workloads, including Databases, Industry Solutions, and Fabric IQ in preview; the division of governance between the OneLake Catalog and Microsoft Purview; shortcuts to Azure Data Lake Storage, Amazon S3, and Google Cloud Storage.
  • Microsoft , Q1 FY2027 earnings call (28 October 2026): 44,000 paid Fabric customers, up more than 50 per cent year over year; more than 19,000 customers running Foundry and Fabric together; 120,000 Foundry customers.
  • Microsoft , Q4 FY2026 earnings disclosures on Fabric (29 July 2026): 40,000 paid customers, up about 60 per cent year over year, with 5,000 added in the quarter; more than 17,000 customers running Foundry and Fabric together.
  • Microsoft , Q3 FY2026 earnings call (29 April 2026): OneLake data volume up nearly four times year over year; 35,000 paid Fabric customers.
  • Microsoft , Q2 FY2026 earnings call (28 January 2026): Fabric annual revenue run rate above two billion United States dollars, with more than 31,000 customers and revenue up 60 per cent year over year.
Data & AnalyticsVoice of Volge25 •
  • Microsoft Learn , medallion lakehouse architecture; Direct Lake overview; Real-Time Intelligence, Eventhouse, and Activator guidance (2025-2026).
  • Microsoft , Synapse and Azure Data Factory to Fabric migration guidance; targeted deprecations (Synapse dedicated SQL pools; Synapse Data Explorer retired 7 October 2025; Azure Data Factory Legacy Trusted Service mode retiring August 2027); assessment-first migration assistants (2025-2026).
  • Microsoft Learn , OneLake Catalog overview and Microsoft Purview guidance , discovery and item governance in the catalogue; information protection, data loss prevention, audit, and compliance in Purview (2026).
  • Microsoft Fabric pricing (verified 8 November 2026) , thirteen F-SKUs from F2 to F8192; pay-as-you-go at about USD 0.18 per capacity unit per hour, giving about USD 263/month at F2, USD 8,410/month at F64, and USD 1,076,000/month at F8192 for capacity running continuously; a one-year or three-year reservation saving about 41 per cent; free report-viewer licences at F64 and above. Pricing is region- and contract-specific and changes; reconfirm against current Microsoft pricing.
  • Databricks , press release, 13 August 2026: revenue run-rate above USD 7 billion, more than 80 per cent year-over-year growth in its second quarter (ended 31 July 2026), more than 1,000 customers consuming above USD 1 million; industry analysis (2026) on Apache Spark and Delta Lake heritage and engineering/ML positioning.
  • Snowflake , Q2 FY2027 results (2 September 2026, quarter ended 31 July 2026): 14,554 customers, with 828 above USD 1 million in trailing-twelve-month product revenue; cloud-data-warehouse positioning and data-sharing strength.
  • Volge , The data estate is the accuracy budget (Topic 7), the companion paper carrying the agent-grounding argument referenced once here.
Data & AnalyticsVoice of Volge26 •

Volge and our expertise on the data platform

Volge con­sol­id­ates ana­lyt­ics land­scapes onto one gov­erned copy of the data, and the work is en­gin­eer­ing rather than ad­vice. The prac­tice sits where this paper loc­ates the money, in the seams between ser­vices: the du­plic­ated cop­ies, the pipelines that exist only to move data, the ca­pa­city re­served sev­eral times over, and the re­con­cili­ation that fol­lows two ver­sions of one re­cord. De­liv­ery is by senior data en­gin­eers. They work in­side the cli­ent's own ten­ant, ca­pa­city and re­pos­it­or­ies, through to the de­com­mis­sion­ing of the leg­acy meters.

The work has a usual shape. A baseline pass quan­ti­fies the in­teg­ra­tion tax on the four counts set out in chapter one, with the raw ex­tracts re­tained and an owner named against each fig­ure, be­cause a con­sol­id­a­tion case that can­not show its baseline does not sur­vive a fin­ance re­view. The build then fol­lows the de­pend­ency order of chapter four. Res­id­ency and clas­si­fic­a­tion come first, then the OneLake found­a­tion and a right-sized ca­pa­city, then the ware­house, then the pipelines and se­mantic mod­els. De­com­mis­sion­ing comes last, and it is where the sav­ing is real­ised.

The re­gional po­s­i­tion is the part a Gulf or In­dian cli­ent usu­ally needs first. Volge is headquartered in the Gulf, and the dis­tinc­tion between a stor­age com­mit­ment and a pro­cessing com­mit­ment set out in chapter five is es­tab­lished per data do­main be­fore a ca­pa­city is pro­vi­sioned, not after. Where the plat­form is not avail­able in coun­try, as in Qatar, the op­tions are set out with their legal con­sequences at­tached rather than de­scribed as a lim­it­a­tion. Where no com­mit­ment ex­ists at all, as in Saudi Ar­a­bia, the choice between a doc­u­mented cross-bor­der po­s­i­tion and in­fra­struc­ture the cli­ent or a local part­ner op­er­ates is taken de­lib­er­ately and re­cor­ded.

Who does the work. A de­ploy­ment strategist holds the busi­ness out­come and the gov­ernance ca­dence with the cli­ent's data lead­er­ship. A trans­form­a­tion ar­chi­tect holds the tar­get design, the ca­pa­city to­po­logy and the se­quence. Senior data en­gin­eers own the work­streams and build in­side the cli­ent's en­vir­on­ment. The same people stay from the baseline to the final de­com­mis­sion­ing, which is what makes the con­trol tests

in chapter four drills the team has run rather than pro­ced­ures it has re­com­men­ded.

An en­gage­ment starts with a scoped as­sess­ment, usu­ally four to six weeks. It pro­duces the quan­ti­fied baseline, a res­id­ency and pro­cessing po­s­i­tion for each data do­main, a copy re­gister with an owner against every in­stance, a ca­pa­city siz­ing taken from ob­served con­sump­tion, and a staged plan whose first stage already has an exit test. Volge is a Mi­crosoft part­ner and be­ne­fits com­mer­cially where a cli­ent ex­pands its Mi­crosoft foot­print, which a reader should weigh when read­ing the re­com­mend­a­tions in this paper. The con­ver­sa­tion be­gins wherever the land­scape stands today.

Data & AnalyticsVoice of Volge27 •
Version · VVM-DAA-002 · 2026
volgeEnable Powerful
Transformations
About Volge

Volge is an engineering firm providing enterprise technology transformation services to clients across the Middle East, India, APAC, Europe, and North America.

Delivery stands on two pillars: Forward Deployed Engineering, with senior specialists embedded in the client’s own team, and Intelligent Command Centre Operations, run from our Global Delivery Centres, which keeps what they build running. Senior engineers stay with an engagement from the first diagnosis through to production.

Our services span Strategy, Data, Platforms, AI and Automation, Cyber Trust, Technology Operations, and Applied Functional Engineering. Engagements begin with a diagnosis. The firm is built on working autonomy: engineers decide, innovate, co-create, and own their work. Motivated engineers build the systems that sustain our clients’ growth.

Terms of use

This edition may be shared with organisations evaluating a Volge engagement. It may not be altered or republished without Volge’s permission, and figures carry their stated basis and date. Volge, Intelligent Command Centre, and Intelligent Command Centre Operations are marks of Volge; other names are the trademarks of their respective owners.