BUSINESS OVERVIEW

Microsoft Corporation

NASDAQ: MSFT · Fiscal year ended 30 June 2026

7 September 2026

Evidence base: Microsoft's own SEC filings — the FY2026 Form 10-K (filed 29 July 2026) and the FY2016–FY2025 10-Ks, the FY2026 Q1 and Q3 10-Qs, the FY2026 Q4 earnings release, the 2025 proxy statement, and the 8-K of 2 September 2026 — supplemented by independent industry evidence from competitors' own filings, Synergy Research, Lawrence Berkeley National Laboratory, EIA, PJM, the US Census Bureau, the Federal Reserve, NBER, and the UK CMA and European Commission.

This is an analysis of how the business works and what drives its economics. It is not a valuation and not a recommendation.

1. Executive Snapshot

Sources: FY2026 Form 10-K; FY2025 10-K; UK CMA; European Commission. Margins computed from segment data in FY2026 10-K Note 18.

2. What the Company Does

Every organisation of any size faces the same problem: its employees need a common set of tools to write, communicate and analyse; its applications need somewhere to run; and both need a single system that knows who each employee is and what they are permitted to see. Historically an enterprise solved this by buying software licences and running them on machines it owned. Microsoft sells the modern version of the same thing — the tools, the place to run them, and the identity system that connects the two — as subscriptions and metered services, invoiced in advance under agreements that typically run three years.

The unit of economics — and why there are now two of them

For most of Microsoft's history the unit was one paid seat. A company buys a Microsoft 365 subscription for each employee; Microsoft bills annually in advance for a multi-year term and recognises the revenue ratably. The cost of serving that seat is a small slice of shared datacenter capacity plus support. Nothing about the arrangement scales with capital, which is why Microsoft could grow the seat business for a decade while capital expenditure stayed near 10% of revenue.

That is still the larger profit pool. Productivity and Business Processes generated $139,996m of revenue and $83,879m of operating income in FY2026 — a 59.9% operating margin, on 42.2% of company revenue producing 54.0% of company operating income (FY2026 10-K, Note 18; margins computed). It is also still improving: that margin was 53.2% in FY2023.

The marginal dollar, however, no longer comes from a seat. It comes from a second unit — one metered hour of rented computing capacity. A customer contracts for Azure capacity and is billed for what it consumes. Microsoft must first buy the servers, build or lease the datacenter, and secure the electricity. The two units are opposites: the seat is an annuity that consumes almost no capital, and the compute-hour is a capital asset rented out over a depreciation schedule. Intelligent Cloud earned a 41.3% operating margin in FY2026 on a segment gross margin of 58.0%, down from 66.9% in FY2023 (computed from FY2026 10-K, Note 18).

Tracing one compute-hour from production to cash explains the whole business model. Microsoft buys a GPU server from a supplier earning roughly 75% gross margin (NVIDIA, Q2 FY2027 results, 26 August 2026), installs it in a datacenter it has either built under a construction commitment or taken on a finance lease with a thirteen-year weighted-average term, and powers it with electricity it must secure years ahead. It then capitalises the server and depreciates it over up to six years, and sells the hours it produces under a contract that is recognised as consumed. The revenue is recognised over years; the cash left the building at the start. That timing gap is the single most important fact about Microsoft's current financial statements.

What is actually sold, and by whom it is bought

Revenue by product line, FY2024 to FY2026 ($m, as reported in the FY2026 10-K, Note 18):

Source: FY2026 10-K, Note 18. Growth percentages computed. These line items were re-cut in the FY2025 recast; the FY2024 figures above are the recast ones and do not match the FY2024 10-K as originally filed.

The buyer is overwhelmingly a business, not a consumer. Microsoft does not disclose a commercial-versus-consumer revenue split, but the consumer lines — XBOX, Microsoft 365 Consumer, Windows and Devices, and part of Search — together account for roughly a fifth of revenue and are flat to declining in aggregate. No customer and no country other than the United States accounted for more than 10% of revenue in FY2026, FY2025 or FY2024 (FY2026 10-K, Note 19). The United States supplied $170,794m of the $331,839m total.

The one exception to that diversification is a related party. Microsoft recorded $24.1bn of revenue from commercial arrangements with OpenAI in FY2026, inclusive of revenue-sharing payments, and carried $6.0bn of accounts receivable from OpenAI at 30 June 2026 (FY2026 10-K, related-party disclosure). That is 7.3% of total revenue and 7.4% of gross receivables from a single counterparty in which Microsoft holds an approximate 25% as-converted interest [computed]. Microsoft does not disclose what the comparable figure was in FY2025, so the contribution of this counterparty to FY2026 growth cannot be established from the filings.

What is being exited, and what that reveals

Two lines are shrinking and the reasons differ. XBOX revenue fell 7% to $21,790m, with hardware down 29% on lower console volume, and the FY2026 10-K refers three times to "impairment and other related expenses in our XBOX business" without ever quantifying the charge. Windows and Devices fell 1%, with first-party Devices declining and Windows OEM licensing growing 5%. In both cases Microsoft's own commentary attributes the segment's improving margin to the shrinkage: More Personal Computing gross margin percentage rose "driven by sales mix shift to higher margin businesses" (FY2026 10-K). A hardware line contracting while segment margin expands is the clearest evidence available that Microsoft's economics come from software and services and are diluted, not enhanced, by selling physical products.

From fiscal year 2027 the reporting changes again. On 2 September 2026 Microsoft announced it will report two segments — Agents and Infra, and Devices and Consumer — replacing the three used throughout this document (8-K furnished 2 September 2026). The recast historical data accompanying that announcement was not available for this analysis. Every segment figure here is on the basis Microsoft used through FY2026 and will not be directly comparable to what it publishes next.

3. Industry, Competitive Position and Moat

The cloud industry exists because computing has fixed costs that most organisations cannot use efficiently. A datacenter is only economic at scale, and almost no enterprise has enough demand to fill one. The industry aggregates that demand and rents the capacity back. What is actually sold is not technology but utilisation: the provider's margin is the difference between the price of a compute-hour and the cost of owning the machine that produces it.

Where the profit actually sits

Ranking the value chain by reported operating margin in the most recent quarter each participant has published makes the structure plain:

Sources: each company's own quarterly release or 8-K, calendar Q2 2026 except Oracle (FY2026, ended May 2026), NVIDIA (FQ2 FY2027) and Salesforce (FQ2 FY2027). Application-layer margins are GAAP; on a non-GAAP basis Salesforce and ServiceNow report 34.1% and 29.5%, the difference being stock-based compensation.

Three conclusions follow, and each one matters for Microsoft. First, the profit is concentrated upstream, at the two layers with the fewest substitutes: NVIDIA's 75% gross margin is not a curiosity, it is Microsoft's cost of goods. Second, the hyperscalers are the second-best-positioned layer, not the best — and Microsoft's Intelligent Cloud margin of 40.6% is the highest of the three disclosed cloud segments, so within that layer it is the strongest. Third, the merchant GPU providers show what rented compute earns without a second business attached: CoreWeave reports a 59% adjusted EBITDA margin and a −2% operating margin on 112% revenue growth, the gap being depreciation, and pays $640m a quarter in interest on $35bn of debt (CoreWeave 8-K, 11 August 2026). That is the economics of Microsoft's Azure business run without Microsoft's balance sheet or its software franchise.

Market structure and what the share data shows

Synergy Research put cloud infrastructure services at $143.4bn in the second quarter of calendar 2026, growing 43% year on year, with Amazon at 28%, Microsoft at 20% and Google at 15% (Synergy Research, 30 July 2026). Against Synergy's own Q3 2025 edition — Amazon 29%, Microsoft 20%, Google 13% — the reading is specific: over three quarters Google gained about two points, Amazon lost about one, and Microsoft was flat, while the Big Three's combined share stayed at 63% in a market growing 43%. Microsoft is holding position in a fast-growing market rather than compounding share within it.

That estimate carries an unavoidable caveat. Microsoft has never disclosed absolute Azure revenue in any 10-K from FY2016 to FY2026 — only growth rates. Azure sits inside the "Server products and cloud services" line and is never broken out. Every published Azure market-share figure is therefore a third-party reconstruction, and should be treated as an estimate rather than a measurement.

Which barriers to entry actually bind

• Power and grid interconnection — binds hard. Lawrence Berkeley National Laboratory's June 2026 queue study finds a median request-to-operation time of over five years for large loads, with only 13% of capacity requested between 2000 and 2020 having reached operation by the end of 2025 and 75% withdrawn. PJM's 2028/29 capacity auction cleared at the regulatory price cap and still fell 6,831 MW short of its reliability requirement — the first RTO-wide shortfall in its history (PJM, 14 July 2026). Chips arrive in eighteen months; a substation does not.
• Contractual lock-in — binds hard, and is the real moat. The UK Competition and Markets Authority's final report of 31 July 2025 named three structural frictions in cloud competition: egress fees, committed-spend agreements, and Microsoft's software licensing practices. That is an independent regulator finding that what keeps customers in place is commercial architecture rather than engineering.
• Chip and advanced-packaging allocation — binds hard. NVIDIA carried $119bn of manufacturing purchase commitments at its most recent quarter end; that queue is effectively pre-booked, and position in it is a barrier no amount of capital can jump.
• Capital scale — binds less than it sounds. CoreWeave reached 1.5 GW of active capacity, 3.7 GW contracted and a ~$104bn revenue backlog on borrowed money. Capital is available to newcomers. The question is the cost of it, not the availability.

Microsoft's specific position

Microsoft names no individual competitor anywhere in the FY2026 or FY2025 10-K. It describes competition entirely by category — "cloud service providers and open-source offerings", "hyperscalers", "AI-first application companies". That is a disclosure choice, and it means the filings cannot be used to establish relative position at all.

What can be established is the asset that would be hardest for a well-funded competitor to reproduce, and it is not the datacenters. Any of four companies can build datacenters. What Microsoft has that they do not is the incumbent position inside the enterprise administrative stack: the identity directory that defines who each employee is, the licensing agreement renegotiated once every three years rather than continuously, and the productivity suite already installed on the desk. The financial signature of that position is analysed in Section 4 — most of Microsoft 365's growth now comes from selling more to customers already under contract rather than from finding new ones. A competitor can match the compute. It cannot easily be the vendor whose agreement is already signed.

Testing Microsoft's own claims against outside evidence

The industry rewards a specific kind of company: one that can fund the build from operating cash rather than debt, one whose contractual position with enterprise buyers survives a price war in compute, and one with a second profit pool that pays the depreciation bill if the first disappoints. Microsoft satisfies all three tests — $155.2bn of operating income against $115.9bn of capital expenditure, funded internally, versus Oracle's negative $23.7bn free cash flow and $43bn of new debt in the same period. The two places outside evidence does not validate its self-description are share, where it is holding rather than gaining, and silicon, where it is a customer rather than a manufacturer.

4. Growth Engine

Revenue rose $50,115m, or 18%, in FY2026. Decomposing that increase is straightforward because there is almost nothing acquired in it: Microsoft spent $1,743m on acquisitions net of divestitures during the year, against $69,132m in FY2024 (FY2026 10-K, cash flow statement). FY2026 growth is essentially entirely organic. Foreign currency contributed a favourable 2% to both revenue and operating income.

That is a material contrast with the recent past and it should not be read backwards. FY2024's 13.6% reported growth included Activision Blizzard, acquired for $75.4bn and closed on 13 October 2023 — the largest acquisition in the company's history. A headline growth rate that is partly acquired describes a different business from the same rate grown, and Microsoft's headline rate has now switched from one to the other.

The drivers, ranked

• Azure consumption — structural. Intelligent Cloud contributed $31,526m of the $50,115m increase, 63% of total growth [computed]. Azure and other cloud services grew 41%, and server products — the on-premises remainder — grew 1%. This is the engine, and its mechanism is that customers buy more compute-hours than they did last year, not that Microsoft raised the price of a compute-hour.
• Revenue per Microsoft 365 seat — structural, and the highest-quality driver in the business. Microsoft 365 Commercial cloud revenue grew 17% on 6% seat growth, which means roughly ten points came from price and mix — Microsoft's own explanation names Microsoft 365 Copilot and Microsoft 365 E5. Selling more to a customer already under contract requires no new datacenter, which is why Productivity and Business Processes gross margin percentage rose in the same year that the company's overall gross margin fell.
• Seat count — structural but slow. 6% growth, driven by small and medium businesses and frontline worker offerings. The knowledge-worker installed base is largely penetrated; incremental seats now come from the parts of the workforce that were never issued a licence.
• Consumer subscription — structural, small. Microsoft 365 Consumer cloud revenue grew 28% on 7% subscriber growth. The same price-and-mix mechanism as the commercial line, on a base a tenth the size. Note that Microsoft removed the Microsoft 365 Consumer subscriber metric from its disclosed metrics in Q1 FY2026.
• Search advertising — cyclical with a structural element. Revenue grew 9%, and 12% excluding traffic acquisition costs, on higher search volume, higher revenue per search and third-party partnerships. Advertising is the most economically sensitive revenue line Microsoft has.
• LinkedIn and Dynamics — structural, steady. +11% and +15% respectively, with Dynamics 365 at +18%. Together they added $3.2bn.
• XBOX — declining, management-driven. The only line that subtracted from growth. Content and services fell 5% against a prior year that benefited from strong first-party performance, offset in part by Game Pass growth; the hardware retreat described in Section 2 accounts for the rest.

What the backlog says, and what it does not

Commercial remaining performance obligation — contracted revenue not yet recognised — rose 84% to $678bn, against Microsoft Cloud revenue of $214.4bn. That is 3.2 times a year of cloud revenue already under contract [computed]. It is the strongest single piece of evidence that demand is real and durable.

Two qualifications matter. The weighted-average duration is approximately 2.3 years and Microsoft expects to recognise only about 30% of it within twelve months — down from about 40% a year earlier, about 45% in FY2023 and about 60% when the disclosure began in FY2018. The backlog is lengthening as it grows, which means a smaller proportion of it is near-term revenue and a larger proportion is a promise about the middle of the decade. And remaining performance obligation is a contract, not cash: it converts only if the counterparty consumes and pays.

5. Margin, Cash and Capital Allocation

Sources: FY2020, FY2023, FY2025 and FY2026 10-Ks. Free cash flow and the conversion ratio are computed as operating cash flow less additions to property and equipment. Comparability note: FY2020 and FY2023 predate the August 2024 segment recast and are on a different segment basis, though the consolidated figures shown here were never restated. FY2021 and FY2023 operating margins each include a benefit from a change in the estimated useful life of server and network equipment — $2.7bn and $3.7bn of operating income respectively — so the margin trend across those years is partly an accounting effect.

Two margins moving in opposite directions

Microsoft's consolidated operating margin of 46.8% is the highest in its history, and it is misleading read alone, because it is the average of two divergent trends. Productivity and Business Processes operating margin went from 53.2% in FY2023 to 59.9% in FY2026. Intelligent Cloud gross margin went from 66.9% to 58.0% over exactly the same period, and its operating margin from 43.2% to 41.3%. The consolidated figure rises because the software segment is improving faster than the infrastructure segment is deteriorating — not because the infrastructure business is getting better.

Microsoft explains the cloud decline consistently and credibly: gross margin percentage fell "driven by continued investments in AI infrastructure and growing AI product usage, offset in part by efficiency gains". The mechanism is simple. Depreciation on newly-installed capacity enters cost of revenue immediately, while the revenue from that capacity arrives over the following years. Microsoft Cloud gross margin has now fallen from a 72% peak in FY2023 to 66%, a decline of six points in three years.

Against that, the operating expense line is the most durable good news in the financial statements. Sales and marketing spend grew 82% in absolute terms over eleven years while revenue grew 289%, taking it from 17.2% of revenue to 8.0% — a release of more than 900 basis points. Research and development fell from 13.2% to 10.7% of revenue on the comparable basis while growing 197% in absolute dollars. Headcount fell by 5,000 in FY2026 to 223,000, the first decline in eleven years, while revenue grew 18%; revenue per employee has risen from roughly $800,000 to roughly $1.49m [computed]. This is genuine operating leverage in a business that sells the same product to more people.

The cash conversion problem

Free cash flow was $66,987m in FY2026, against $71,611m in FY2025 and $74,071m in FY2024 [computed]. Microsoft's free cash flow has now declined for two consecutive years while net income rose 52% over the same period. Cash conversion — free cash flow as a percentage of net income — has fallen from 102% in FY2020 to 84% in FY2024 to 50% in FY2026.

The cash flow statement understates the capital being committed in two ways. Microsoft obtained $24,608m of right-of-use assets under finance leases in FY2026, which is capital expenditure in economic substance but does not appear as cash capex (FY2026 10-K, Note 13). And purchases of property and equipment still sitting in accounts payable rose from $6.9bn to $26.7bn (Note 6). Adding finance-lease additions to cash capex gives roughly $140.6bn of capital committed in one year, 42.4% of revenue [computed], and reduces free cash flow after finance leases to about $42.4bn — against $48,716m returned to shareholders in dividends and buybacks. Cash and short-term investments fell $17.7bn during the year, to $76,843m.

Where the cash has gone over eleven years

Ranked, FY2016 through FY2026 (summed from the cash flow statements; no cash flow figure was ever restated):

The ranking has inverted. For most of the eleven years Microsoft was a company that returned more cash than it invested; cumulative shareholder returns of $412.5bn represent 62% of cumulative free cash flow of $665.6bn. In FY2026 capital expenditure alone ($115.9bn) exceeded dividends and buybacks combined ($48.7bn) by a factor of 2.4. Management has redirected the company's cash from shareholders to the balance sheet, has been explicit about doing so, and has funded it from operating cash flow rather than debt — a distinction that separates Microsoft from every other large builder in the industry.

What is committed but not yet spent

The contractual obligations disclosure is where the scale of the commitment becomes visible, and it nearly doubled in one year:

Sources: FY2026 and FY2025 10-Ks, MD&A contractual obligations table and Note 13. Purchase commitments "primarily relate to datacenters and include open purchase orders and take-or-pay contracts". Leases not yet commenced at 30 June 2026 are stated to commence between fiscal 2027 and fiscal 2033.

Adding the $329.1bn of leases signed but not yet commenced to the $743.8bn of contractual obligations gives roughly $1.07 trillion of committed future outflow, against $442.4bn of shareholders' equity and $331.8bn of annual revenue [computed]. The trajectory of the not-yet-commenced figure through the year is the sharpest single series in the filings: $92.7bn at June 2025, $106.2bn at September 2025, $196.6bn at March 2026, $329.1bn at June 2026. One qualification is new in the FY2026 wording — the leases are described as "with some arrangements subject to certain contractual conditions being met", a clause absent in FY2025. Microsoft does not say what proportion is conditional.

6. Cyclicality, Constraints and What to Monitor

Where the business sits in its own cycle

Microsoft's revenue is unusually well insulated, and its capital position is not. On volume and price, the business is closer to a peak than a trough on every measure it discloses. Azure growth of 41% in FY2026, and 43% in the June quarter, is a re-acceleration from a cyclical trough of 29% in FY2023 — though far below the 113% of FY2016, since the base is now vastly larger. The eleven-year Azure growth series runs 113%, 99%, 91%, 72%, 56%, 50%, 45%, 29%, 30%, 34%, 41%: a long structural deceleration, a trough in FY2023, and two years of reacceleration since.

Operating margin at 46.8% is an all-time high, and capital expenditure at 34.9% of revenue is more than triple the 9.8% of FY2016. The two records were set in the same year, and that combination is the cycle position: the margin describes capacity installed years ago and already paid for, while the capital intensity describes capacity whose revenue has not yet arrived. A record consolidated margin recorded at a record level of capital intensity is a different fact from a record margin recorded at a low one.

The revenue base itself is genuinely defensive. Roughly $72,965m of short-term unearned revenue sits on the balance sheet, invoiced and collected in advance; $678bn of commercial backlog is contracted; and the volume licensing model means most enterprise customers renegotiate once every three years rather than continuously. Microsoft's own history supports this: FY2023 was a weak year for the technology sector and Microsoft's revenue still grew 7%, with Windows OEM revenue down 25% while the commercial cloud grew 22%. The consumer and OEM lines are cyclical; the commercial contract base is not.

The constraints that actually bind

Microsoft's FY2026 risk disclosures are unusually specific about physical limits, and the outside data confirms them. "Our datacenters depend on the availability of permitted and buildable land, predictable energy, networking supplies, and servers, including graphics processing units." The filing adds a category that did not appear in prior years: "community opposition, state and local moratoriums, and hyper-local dissent, as well as increasingly coordinated opposition to infrastructure development across jurisdictions".

The independent evidence is consistent. Lawrence Berkeley National Laboratory's June 2026 update puts US datacenter electricity use at 192 TWh in 2024, 4.7% of US electricity, rising to roughly 235 TWh in 2025, with a 2030 reference case of 649 TWh or 11.8% — a forecast, and one whose own authors note that newer shipment data projects up to 50% fewer GPU units than earlier estimates. On the supply side, Texas suspended large-load interconnection studies on 3 August 2026 pending a statewide audit, with no completion date set, against roughly 474 GW of pending large-load requests in ERCOT of which about 90% are datacenters. The EIA responded by cutting its forecast for Texas 2027 electricity demand growth from 14% to 6%. Power, not chips, is the constraint that determines whether committed capital becomes revenue on schedule.

The downside mechanism

The relevant historical analogue is the telecommunications buildout of 1996–2001, and its mechanism transfers precisely. Communications equipment investment grew from about $62bn to over $135bn in constant dollars over five years, then fell to 69% of the prior year in a single year; telecom employment fell 22% from its March 2001 peak; and roughly $700bn of market value was lost, over 3.5% of total US corporate equity value at the peak (Federal Reserve Bank of Richmond, Spring 2003). Meanwhile aggregate real US investment in information processing equipment and software fell only 0.4% in 2001 and 2.9% in 2002, and had fully recovered by 2003 (BEA). The buyers of technology were largely fine. The builders of the capacity were not.

Applied to Microsoft, the mechanism is a depreciation schedule meeting a demand curve that slows. Servers and network equipment are carried at $215,874m gross and depreciated over two to six years. Depreciation expense was $34.3bn in FY2026, against $22.0bn in FY2025 and $15.2bn in FY2024. If the average depreciable life on that base proved to be five years rather than six, annual depreciation would rise by roughly $7bn; at four years, by roughly $18bn — 4.6% and 11.6% of FY2026 operating income respectively [computed, illustrative: applies a single life to the full gross balance]. Microsoft's own risk factor states the consequence directly: "Overestimation of demand or misalignment of capacity investments may result in underutilisation of infrastructure and may lead to impairment of assets on our balance sheet."

Durable versus borrowed

Leading indicators, and where each is published

• Microsoft Cloud gross margin percentage — quarterly, in Microsoft's earnings release and 10-Q MD&A. It fell from 72% to 66% in three years; whether it stabilises is the central question about the economics of the build.
• The proportion of commercial remaining performance obligation recognisable within twelve months — annually in the 10-K, Note 11. It has fallen from about 60% in FY2018 to about 30%. A further fall means the backlog is lengthening faster than it is converting.
• Leases not yet commenced — quarterly, in the leases note of each 10-Q and 10-K. This is the forward capital commitment, and it moved from $92.7bn to $329.1bn in one year.
• Free cash flow after finance-lease additions — computed from the cash flow statement and the leases note. Cash capex alone now understates capital committed by about $25bn a year.
• Grid interconnection queue times and regional moratoriums — LBNL's annual "Queued Up" report, FERC filings, and ISO announcements such as ERCOT's August 2026 pause. This determines the timing of revenue from capital already committed.
• The CMA's strategic market status decision on Microsoft's business software ecosystem — statutory deadline February 2027, published by the CMA. It covers Windows, Office, Teams, Copilot, server software, security and identity.
• US Census Bureau Business Trends and Outlook Survey AI-use series — biweekly. It is the only government measurement of whether enterprise AI adoption is broadening; it stood at 19.8% on 3 May 2026 and is flat for firms under twenty employees.

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality and capacity constraints are covered in Section 6. What follows is what cyclicality does not capture, ordered by how much each compounds with the others.

• The OpenAI relationship changed in FY2026, and the filing does not explain why. The FY2025 10-K stated that "the OpenAI API is exclusive to Azure" and that Microsoft held "a right of first refusal on OpenAI's new capacity needs". Both sentences are absent from the FY2026 10-K, and "reciprocal revenue-sharing arrangements" became "will continue to receive revenue-sharing payments". The filing does not say whether these terms were renegotiated, expired, or merely dropped from disclosure. In the same year OpenAI became a $24.1bn customer, a $6.0bn receivable, and an equity-method investee whose recapitalisation produced a $6.5bn pre-tax gain that Microsoft created a new non-GAAP measure specifically to exclude. The mechanism that matters: if that counterparty's compute purchases slow or move elsewhere, roughly 7% of revenue, a share of Azure growth that cannot be sized from the filings, and a $6.0bn receivable are all exposed at once — and Microsoft's own risk factor concedes it, noting that strategic partners "are significant customers of Azure" and that expected consumption "may not materialise, may be delayed or reduced".
• The moat and the regulatory exposure are the same object. Section 3 identified contractual lock-in as the real barrier to entry. Three authorities are currently examining precisely that. The CMA opened a strategic market status investigation into Microsoft's business software ecosystem on 14 May 2026 — explicitly covering Windows, Office, Teams, Copilot, server software, security and identity — with a statutory decision due in February 2027. The European Commission opened a market investigation in November 2025 into whether Azure should be designated a gatekeeper under the Digital Markets Act despite not meeting the quantitative thresholds. The FTC was reported in February 2026 to have issued civil investigative demands to at least six Microsoft competitors covering licensing terms and the cost of running Microsoft software on rival clouds. A remedy that unbundled licensing or capped egress and committed-spend terms would attack the mechanism, not the margin — which is why it compounds rather than merely costs.
• The depreciable life is a judgement made before the current hardware generation. Microsoft set the six-year life for server and network equipment in July 2022 and has not revisited it, while capital expenditure has grown from $28.1bn to $115.9bn. The two prior changes to this estimate both increased reported profit — by $2.7bn of operating income in FY2021 and $3.7bn in FY2023 — and the FY2023 10-K concedes that excluding that year's change, gross margin percentage decreased by a point. The estimate has only ever moved in the direction that helps. A reassessment in the other direction would be the first, and it would land on a gross asset base of $215,874m.
• A meaningful share of industry demand is financed by its own suppliers. NVIDIA holds $42.3bn of non-marketable equity securities in its own ecosystem and $30bn of commitments to buy cloud capacity from its customers; Amazon booked $53.4bn of other income in one quarter primarily from its Anthropic investment; Microsoft booked $3.2bn from the same holding. Anthropic committed to $30bn of Azure compute alongside investment from Microsoft and NVIDIA. Microsoft's own disclosure that roughly 90% of its $678bn commercial backlog comes from customers outside frontier-model companies is the only quantified bound any major has published, and it is reassuring. But it also implies roughly $68bn that is not, and the equivalent figure for its competitors' backlogs is not published at all.
• The IRS transfer-pricing claim remains unresolved after three years. The IRS is seeking $28.9bn of additional tax plus penalties and interest for tax years 2004 to 2013, with tax years 2014 to 2017 still under audit. Microsoft states its allowances are adequate and carries $28,647m of long-term income taxes. No settlement, payment or filing is disclosed. Deloitte's critical audit matter notes that resolution of the remaining matters "could have a material impact".
• Incentive pay is measured entirely on growth, not on the return from the capital that produces it. The chief executive's equity award is 100% performance stock, and all four of its metrics are revenue-growth metrics — Azure growth at 35%, other Microsoft Cloud growth at 35%, consumer services at 15%, XBOX content and services at 15%. The cash incentive's two financial metrics are revenue and operating income. There is no margin, return-on-capital, free-cash-flow or capital-efficiency metric anywhere in either plan (2025 proxy statement). In a year when capital expenditure reached 34.9% of revenue, that is a structural asymmetry worth naming.

What the sources could not answer

• Absolute Azure revenue in any year. No 10-K from FY2016 to FY2026 discloses it — growth rates only. This makes every Azure margin, share and return calculation an estimate rather than a measurement. Microsoft has been reported as intending to begin disclosing it from FY2027; that report is not primary-sourced.
• FY2025 revenue from OpenAI. Only FY2026's $24.1bn is disclosed, so OpenAI's contribution to FY2026 growth cannot be isolated.
• Any dollar figure for OpenAI's Azure compute commitment. The $194.1bn of purchase commitments is described only as relating "primarily to datacenters" with no counterparty attribution.
• The size of the FY2026 XBOX impairment. It is named three times in the MD&A and never quantified; Note 9 confirms it is not an intangible-asset impairment and goodwill was essentially unchanged.
• What proportion of the $329.1bn of not-yet-commenced leases is subject to the "certain contractual conditions" clause introduced in FY2026.
• Forward capital expenditure guidance. Microsoft provides it on the earnings call rather than in the release or the 10-K, and it could not be confirmed from a primary source for this document.
• Recast historical financials on the two-segment FY2027 basis announced on 2 September 2026. The exhibit was furnished to the SEC but its content was not available here.

Three questions would need resolving before an investor could form a view. What is Azure's actual revenue base and unit margin, without which the return on $355bn of eleven-year capital cannot be computed? What are the current terms of the OpenAI arrangement, given that two material terms disappeared from disclosure in one year? And what does Microsoft believe the useful economic life of an AI accelerator to be, as distinct from the six-year accounting life set in 2022?

8. Investor Takeaways

• This is a software annuity that has bolted a capital-intensive utility onto itself. The seat business earns a 59.9% operating margin and needs no capital; the compute business earns 41.3% on a gross margin falling three points a year. Both are inside one income statement, and the consolidated margin hides the divergence.
• The economic engine is the incumbent position, not the datacenters. Microsoft 365 Commercial cloud revenue grew 17% on 6% seat growth — ten points of price and mix sold into customers already under contract, requiring no incremental capital. That is what a moat looks like in the financial statements.
• The main growth lever is Azure consumption, and it is working. Intelligent Cloud supplied 63% of FY2026's revenue growth, Azure grew 41%, and commercial backlog rose 84% to $678bn. FY2026 growth was almost entirely organic.
• What could break the story is the gap between when the capital leaves and when the revenue arrives. Cash conversion has halved to 50% of net income; roughly $1.07 trillion of future outflow is committed; and the historical precedent for this pattern is a buildout in which the buyers of the technology were fine and the builders were not.
• Watch three numbers: Microsoft Cloud gross margin, the twelve-month conversion rate of the backlog, and leases not yet commenced. They are the earliest available evidence on whether $115.9bn a year of capital is buying an annuity or an inventory.