Microsoft Corporation

01 — Executive summaryRevenue up 18%. Operating income up 18%. Net income up 31%. Only two of those are the same business.

Microsoft closed fiscal 2026 with revenue of $90.0 billion in the June quarter, up 18%, and $331.8 billion for the year, also up 18%. Operating income rose 21% to $155.2 billion. Azure and other cloud services grew 43%, the fastest rate since early 2022, and crossed $100 billion in annual revenue for the first time. Microsoft Cloud reached $59.3 billion in the quarter, up 27%, at a 65% gross margin. Commercial remaining performance obligation — contracted work not yet billed — closed the year at $678 billion, up 84%. Microsoft 365 Copilot passed 30 million paid seats, having doubled in two quarters.

Those are the facts, and they are not in dispute. What is in dispute is which of them describe a business and which describe an accounting year.

The reported net income figure of $133.7 billion, up 31%, grew roughly twice as fast as operating income for a specific reason that the company disclosed itself: a full-year net gain of $5.0 billion on its OpenAI investment, plus a $3.2 billion gain on its Anthropic stake in the June quarter. Strip the non-operating items and the clean comparison is the non-GAAP line — net income up 22% — which is a better number than the headline, not a worse one, because it is the one that repeats.

Our question in this report is narrower than whether Microsoft is a good company. It plainly is. The question is this: Microsoft has committed roughly $175 billion of calendar 2026 capital expenditure to assets whose economic life it has just declared to be 67% longer than it believed a year ago. What has to be true for that to be the right call, and does the current price already assume it?

1. The operating performance is the best in the company's history, and it is broad

Revenue growth has held in a 15–18% band for twelve consecutive quarters. At $331.8 billion of annual revenue that is not a growth rate; it is a compounding machine, and it is the most unusual fact in the large-cap technology complex. Microsoft Cloud grew 27% in the year to $214.4 billion — 64.6% of total revenue, up from 46% in FY2022. Operating margin expanded 470 basis points over four years to 46.8%.

The composition matters more than the headline. Productivity and Business Processes — Microsoft 365, LinkedIn, Dynamics — generated $83.9 billion of operating income on $140.0 billion of revenue in FY2026, a 59.9% segment margin. That single segment produced more profit than the entire company did in FY2023. It is the franchise that funds everything else, and it grew operating income at 17% while adding Copilot cost into the same segment.

Intelligent Cloud is the growth engine and the capital sink simultaneously: $137.8 billion of revenue, $57.0 billion of operating income, a 41.4% margin. Azure is accelerating, not plateauing, against every published expectation going into the quarter. And More Personal Computing, at $54.1 billion, has now declined for two consecutive years — down 4% in the June quarter, with Xbox content and services down 10% and Windows OEM down 7%.

2. The backlog is the strongest evidence and the largest single concentration

Commercial RPO of $678 billion is equivalent to more than two full years of total company revenue, and it grew 84% in twelve months. Weighted average duration is 2.3 years, with roughly 30% converting to revenue inside the next twelve months — call it $200 billion of contracted work flowing through fiscal 2027 before a single new contract is signed. Management also disclosed that the entire $51 billion sequential increase came from customers other than the large AI model developers, and that the backlog grew 25% excluding OpenAI.

That disclosure cuts both ways, and the second edge is the one the market has not fully priced. Working the arithmetic backwards: 25% growth applied to the prior-year $368 billion balance gives a non-OpenAI backlog of roughly $460 billion, leaving approximately $218 billion — about 32% of the total — attributable to OpenAI. Microsoft disclosed in January 2026 that 45% of the then-$625 billion balance was tied to OpenAI. The direction of that ratio is improving. The level is still a third of the company's contracted future revenue resting on one privately held counterparty that has never reported audited financials.

3. Capital expenditure has nearly quadrupled and free cash flow has not kept up

Microsoft spent $115.9 billion on property and equipment in FY2026, up 80% from $64.6 billion, and added 88 data centres. Operating cash flow grew 34% to $182.9 billion — an excellent result — but capital expenditure grew faster, and free cash flow fell 6.5% to $67.0 billion. Free cash flow per share was $9.02 against trailing GAAP earnings per share of $17.95. The company is now converting 63 cents of every operating dollar into property and equipment, against 32 cents three years ago.

Management has been explicit that this continues: FY2027 capital expenditure is guided to grow again, with more than $50 billion in the September quarter alone. Amy Hood has also made the strongest available argument that this is discretionary rather than committed: two-thirds of the spend is on short-lived assets — CPUs and GPUs — with short lead times, which means a demand shock can be absorbed by slowing procurement rather than by writing down concrete. That is a real option and it deserves to be weighed. It is also an option whose exercise would reduce the revenue that justifies the share price.

4. The useful life change is the most consequential accounting decision in the report

Effective at the start of FY2027, Microsoft extended the estimated useful life of its data centres and office buildings from 15 years to 25 years. The company's explanation is operating history and expected use, and it is entirely legitimate to extend an asset life when evidence supports it. The timing deserves scrutiny: the change occurs in the first year in which the depreciation from three years of unprecedented building begins to flow through the income statement.

Management stated that the effect on FY2027 operating income will be minimal, because buildings are a small share of the depreciable base. The larger effect is on capital expenditure as measured — a portion of data centre leases shifts from finance leases (counted in capex) to operating leases (not counted), reducing the guided calendar 2026 figure to approximately $175 billion. On a like-for-like basis, the capital programme is larger than the headline suggests, and the capex-to-revenue ratio is higher than the reported number implies.

The distinction we would draw is this: buildings depreciated over 25 years is a defensible judgement. The GPUs inside them are depreciated over roughly six, and the useful-life change does nothing to lengthen the life of the equipment that generates the revenue. Two-thirds of the capital programme is the part with the short life.

5. On operating earnings the shares are expensive; on the balance sheet they are not

At $497.75 Microsoft trades at 27.7× trailing GAAP earnings and 25.2× forward. Our estimate of FY2026 operating earnings per share, excluding the non-operating investment gains, is $24.33, which puts the shares at roughly 20.5× current-year operating earnings and 21.2× our FY2027 base case of $23.42. Those are reasonable multiples for a business growing revenue at 18% with a 47% operating margin, and materially cheaper than the 27× operating multiple we calculate for Amazon on the same basis.

Unlike Amazon, Microsoft has not funded the build with debt. Long-term debt fell 22.6% to $31.1 billion while book equity rose 28.8% to $442.4 billion. Debt-to-equity is 0.29; interest coverage is roughly 50×. Return on invested capital of 26.2% still exceeds our estimated 10.1% cost of capital by more than sixteen percentage points. The capital is being spent, but the balance sheet that funds it is not under strain.

That is the fundamental difference between this report and our work on Amazon: Amazon is converting a fortress balance sheet into a leveraged capital cycle. Microsoft is running an enormously expensive capital cycle out of retained earnings and still returning $48.7 billion a year to shareholders. The risk profile is lower. The valuation reflects that.

6. The two models are converging, and Microsoft is the one giving up exclusivity

In April 2026 Microsoft and OpenAI amended their partnership for the third time in eighteen months. Microsoft's IP licence to OpenAI models runs to 2032 but is no longer exclusive; OpenAI may serve all of its products on any cloud; the revenue share Microsoft pays to OpenAI has been eliminated, OpenAI's share to Microsoft continues through 2030 subject to a cap; and the AGI clause — which would have triggered a change in Microsoft's rights on a board declaration — has been removed entirely. Microsoft shares fell about 2% on the announcement.

Read generously, this converted a philosophical bet on an undefined threshold into a normal commercial contract with a fixed sunset, and Microsoft kept the Azure-first deployment posture, the $250 billion compute commitment, and a 27% equity stake. Read ungenerously, Microsoft traded away the exclusivity that made its AI position unreplicable, in exchange for certainty about a right it might otherwise have retained indefinitely. Both readings describe the same document. The market's immediate reaction was the second.

Since then Microsoft has leaned harder into its own stack: the Maia 200 inference accelerator, the MAI family of in-house models spanning image, voice, transcription, coding and security, and an explicit public position advising enterprises not to depend on any single frontier lab at the application layer. That advice applies as much to Microsoft's own dependency as to anyone else's.

What would change our mind

We would move to a more positive rating on any of three developments: Azure growth holding above 45% for two consecutive quarters while capital intensity, measured as capex to revenue, stops rising; a disclosure that reduces the OpenAI share of the backlog below 20% without a corresponding decline in bookings growth; or evidence that Copilot seat growth is translating into measurable revenue per seat rather than into bundled licence upgrades at no incremental price.

We would move to a negative rating on any of three others: a quarter in which the backlog declines sequentially without a stated explanation; a depreciation restatement or a further extension of useful life; or evidence that the OpenAI counterparty risk is crystallising — most cleanly through a restructuring of the compute commitment or a material revision to the contracted profile.

02 — Company & business modelA software annuity wrapped around a hardware factory

Microsoft is best understood as two businesses with opposite economics bolted together, which is why almost every debate about the stock is really a debate about which half is setting the price.

The first is a subscription software annuity. Roughly 464 million people use Microsoft 365 commercially; Microsoft 365 Commercial seat growth has been a steady 6% year over year for eight consecutive quarters; renewal rates are not disclosed but the segment margin of 59.9% tells you what happens when a customer base renews at near-total rates at prices that rise on a schedule. Windows, Office, Teams, Entra, Defender, Dynamics, LinkedIn — this is a portfolio of products where the customer's switching cost is measured in years of integration work, and where the marginal cost of serving an additional seat is close to zero.

The second is a capital-intensive infrastructure factory. Azure sells compute, storage and increasingly inference by the hour, at a gross margin that is good but structurally lower than software, and it requires the construction of buildings, the purchase of accelerators, and the financing of a supply chain measured in years. Microsoft Cloud gross margin was 65% in the June quarter, down from 68% a year earlier and 72% in FY2022. That decline is partly mix and partly the cost of serving AI workloads, which is a real economic difference and not an accounting artefact.

The strategic insight that has driven Microsoft for the last decade is that these two businesses can be made to feed each other. The software annuity generates the cash; the factory converts it into capacity; the capacity is then inserted into the annuity through Copilot, which raises the price of the annuity without adding a new customer. This is why the E7 Frontier Suite at $99 per user per month, launched in May 2026, is the most strategically interesting product Microsoft has shipped in years. It is not a new product category. It is a price increase dressed as a bundle, sold to an installed base that mostly will not leave.

The reporting structure, and its limitations

Microsoft reports three segments, and the boundaries between them have become progressively less informative as AI has spread across all three.

Two limitations are worth stating plainly. First, segment operating income excludes corporate-level costs, so the three segments sum to $155.3 billion against a consolidated $155.2 billion only by coincidence — the corporate allocation adjusts the total. Second, and more importantly, Copilot revenue is booked inside Productivity, while the infrastructure that serves Copilot is booked inside Intelligent Cloud. There is no disclosure that allows an outside analyst to determine the margin on an AI workload. The company has effectively placed the AI cost in one segment and the AI price in another, and both segments look fine.

Where the growth has come from

Over five years Microsoft has added $151.6 billion of revenue. Microsoft Cloud accounted for $123.2 billion of that — 81% of all growth — against $28.4 billion from everything else combined. The company that was once described as a mature enterprise software vendor has become a cloud provider with a software attach.

What that shift has done to the profit and loss account is visible in the margin structure, and it is more subtle than the growth numbers suggest. The operating margin has expanded 470 basis points over four years while the gross margin has contracted 50 basis points. The mechanism is that research and development and sales and marketing have grown more slowly than revenue, while cost of revenue has grown faster. Microsoft has been converting a high-margin software licence business into a slightly lower-margin cloud business and then taking the difference out of operating expense.

That is a rational strategy, and it has a finite runway. Operating expense was 21.6% of revenue in FY2026 against 25.4% in FY2022. It cannot halve again. Once operating leverage stops contributing, margin expansion has to come from the gross margin line, and the gross margin line is where the cost of inference and the cost of depreciation sit.

The three questions this report is trying to answer

Everything that follows is organised around three questions, and we state them here so that the reader can hold us to them.

First, is the capital programme earning its cost of capital? Microsoft spent $115.9 billion on property and equipment in FY2026 and will spend more in FY2027. Return on invested capital is 26.2% and falling; it was 29.8% in FY2023. The spread over cost of capital has narrowed from roughly 20 percentage points to 16. That is still an enormous spread and it is the single most important reason to own the shares. The question is the second derivative.

Second, is the backlog a customer base or a counterparty? $678 billion of contracted revenue is only as good as the ability of the counterparties to pay it. A third of it traces to a company that has never filed audited accounts, is loss-making at the operating level, and has raised capital at successively higher valuations in a market that could reprice. This is not a reason to dismiss the backlog. It is a reason to discount it, and to ask by how much.

Third, is the accounting catching up or keeping up? Two decisions in FY2026 bear directly on this. Microsoft extended the useful life of data centres and office buildings from 15 to 25 years. It also reclassified a portion of data centre leases from finance to operating leases, which reduces reported capital expenditure without changing the underlying commitment. Neither is improper. Both make the reported numbers look better than the underlying economics, and the two together are worth approximately the difference between a capital programme that looks manageable and one that looks like a bet.

03 — The financial recordFive years, and the point at which the cash stopped coming out the other side

The FY2026 income statement is the best in Microsoft's history on every conventional measure. It is also the year in which the cash flow statement and the income statement began telling materially different stories, and that divergence is the analytical starting point for everything that follows.

Read the operating income line. It has grown from $83.4 billion in FY2022 to $155.2 billion in FY2026, a 16.8% compound rate, almost exactly matching revenue growth of 13.7% with a modest margin assist. This is a well-run company executing a plan, and the consistency of the margin expansion — 42.1%, 41.8%, 44.6%, 45.7%, 46.8% — is the signature of a business with genuine pricing power.

Now read the other income line. It went from negative $1.4 billion in FY2023 to positive $10.7 billion in FY2026. That $12.1 billion swing is larger than the entire operating income of most S&P 500 companies, and it is not operating performance. It is substantially the mark on the OpenAI and Anthropic stakes, plus interest income on a larger investment portfolio. In FY2026 the OpenAI investment alone contributed a net gain of $5.0 billion, or $0.67 per share.

The distinction is not academic. Microsoft trades at 27.7× trailing GAAP earnings because GAAP earnings are $17.95. On operating earnings of $24.33 the multiple is 20.5×. Reported earnings growth of 31% flatters a business growing operating income at 21%. The company disclosed all of this prominently and adjusted for it in its own non-GAAP presentation, which is to its credit. Our point is simply that the headline number the market quotes is not the number that repeats.

The cash flow statement, which is the more important document

Free cash flow has now declined for two consecutive years while operating cash flow has grown 54%. Free cash flow margin has fallen from 30.2% in FY2024 to 20.2% in FY2026, a ten-point compression in two years. The cause is unambiguous and is visible in a single ratio: capital expenditure was 32% of operating cash flow in FY2023 and 63% in FY2026.

The most interesting fact on this page, though, is the bottom line. Microsoft returned $48.7 billion to shareholders in FY2026 through dividends and buybacks, funded from free cash flow of $67.0 billion. The capital programme and the shareholder return programme are currently being funded from the same pool, and the pool is shrinking as a share of revenue. Microsoft has not had to choose between them, and it is not close to having to choose. But the arithmetic is worth watching: at guided FY2027 capital expenditure the company will spend more on property and equipment than it returns to shareholders for the first time in its history.

The positive reading of these ratios is that Microsoft is doing the most valuable thing a compounder can do: reinvesting at a return above its cost of capital rather than distributing cash it has no use for. Operating cash flow to net income of 1.37× means the reported profit is real and cash-backed. Interest coverage of roughly 50× means there is no financing risk. Return on invested capital of 26.2% against a 10.1% cost of capital means the reinvestment is currently creating value.

The negative reading is that all four of those ratios are moving in the same direction at once. Cash backing is down from 1.57× to 1.37×. The reinvestment ratio is up from 0.32 to 0.63. The capital spread is down from 20 points to 16. And free cash flow per share is flat to down while reported earnings per share compounds at 17%. Those are not independent observations. They are four views of the same fact.

04 — Segment deep diveIntelligent Cloud: the fastest $100 billion in the history of enterprise software

Azure and other cloud services passed $100 billion of annual revenue in FY2026, growing 41% for the year and 43% in the June quarter. It took Microsoft roughly a decade to get there. It is the second cloud platform in the world by revenue and the fastest-growing of the three at scale in the most recent quarter, ahead of Google Cloud at 28% on a larger base and well ahead of AWS at 18%.

The acceleration is the story. Azure grew 33% in the September 2025 quarter, 31% in December, 33% in March, and 43% in June. For a business at a $100 billion run-rate, reaccelerating from the low thirties to the low forties is not a normal competitive dynamic. It is what happens when a demand curve shifts faster than a supply chain can respond, and the supply chain in question belongs to the customer as much as to the vendor.

How the segment actually makes money

Intelligent Cloud reported $137.8 billion of revenue and $57.0 billion of operating income in FY2026 — a 41.4% segment margin, up from 39.4% in FY2024. That margin is worth dwelling on because it looks inconsistent with the idea that AI infrastructure is margin-dilutive.

The reconciliation is that Azure is not one business. It is at least four, with very different economics, and Microsoft discloses none of the split.

• Core IaaS and PaaS — compute, storage, databases, networking. Mature, high-utilisation, and improving as the installed base amortises. This is where the classic cloud operating leverage lives.
• Azure OpenAI inference — serving frontier models to enterprises. Low gross margin, capital-hungry, and currently the fastest-growing part of the mix. This is the part that dilutes.
• First-party AI applications — Copilot, GitHub Copilot, Security Copilot. These consume Azure capacity internally and are booked in other segments, so their infrastructure cost lands here and their revenue lands there.
• Microsoft 365 and Xbox infrastructure — an internal customer with a transfer price. Nothing about the economics is disclosed.

Because the mix within Azure is undisclosed, the 41.4% segment margin tells you less than it appears to. It is entirely possible for Azure's segment margin to be rising — as it is — while the margin on incremental AI revenue is below the segment average. In that case the reported margin improves on the base while the incremental economics deteriorate, and the only way to detect the difference from outside is to watch for the point at which segment margin stops expanding.

Capacity is the constraint, and it is deliberate

The strongest evidence that Microsoft is demand-constrained rather than supply-constrained is the guidance. For the September quarter Microsoft guided Azure to approximately 45% constant-currency growth, above the 43% just reported, and said on the call that it expects growth to accelerate in the first half of FY2027. A company guiding a $100 billion-plus business to accelerate in the low forties is telling you it has already sold capacity it has not yet built.

That is consistent with the backlog. RPO of $678 billion against Microsoft Cloud revenue of $214.4 billion implies contracted work equal to more than three years of cloud revenue at current run-rate. Management has said most of the FY2027 capacity is already reserved. And the company added 88 data centres in the fiscal year, with roughly two-thirds of capital expenditure going to the CPUs and GPUs that fill them.

05 — Segment deep diveProductivity and Business Processes: the asset that is not being debated

Productivity and Business Processes is the least discussed and most valuable part of Microsoft. In FY2026 it generated $140.0 billion of revenue — up 16% from $120.8 billion — and $83.9 billion of operating income, a 59.9% segment margin as reported by the company.

To put that in perspective: this single segment produced 54% of Microsoft's total operating profit, and its operating income alone would rank it among the thirty most profitable companies in the world. It has compounded operating income at 17% over two years while absorbing the incremental cost of Copilot across both the commercial and consumer cloud lines.

What is inside the number

The composition has shifted in a way that matters for the durability of the margin. Microsoft 365 Commercial cloud is now the largest component at an estimated $118.6 billion on a Farstar derivation from the disclosed cloud line; Microsoft 365 Consumer cloud grew 28% for the year, its fastest rate since the pandemic, on price increases and the Copilot Pro attach; LinkedIn grew 11%; Dynamics 365 grew 18%, decelerating notably from 22% in the March quarter to 13% in June.

The most important number in the segment is the one that has not moved. Microsoft 365 Commercial seat growth has been 6% year over year for eight consecutive quarters. That is the definition of a saturated market, and it is precisely what makes the productivity segment such a formidable asset: it is a nearly-fixed customer base to which Microsoft can sell new things at higher prices. The seat count is no longer the growth driver. The growth driver is revenue per seat.

LinkedIn, Dynamics, and the parts that are not Copilot

LinkedIn generated an estimated $19.4 billion in FY2026, growing 11%. It is a genuinely differentiated asset — the only scaled professional graph in existence — but it has been a mid-teens or lower grower for three years and there is no evidence that AI is changing that trajectory. It is a good business with a stable contribution, not a swing factor.

Dynamics 365 grew 18% for the year but decelerated to 13% in the June quarter, and Microsoft guided to "low teens" for September. This is the most competitive part of the productivity portfolio: it competes directly with Salesforce, which is simultaneously suing Microsoft in the London High Court over Teams bundling. Dynamics is a real business, and it is not on a path to displace the category leader.

The consumer business deserves a sentence. Microsoft 365 Consumer cloud grew 28% for the year, driven by a $3 price increase on the base subscription and the Copilot Pro attach. That is a meaningful contribution and it is also the most cyclically exposed line in the segment, because consumers can and do cancel. Its 5.7% share of total revenue means it is not a material swing factor either way.

06 — Segment deep diveMore Personal Computing: the part of Microsoft that is shrinking

More Personal Computing reported $54.1 billion of revenue in FY2026, down 1% from $54.6 billion — the second consecutive year of decline. Operating income of $14.4 billion was a 26.6% margin, down from 28.6% in FY2025.

Within the segment, every line declined in the June quarter: Windows OEM and devices fell 7%, Xbox content and services fell 10%, and only search advertising — up 10% excluding traffic acquisition costs — grew. Management guided Windows OEM to a low-twenties decline in the September quarter, which is not a guidance range so much as a description of a market that is contracting.

Gaming: the most expensive acquisition in Microsoft's history is being unwound

Microsoft's $68.7 billion acquisition of Activision Blizzard closed in October 2023. In the three years since, gaming has been restructured twice and is now being restructured a third time, and the scale of the reset is the clearest available evidence about the returns on that capital.

On 6 July 2026 Microsoft announced approximately 4,800 job eliminations across the company, of which 1,600 were at Xbox on the day and up to 3,200 in total over the coming year — approximately 20% of Xbox headcount. Four studios were divested: Ninja Theory and Undead Labs to new owners, and Compulsion Games and Double Fine Productions returned to their management teams with their IP. A fifth, Arkane France, entered works council consultation.

The memo from new Xbox CEO Asha Sharma is the most candid document Microsoft has published about any of its businesses, and one sentence in it deserves to be quoted verbatim rather than paraphrased:

"Our business today is not healthy. We are operating at margins that are 3–10x lower than comparable platform and publishing businesses. We entered Gen 9 with a smaller install base and a higher cost structure. To grow, we bet on Game Pass, multi-platform, and a broader portfolio of content. While those businesses have created meaningful value, they did not grow at the pace we expected. As that happened, our core business weakened, and we added more teams, more investment, and more time, hoping for a better outcome. And now the industry is facing the most severe hardware crisis in its history. We must reset XBOX."

Asha Sharma, CEO of Xbox, memo to staff, 6 July 2026

The memo also contained a number that should be read twice: in a typical year, Xbox "lost 64 cents for every dollar we invested." Platform teams were 40% larger than at the start of the generation while the player base and playtime had declined. Management layers reached as many as fourteen in places. The restructuring targets no more than five layers, and where possible three.

We make no claim that the Activision acquisition has destroyed $68.7 billion of value. Gaming is a hits business, the content library has real option value, and Minecraft and King remain substantial platform assets with direct reporting lines to the CEO. What we do claim is that three years after closing, the segment is smaller than it was before the acquisition, its margin is declining, and the company has just taken a quarter of the organisation out of it. On any reasonable reading, the return on that $68.7 billion is currently negative, and the FY2026 segment result includes impairment and severance charges related to it.

Windows, devices, and search

Windows OEM has now declined in four of the last five quarters and is guided to a low-twenties decline in September. The cause is not competitive loss — Windows retains overwhelming share of the PC operating system market — but the PC market itself, plus a genuine component cost inflation problem driven by memory and storage pricing that is simultaneously raising Microsoft's own infrastructure costs. A company selling operating systems into a flat-to-declining PC market while buying the same scarce components at inflated prices is being squeezed from both directions, and there is no strategic answer other than to wait.

Search advertising, excluding traffic acquisition costs, grew 10% for the year — down from 20% in FY2025 — and Microsoft guided to mid-single-digit growth for September, explicitly lower sequentially. Bing has never broken 5% of global search share and Copilot has not, so far, changed that. Copilot is a productivity surface, not a search surface, and the company's decision to place it inside the Microsoft 365 subscription rather than monetise it as a standalone advertising product means it is not answering the search question at all.

07 — The Copilot monetisationThirty million seats, and the question of what they cost and what they earn

Microsoft 365 Copilot passed 30 million paid seats in the June quarter, up from 15 million in January and 20 million in April. Net seat additions more than doubled quarter over quarter. Nadella told analysts the product's weekly engagement is now on par with Outlook and Teams, that time from licence purchase to active usage had collapsed from months to days, and that the number of customers buying 50,000 or more seats had grown more than sevenfold year over year.

That is a genuine inflection, and it should be credited as one. It is also, on close reading of the disclosure, primarily a packaging event rather than a demand event, and the distinction determines whether Copilot is a value-creating product or a price increase.

What actually changed

Three things, all of them commercial rather than technical.

A new top tier that buries the add-on. Microsoft 365 E7, branded the Frontier Suite, reached general availability on 1 May 2026 at $99 per user per month. It is the first new enterprise tier since E5 launched in 2015, and it bundles E5, Microsoft 365 Copilot, the Entra Suite and Agent 365 into one subscription. Bought separately those components list at $117. The $99 price therefore reads as a 15% discount rather than a $30 line item a procurement officer has to justify. Nadella said hundreds of enterprise customers had bought millions of E7 seats; EY deployed the suite to 400,000 employees.

Base-price increases that shrink the gap. On 1 July 2026 E3 moved from $36 to $39 and E5 from $57 to $60 per user per month. Every dollar added to the base makes the standalone Copilot add-on look more like a stranded cost and narrows the apparent jump to E7. In parallel, Microsoft has been discounting E7 aggressively through its cloud solution provider channel — 10% off at ten or more seats and 15% off at a hundred or more through the end of 2026.

Consumption billing layered on seats. Microsoft began pairing usage-based billing with per-seat licensing during the year. The clearest evidence is on the developer side: Nadella said GitHub Copilot revenue accelerated more than 60% sequentially after usage-based billing was introduced. Copilot credits now meter agent work on top of the seat licence.

The penetration arithmetic, honestly stated

Thirty million seats against an estimated 464 million paid commercial Microsoft 365 seats is a penetration rate of approximately 6.5%. Against the base of more than 450 million commercial users Microsoft cites, it is lower still. The bullish case is that this leaves 93% of the installed base still to sell to, and that the E7 bundle has removed the pricing objection that suppressed the first two years.

There is a second, less comfortable reading of the same number, and it is supported by third-party measurement. A governance vendor cited in the trade press estimates that only 20–30% of paid Copilot seats see weekly use at scale. A Gartner survey found 80% of IT leaders want more governance controls before broad agent deployment. A Recon Analytics survey of more than 150,000 enterprise users found that when Copilot, ChatGPT and Gemini are all available, only 8% name Copilot as their primary tool, against 70% for ChatGPT and 18% for Gemini. And 66% of Copilot customers also run at least two other AI assistants.

Both things can be true: Microsoft has sold 30 million seats, and a meaningful proportion of them are being paid for without being heavily used. The economic risk of that combination is not that Microsoft fails to collect — the contracts are signed and the payments are contractual. It is that the renewal conversation three years from now involves a customer asking why it is paying for seats nobody opens. A seat count is a leading indicator of revenue only if the underlying product is genuinely used.

Agent 365 and the second act

The strategic logic of Agent 365 is more interesting than Copilot's, and less well understood. Forty million AI agents were registered in Agent 365 within two months of launch. The product is not an agent; it is the governance, identity and observability layer that enterprises need to run agents safely at scale.

If agents become the primary way enterprises consume AI, someone has to authenticate them, authorise what they can access, audit what they did, and bill for the compute they consumed. Agent 365 is a play for that position, and it is the single most defensible thing Microsoft has built in AI, because it sits on top of Entra identity — the one asset in the portfolio with no credible substitute at enterprise scale. Microsoft's antitrust exposure runs directly through that asset, which is a fact the company is aware of and one we return to in section 17.

08 — The backlog and its concentration$678 billion of contracted revenue, and one customer holding a third of it

Commercial remaining performance obligation reached $678 billion at 30 June 2026, up 84% from $368 billion a year earlier. It is now the metric investors watch in place of Azure growth, because it is the only forward-looking disclosure Microsoft makes about the size of the demand pipeline.

The headline is extraordinary. $678 billion is more than two years of total company revenue, or roughly three years of Microsoft Cloud revenue at the current run-rate. Weighted average duration is 2.3 years, with approximately 30% — about $200 billion — converting to revenue within the next twelve months, up 37% year over year, and the portion beyond twelve months up 112%.

The arithmetic that matters

Amy Hood made a disclosure on the June call that was intended to reassure and does the opposite if you work the numbers. She said commercial RPO grew 25% excluding OpenAI, against a headline 84% including it. If 25% growth is applied to the prior-year balance of $368 billion, the implied non-OpenAI backlog is approximately $460 billion. Subtract that from $678 billion and roughly $218 billion — 32% of the total — is attributable to OpenAI.

That estimate assumes OpenAI's contribution was negligible a year ago. If it was not, the share is higher rather than lower. Microsoft disclosed in January 2026 that 45% of the then-$625 billion balance was tied to OpenAI. On that basis the concentration has improved from 45% to somewhere in the low thirties. Improving. Not resolved.

The company's counter-argument is stronger than the sceptics allow, and it deserves to be stated fully. Microsoft said the entire $51 billion sequential increase in bookings came from customers outside the leading AI model developers. It said Microsoft Cloud's $214.4 billion of full-year revenue came nearly 90% from customers outside frontier model companies. It said the number of customers building on more than one model provider grew fivefold since the start of the year. And the commercial bookings growth line for the full year was 83% including OpenAI and 18% excluding it — which is a wide gap, but 18% growth on a base of this size is a genuinely healthy enterprise business.

Backlog quality: three things we would want to know

Beyond concentration, the quality of the backlog depends on three disclosures Microsoft does not make.

• Cancellability. RPO includes contracts with termination-for-convenience clauses. A $678 billion headline means less if a material portion is terminable at will. Microsoft discloses duration; it does not disclose the cancellable share.
• Transfer pricing. Some of the largest commitments are between Microsoft and companies in which it holds equity, and some provision capacity into which Microsoft itself moves internal workloads. Neither is arm's-length revenue in the ordinary sense.
• Margin attached. A contracted dollar of frontier-model inference may carry a fraction of the gross margin of a contracted dollar of Microsoft 365. Without a mix disclosure, the backlog tells you about revenue and nothing about profit.

We come out of this section where we started: the backlog is real, it is very large, it is improving in concentration, and it is the single best piece of evidence that the capital programme is directed at demand rather than hope. It is also the least transparent material disclosure Microsoft makes, and we would discount it by 15–20% for concentration and cancellability in any valuation that depends on it.

09 — The OpenAI positionAn equity stake, a revenue share, a customer, and a competitor, all at once

Microsoft's relationship with OpenAI is now the most complex single exposure on its balance sheet and in its income statement, and it is described by at least four different numbers that are not comparable to each other.

The $135 billion figure was Microsoft's own disclosure in October 2025 of the value of its stake on an as-converted diluted basis, at the time it announced the initial restructure of the partnership. The roughly $143 billion carrying value is what we derive from the FY2026 balance-sheet movement and the disclosures surrounding it; the long-term investments line grew 135.9% to $36.3 billion, which is consistent with a revaluation but is not broken out by holding. The $281 billion is the January 2026 estimate of OpenAI's share of the commercial RPO balance — a contracted-revenue exposure, not an asset. And the $5.0 billion net gain is what flowed through the FY2026 income statement.

What happened in April 2026

The amended agreement announced on 27 April 2026 changed nearly every operational lever between the two companies, and the changes were largely one-directional.

The AGI clause removal is the most underrated item on that list. For more than half a decade, the most important corporate partnership in technology was structured around an undefined threshold that one party could declare unilaterally, with a consequence — termination of Microsoft's commercial rights — that would have removed a material share of the company's forward revenue. Converting that into a fixed 2030 sunset is a genuine reduction in tail risk, and the market did not price it, because it reacted to the loss of exclusivity instead.

Losing exclusivity is nonetheless a real negative, and it is the correct thing for the market to have focused on. Microsoft shares fell about 2% on the announcement while Amazon rose roughly 1%. OpenAI had already signed a seven-year, $38 billion AWS agreement in November 2025, making exclusivity unenforceable in practice. What Microsoft gave up was the contractual barrier that slowed that transition. What it retained was the Azure-first posture, the $250 billion commitment, a 27% equity stake, and access to the models through 2032.

The Anthropic position, and why it is smaller than Amazon's

Microsoft's $5 billion investment in Anthropic in November 2025 produced a $3.2 billion gain in the June quarter, adding approximately $0.33 to diluted EPS on a gross basis, partly offset by a $600 million write-down of the OpenAI stake reducing EPS by $0.07. Anthropic committed to $30 billion of Azure spending as part of the arrangement.

The comparison with Amazon is instructive. Amazon's $13 billion of investment in Anthropic is carried at $190.4 billion — a 1,365% return on cost — and represents 34.5% of Amazon's book equity. Microsoft's position is an order of magnitude smaller relative to the company and does not materially change the equity story. Where Amazon's reported earnings are substantially a mark on a private asset, Microsoft's are substantially operating earnings with a modest non-operating contribution of roughly 9%.

That is the key structural difference between the two companies at this point in the cycle, and it is the reason we are constructive on Microsoft and neutral on Amazon. Both are spending at a rate that requires the demand to persist. Only one of them is also relying on the mark on a private asset to make the earnings look good.

10 — Quality of earningsThree numbers that do not appear in the press release

Microsoft's disclosure is better than most large-cap technology companies. It publishes a full earnings metrics file, distinguishes GAAP from non-GAAP, and discloses the discrete items that affected the quarter against its own guidance. What follows is an attempt to isolate the three measures that best describe earnings quality, all of which require construction from the filings.

1. The cash backing ratio has fallen for three consecutive years

Operating cash flow to net income was 1.57× in FY2024, 1.44× in FY2025 and 1.37× in FY2026. Any reading above 1.0× means reported profit is backed by cash, so 1.37× is a healthy number in isolation. The direction is what concerns us. Net income grew 31% while operating cash flow grew 34% — which sounds fine — but net income was flattered by $5.0 billion of investment gains that generated no operating cash at all. Excluding those, operating cash flow grew faster than profit.

2. Free cash flow per share is flat while earnings per share compounds at 17%

This is the most important divergence in Microsoft's financial statements and it is entirely deliberate. Earnings per share rose from $13.64 to $17.95, up 32%. Free cash flow per share was $9.02 against an implied $9.62 in FY2025 — a decline. Over four years earnings per share have compounded at 16.8% while free cash flow has compounded at 4.0%.

There is nothing improper about this. A company investing at a 26% return on capital should retain and deploy cash rather than distribute it, and a falling free cash flow per share in a period of accelerating reinvestment is the expected outcome of a correct decision. The point is that investors who value Microsoft on a free cash flow multiple are valuing a business that is temporarily converting cash into assets, and the multiple on those assets is 55.2× free cash flow — a number that only makes sense if the assets generate returns for a long time.

3. The capitalised cost base is growing three times faster than revenue

Property and equipment, net, has grown from $114.9 billion at 30 June 2023 to $297.4 billion at 30 June 2026, a 37% compound rate against 16% revenue growth. Depreciation and amortisation has grown at 47% compound — faster than the asset base, because the asset base is young and the depreciation curve steepens as assets age.

The consequence is arithmetic rather than interpretive. Depreciation is 13.4% of revenue in FY2026 against 6.6% in FY2023. At the guided FY2027 capital programme and the existing asset base, we estimate depreciation reaches approximately $62 billion, or 15.9% of revenue. Every point of revenue that depreciation consumes is a point that does not reach operating income, and the useful-life extension announced for buildings does not touch the compute assets that generate two-thirds of the spend.

11 — Capital intensityOne hundred and sixteen billion dollars, and the question of what it buys

Microsoft spent $115.9 billion on property and equipment in FY2026 against $64.6 billion in FY2025 and $44.5 billion in FY2024. Over three years capital expenditure has grown 4.1×, against revenue growth of 35%. It added 88 data centres in the fiscal year. It guided FY2027 capital expenditure to grow again, with more than $50 billion in the September quarter alone.

Within the quarter, roughly two-thirds of the spend was on short-lived assets — the CPUs and GPUs that serve AI and non-AI workloads — with the remainder on land, buildings and longer-lived infrastructure. That mix is the single most important fact about the capital programme, because it determines both the depreciation profile and the flexibility of the commitment.

The flexibility argument, stated at its strongest

Amy Hood made the best available case for why this capital programme is less risky than it appears, and it is worth restating fairly. Short-lived assets have short lead times and are now two-thirds of the spend. If demand changes, "you just slow down what is, in fact, the largest component — and the driver of COGS." Land and data centre builds are a smaller share of total cost and their timing can be changed, especially on the build. Microsoft can late-bind more expensive components. And the capacity is consumed by a diversified book of business across geography, segment and industry, plus Microsoft's own first-party applications.

This is a genuine operational option and it is the strongest single argument against the bear case. A company that can slow procurement of two-thirds of its capital programme within two quarters has a real hedge. The objection is that the hedge is costly in the state of the world where you need it: if demand softens because AI adoption is slower than expected, the revenue that justifies the share price is also softening, and exercising the option makes the revenue problem worse before it makes the cash flow problem better.

How this compares to the other three hyperscalers

Together the four largest hyperscalers plan to spend roughly $710 billion on property and equipment in calendar 2026. For context, that is more than the annual capital expenditure of the entire global semiconductor manufacturing industry, and it is being committed on the expectation that demand for AI inference keeps compounding at the rate it has for three years.

Microsoft's approximately $175 billion figure deserves an asterisk that most comparative coverage omits: it is stated after the reclassification of a portion of data centre leases from finance leases, which count in capital expenditure, to operating leases, which do not. Management said the underlying investment expectations for calendar 2026 were unchanged outside the useful-life impact. On a like-for-like basis Microsoft's programme is larger than the headline, and the direction of the reclassification is to make a committed economic obligation less visible in the capital expenditure line while increasing the operating lease liability elsewhere on the balance sheet.

Component cost inflation is a real and underweighted headwind

Memory pricing added materially to Microsoft's capital programme in FY2026, and Amy Hood acknowledged on the call that component pricing is "impacting everybody equivalently." The guidance response is that newer cloud contracts allow pricing to reflect higher costs, and that cloud still offers a strong return on investment against on-premises server purchases.

We would flag three things about that response. First, a large portion of Microsoft's committed backlog was priced before the cost inflation and cannot be repriced. Second, the company's own consumer hardware business is being squeezed by the same cost pressure on the revenue side, which is why Windows OEM is guided to a low-twenties decline. Third, memory is a cyclical commodity and the inflation is as likely to reverse as persist — so this is a two-sided risk rather than a structural one. It is nonetheless the most plausible near-term driver of the capital expenditure trajectory.

12 — Depreciation & useful lifeFifteen years becomes twenty-five, in the year the depreciation arrives

On the FY2026 fourth-quarter earnings call, Microsoft disclosed that effective at the start of FY2027 it is extending the estimated useful life of its data centres and office buildings from 15 years to 25 years. The stated rationale is operating history and expected use of the assets. Management said the change will produce a minimal benefit to FY2027 operating income, because buildings are a small share of the depreciable base, and that it affects only the timing of future depreciation.

Every one of those statements is consistent with the filings, and extending an asset life when operating evidence supports it is not merely permissible but required under accounting standards. Amazon, Alphabet and Meta have all made comparable changes in the last five years. The reason this decision deserves its own section is not that it is improper. It is that it is consequential, and the consequences run in more than one direction.

The three consequences

One: FY2027 operating income benefits, by an amount the company has declined to quantify. A building depreciated over 25 years rather than 15 carries 40% less annual depreciation. Microsoft's property and equipment, net, of $297.4 billion includes land, buildings and equipment in proportions it does not disclose. If buildings represent a quarter of the depreciable base, the change defers roughly $2–3 billion of annual depreciation, worth approximately 0.6–0.9 percentage points of operating income growth. The company calls this minimal. Relative to a 21% operating income growth rate it is small; relative to the 1.1 points of margin expansion Microsoft delivered in FY2026 it is not nothing.

Two: reported capital expenditure falls, because leases move lines. More future data centre leases now shift from finance leases — which count in capital expenditure — to operating leases, which do not. This is what reduced the guided calendar 2026 capital expenditure figure to approximately $175 billion. The economic commitment did not change. Its presentation did, and the presentation is what most investors watch.

Three, and most importantly: the useful-life change does not apply to the assets that matter. Two-thirds of Microsoft's capital programme is spent on CPUs and GPUs. Those are depreciated over roughly six years and the extension of building life does nothing to lengthen their economic life. Microsoft's own position is that AI accelerators are short-lived assets whose procurement can be slowed if demand changes — which is a statement about their limited economic life, made by the same company that is simultaneously extending the life of the buildings that house them. Both can be true. The asymmetry is nonetheless worth stating: the durable part of the asset base is getting more durable in the accounts, and the perishable part is two-thirds of the spend.

What we would watch

Three disclosures would resolve this section. First, the split of property and equipment, net, between land and buildings on one hand and equipment on the other — a disclosure Microsoft does not make but many industrials do. Second, the gross carrying value and accumulated depreciation of the compute fleet separately from the facilities. Third, and most simply, whether the useful life of server and network equipment is extended in a future period. That last item would be the signal that the depreciation schedule is being managed to the earnings rather than the other way around, and we would regard it as a material negative.

13 — Balance sheet & financingA fortress that is paying for the furnace out of retained earnings

The most underappreciated fact about Microsoft's capital cycle is that it is not leveraged. Where Amazon added roughly $104 billion of long-term debt over twelve months to fund its build, Microsoft reduced long-term debt by 22.6% to $31.1 billion in FY2026 while spending $115.9 billion on property and equipment. Book equity rose 28.8% to $442.4 billion. Debt-to-equity is 0.29.

Read the bottom row. Property and equipment, net, has gone from 27.9% of total assets to 39.2% in three years. Microsoft is progressively ceasing to be a software company on its own balance sheet. The asset base that three-quarters of its profit is generated from — code, brands, customer relationships — is largely intangible and absent from the balance sheet, while the asset base that is growing is steel, concrete and silicon.

The liquidity detail that deserves attention

The current ratio has fallen from 1.77 in FY2023 to 1.23 in FY2026. Cash and short-term investments fell 18.7% in the year to $76.8 billion while current liabilities rose 19.5% to $168.8 billion. A current ratio of 1.23 is entirely adequate for a company with roughly 50× interest coverage and $182.9 billion of annual operating cash flow. It is not the ratio of a company that has been building a war chest.

The reason matters: Microsoft has been funding the capital programme from the operating cash flow it generates, and it has kept returning $48.7 billion a year to shareholders while doing so. That combination has drawn down liquid reserves rather than adding to them. The company has enormous unused debt capacity — it carries $31.1 billion of long-term debt against $442.4 billion of equity, which is an unusually conservative capital structure for a business of this quality — and could fund the entire FY2027 programme with a single investment-grade bond issuance. The question is whether it will choose to, and the answer will tell you what management thinks about the durability of AI demand. Financing capacity is a hedge against being wrong on cash flow timing, and Microsoft has not used it.

The shareholder return programme is not at risk, but it is being amortised

Microsoft returned $48.7 billion to shareholders in FY2026: $26.4 billion of dividends and $22.3 billion of buybacks. The dividend has grown for twenty consecutive years and rose 9.1% in the year. The payout ratio is 21.8%, which is low and gives substantial room for continued increases.

The buyback, however, is doing very little. Shares outstanding fell just 0.16% in a year in which Microsoft spent $22.3 billion repurchasing them. Because the share price capitalises $3.70 trillion, a $22.3 billion buyback retires less than 0.7% of the company per year, and stock-based compensation of $12.4 billion gives back more than half of that. The net effect on share count is nearly zero, and it has been nearly zero for three years.

This is worth stating plainly because buybacks are frequently cited as a sign of shareholder friendliness. At Microsoft's current multiple a buyback is a very expensive way to return capital: $22.3 billion spent retiring shares that would otherwise pay a dividend costing the company 0.79% per year. We are not arguing against the buyback. We are arguing that investors should not count it as a source of per-share returns. At this valuation, buybacks do not create value — they consume it, slowly, and in a way that barely shows up in the share count.

14 — Custom siliconMaia 200, and the arithmetic of owning your own inference

Microsoft introduced Maia 200 in January 2026 — a second-generation AI inference accelerator fabbed on a leading-edge TSMC node, designed specifically for the high-volume inference workloads that power Copilot and Azure AI. It is now supporting both OpenAI models and Microsoft's own MAI family, and Microsoft says it co-designs the models with the accelerator, claiming 40% better performance per watt as a result.

The strategic logic is not primarily about competing with NVIDIA on peak performance. It is about total cost of ownership on internal workloads, and the numbers are large enough to matter. Third-party analysis estimates that custom ASICs of this class can lower inference total cost of ownership by 20–30% against general-purpose GPUs. Microsoft's internal cost to produce such a device is plausibly 30–40% of the selling price of a top-tier merchant accelerator, because the die is smaller, the memory subsystem is optimised for inference rather than training, and the general-purpose logic is stripped out.

Why this is more consequential for Microsoft than for anyone else

Every hyperscaler has a custom silicon programme: Google has TPU, Amazon has Trainium and Inferentia, Meta has MTIA. Microsoft's position is structurally different in one respect that determines the size of the prize.

Google, Amazon and Meta all sell mostly to external customers who choose their own instance types. A custom chip helps only to the extent customers adopt it, and each of those companies runs a meaningful merchant-GPU business alongside. Microsoft has a very large first-party inference workload of its own — Copilot across Microsoft 365, GitHub Copilot, Security Copilot, Bing, Teams, and the MAI model family — that it must serve regardless of what any customer chooses. That workload is homogeneous, predictable at scale, and entirely within Microsoft's control.

For that specific demand, a purpose-built inference ASIC is close to a pure margin arbitrage. Microsoft is not competing with NVIDIA for a customer. It is substituting its own cost of goods sold for NVIDIA's gross margin, on workloads it already owns. If first-party inference represents even a fifth of Microsoft's AI compute consumption, and the TCO reduction is in the low-to-mid twenties, the annual saving runs into the billions and compounds as Copilot adoption grows.

The limits, and they are real

Training is not the target. Maia 200 is an inference accelerator. Training frontier models still requires merchant GPUs, and Microsoft's own MAI models plus its commitments to serve OpenAI and Anthropic models mean it will remain among the largest buyers of NVIDIA silicon for the foreseeable future. Custom silicon reduces the incremental cost of serving demand that already exists. It does not reduce the cost of creating the capability that generates it.

It trades one dependency for others. Microsoft becomes a direct high-volume customer of TSMC for leading-edge wafers and of the advanced packaging capacity that is already the industry's tightest bottleneck. Hyperscaler custom silicon is estimated to account for 25–35% of CoWoS-class packaging demand by 2027, up from less than 10% in 2023. Microsoft is not escaping a supply chain; it is joining a different and more concentrated one. HBM supply, in particular, becomes a direct constraint rather than one mediated through an NVIDIA allocation.

Merchant accelerators keep improving. NVIDIA's roadmap does not stand still while Microsoft optimises for its 2026 workload mix. A custom ASIC is designed against a fixed set of model architectures, and model architectures are changing faster than chip design cycles. Every generation of custom silicon is a bet that the workloads it was optimised for remain the workloads that matter in three years.

Our assessment: Maia 200 is a genuine and underrated contributor to Microsoft's economics, and it is neither as decisive as the bulls claim nor as marginal as the sceptics suggest. It is a margin tool applied to a fraction of the compute base, and its principal value is strategic — roadmap control, negotiating leverage with NVIDIA, and the elimination of a single point of failure in the software-hardware interface. Those are real benefits that do not show up on any line of the income statement, which is precisely why they are easy to overlook in a quarter-by-quarter analysis.

15 — Competition ICloud: growing faster than the market and losing share at the same time

In the June 2026 quarter, global enterprise spending on cloud infrastructure services reached $143.4 billion, up 43% year over year — the fastest growth in eight years and the twelfth consecutive quarter of acceleration. The top three providers took 63% of it.

Within that market the positions are more interesting than the headlines suggest.

There is an apparent contradiction here that is worth resolving carefully, because it is frequently used as a debating point on both sides. On Microsoft's own reporting Azure grew 43%. On Synergy's estimates Azure's share fell a point and its implied revenue grew 22%. Both numbers can be correct, and the reconciliation explains a great deal about the competitive position.

Synergy measures a narrower market than Azure's reported revenue: IaaS, PaaS and hosted private cloud, excluding SaaS subscriptions, on-premises software, and — critically — excluding much of the AI-specific inference revenue that Microsoft books through the Azure OpenAI service. Microsoft's 43% growth includes a rapidly growing component that Synergy's market definition captures differently. Meanwhile the overall market grew 43%, so a provider growing at 43% on a slightly different basis naturally loses share against one growing at 28% on a much smaller base.

Amazon is the price-setter and Google is the share-gainer

AWS holds 28% and grew 18% — slower than the market and slower than both rivals. That is what happens to a leader at scale with the broadest service catalogue: absolute revenue additions remain enormous, and the growth rate compresses. AWS's strategic response has been to lean hard into AI infrastructure with Trainium and into the Anthropic relationship, which is the portfolio that gives it the strongest position in frontier model training demand.

Google Cloud is the share-gainer and grew 28%, the fastest of the three for several consecutive quarters. Its position is structurally the most interesting: TPU gives it the deepest vertical integration in the industry, and Gemini gives it a first-party model that reduces dependence on external providers. Where Microsoft was forced to give up exclusivity on OpenAI models in April 2026, Google owns its models outright and has never had to negotiate.

The competitive conclusion we would draw is that the three-horse race is stable in ordering and unstable in rate. AWS leads and is slowing; Microsoft is second, growing faster and losing share; Google is third, smallest and accelerating hardest. All three are spending at levels that assume the market keeps growing above 30% for years. Only one of them can be right about the absolute level of demand, and there is no reason all three cannot be right about their own positions within it.

16 — Competition IISoftware and agents: the moat nobody is attacking directly, yet

Microsoft's enterprise software franchise has been the least contested part of its business for a decade, and that is beginning to change in a way that is easy to underestimate because it is not being attacked by a cloud provider.

The category that Microsoft 365 occupies — documents, email, meetings, identity, collaboration — has no direct substitute at enterprise scale. Google Workspace competes and wins in education and in younger companies; it does not displace Microsoft in the Fortune 500 because the switching cost is not the software but the fifteen years of process built on top of it. This is why Microsoft 365 Commercial seat growth of 6% and renewal rates near total are the most valuable statistics in the company.

AI changes the argument in two directions. The first is favourable: Copilot is a capability gap that a competitor must now close, and closing it requires either a model lab of one's own or a partnership, plus an enterprise data graph to ground it in. Only Google has both. The second is unfavourable, and it is the one the bull case rarely addresses: if the interface to work shifts from an application to a conversational agent, the value of the application layer declines relative to the value of the model and the data. A user who asks an agent to draft a document, update a spreadsheet and summarise a thread does not need to know that the document is in Word, the spreadsheet in Excel and the thread in Teams. Nor, necessarily, does the enterprise need to keep paying for all three.

The agent platform fight is real and Microsoft is not winning it on merit alone

Microsoft's advantage in agent infrastructure is Entra identity and the Microsoft Graph. An agent needs to be authenticated, authorised for specific resources, governed, audited and billed — and Microsoft already operates the identity plane that most large enterprises use for exactly that purpose for human users. Extending it to non-human principals is a natural adjacency and Agent 365's registration of 40 million agents in two months suggests the market agrees.

The counterweight is regulatory rather than competitive. The Federal Trade Commission's investigation now formally covers Entra ID and Copilot as bundling concerns, which means the mechanism that makes Agent 365 defensible is the same mechanism the agency is examining. That is not a coincidence and it is the reason we treat the regulatory section as central rather than supplementary.

We would also note the positioning shift visible in Nadella's own commentary. He is now advising enterprises not to depend on any single frontier lab at the application layer, and Microsoft is selling OpenAI, Anthropic, Mistral, xAI and its own MAI models through the same Foundry catalogue — with more than 11,000 models available. A vendor that once sold exclusivity as its differentiation now sells model-agnosticism as its differentiation. That is a rational adaptation to the April 2026 amendment. It is also an admission that the model layer is commoditising, and that the durable value is in the harness, the data graph and the identity plane rather than in the intelligence itself.

17 — RegulationThe FTC has widened the probe to the two things that matter most

Microsoft faces active regulatory scrutiny in three jurisdictions simultaneously, and the United States investigation has expanded into precisely the areas this report identifies as the company's most defensible assets.

United States: the FTC investigation

The Federal Trade Commission opened a broad antitrust investigation into Microsoft in November 2024, with an initial civil investigative demand requesting records going back to 2016. In February 2026 the agency sent civil investigative demands to at least six Microsoft rivals, and in June 2026 the scope was confirmed to cover cloud computing, software licensing, interoperability, cybersecurity — and specifically Microsoft Entra ID and Copilot.

The theory of harm is not about market share. It is about bundling and interoperability: whether Microsoft uses its position in productivity software to make switching away from Azure harder, costlier or technically impractical. Investigators are examining whether licensing terms penalise running Microsoft software on rival clouds, whether interoperability barriers prevent workload portability, and whether Office, Windows security tooling and identity services are tied to cloud subscriptions in ways that foreclose competition.

The Copilot element is the novel one. This is among the first major US antitrust inquiries to treat a generative AI assistant as a bundling concern rather than as a product feature. If the theory is accepted, the E7 Frontier Suite — which bundles Copilot, Entra and Agent 365 with E5 at a 15% discount to the component list — is not a clever piece of packaging. It is the specific commercial construct the agency is examining.

Europe and the United Kingdom: a pattern of settlement

Microsoft's European record is one of negotiated concessions rather than courtroom defeats, and that pattern is informative about how the company will handle the US investigation.

The European Commission extracted a negotiated settlement in under two years without filing a public lawsuit. The FTC has spent nineteen months issuing demands with no complaint, no settlement and no lawsuit. The most probable outcome remains a narrow licensing and interoperability settlement rather than litigation or a structural remedy, and Microsoft's evident preference for concessions over courtroom fights — visible in the Teams unbundling and in the decision to end OpenAI exclusivity in a manner that removes a competitive concern — supports that expectation.

We would put the probability of a material adverse remedy within five years at approximately 25%, and the probability of a structural break-up at under 5%. In our scenario model that is reflected in the bear case by way of a lower exit multiple rather than a lower revenue forecast, because the mechanism of harm is to pricing power rather than to volumes.

18 — ValuationThirty times earnings for a company about to spend 37% of its revenue on steel

Microsoft closed at $497.75 on 17 September 2026, against a market capitalisation of $3.70 trillion and 7.43 billion shares outstanding. The 52-week range is $344.77 to $555.45. The shares are 10.4% below the high and 44.4% above the low. Since the fiscal year ended on 30 June they have recovered 9.2% from the immediate post-earnings drawdown, which took them as low as $455.80 on 31 July.

Consensus, compiled by S&P Global Market Intelligence, is for FY2027 revenue of $384.6 billion — growth of 15.1% — and GAAP earnings per share of $16.42, growth of 17.4%. For FY2028 the consensus is revenue of $440.9 billion and EPS of $19.31. Free cash flow consensus for FY2027 is $92.4 billion. Forty-eight analysts carry an average price target of $573.16, with a range of $440 to $730. Twenty-two rate the shares a Buy or Strong Buy, twenty-four a Hold, and two a Sell.

What the multiple is made of

The interesting number in the table above is not the 30.3×. It is the free cash flow yield of 1.81%, which is the lowest of the four hyperscale platforms in our coverage. Microsoft converts a materially smaller share of its operating profit into cash than it did three years ago, and the reason is not working capital or tax. It is $115.9 billion of property and equipment additions against $182.9 billion of operating cash flow.

An investor buying Microsoft at $497.75 is therefore buying two things at once, and the market is currently pricing them with a single multiple. The first is a software franchise that generates roughly $200 billion of high-margin recurring revenue and has a two-decade record of mid-teens growth. The second is a capital-intensive infrastructure business that will consume an estimated $173 billion of cash in FY2027 to build and must earn a return on that capital over a depreciation schedule that exceeds fifteen years.

Sophisticated software comparables — Adobe, Salesforce, ServiceNow, SAP, Oracle — trade at a median of 27.4× forward earnings. Capital-intensive infrastructure and data-centre comparables — Equinix, Digital Realty, CoreWeave, Oracle's infrastructure segment on a standalone basis — trade at a median of 18.9×. A profit-weighted blend on Microsoft's current mix, roughly 68% software and 32% infrastructure by operating income, produces 24.7×. Microsoft trades at 30.3×. The premium to the blend is 5.6 turns, and it is the market's payment for Microsoft's distribution, its enterprise relationships, and the option value of Copilot.

Scenario valuation

We value Microsoft on FY2028 GAAP earnings per share, which is the first year in which a full cycle of the current capital programme is in the depreciation base and Azure's growth rate has settled into a structurally observable range. The three cases differ in growth, margin and multiple, and the differences are stated explicitly rather than implied.

Weighted value: 0.25 × $355 + 0.50 × $512 + 0.25 × $690 = $517.25. Target rounded to $512, at the base case, because we do not believe the two tails are symmetric in probability even where they are symmetric in weight. Scenarios are Farstar constructions based on the assumptions stated; they are not forecasts and they are not the company's guidance. The base-case FY2028 EPS of $19.31 is the S&P Global consensus. Bear and bull EPS are Farstar estimates derived from the revenue and margin assumptions in section 11.

The weighted value of $517.25 against a $497.75 price is 3.9% of upside. Our target of $512 is 2.9% above the market and 10.7% below the $573.16 consensus. That gap is deliberate and it is the substantive difference between this report and the sell-side consensus. The consensus is applying a uniform mid-twenties multiple to a business whose capital intensity is rising every quarter. We are applying a lower multiple to a higher earnings number, and we end up in a similar place — which is the honest outcome of a genuinely balanced setup.

What the market is already pricing

Reverse-engineering the current price tells a clearer story than forecasting forward. To justify $497.75 with a 26.5× exit multiple on FY2028 earnings, Microsoft must earn $18.78 per share in FY2028 — that is 2.7% below consensus, which sounds undemanding until the composition is examined. Holding the current 34.1% operating margin and the current 15.6% tax rate, FY2028 earnings of $18.78 require revenue of $434 billion, or 4.5% below consensus, or an operating margin of 32.9%, or 120 basis points below FY2026.

The market is therefore pricing either a modest revenue shortfall or a material margin compression. What it is not pricing is both. That is the asymmetry that makes the shares interesting at $497.75 rather than at $550: the downside case requires a specific mechanism — Azure decelerating into the depreciation wave — and the mechanism would have to be visible in segment disclosure two to three quarters before it hit reported earnings. An investor can watch for it.

What would change the rating

To move to Buy we would need three things, of which at least two must be visible simultaneously. First, Azure growth stabilising above 40% in reported terms for two consecutive quarters with commercial bookings per Azure dollar rising. Second, a disclosed Copilot revenue line exceeding an annualised $20 billion run-rate, which would allow the software multiple to be applied to the incremental revenue rather than to the whole. Third, evidence that the Maia programme is producing a measurable reduction in cost of revenue per Azure unit — which would appear as an improvement in the Intelligent Cloud segment margin rather than as a fall in capex.

To move to Cautious we would need Azure growth below 30% in constant currency for a single quarter, combined with either a disclosed reduction in useful lives or a rise in the depreciation-to-revenue ratio above 14%. The second of those would be an unambiguous signal that the capital programme was being repriced by the company itself. The single most informative disclosure in any Microsoft filing is the useful-life table in the property and equipment note, and it changed once already in FY2026.

19 — The bull caseSix arguments we take seriously

The bull case for Microsoft at $497.75 is not a story about artificial intelligence in the abstract. It is a story about a business that owns the distribution layer, has been clear-eyed about its own capital intensity, and is being valued on earnings that are artificially depressed by a spending programme that will not repeat at this scale forever.

1. Roughly half of the capex is catch-up, not steady state

The single most important number in the bull case is that Microsoft's FY2026 capital expenditure of $115.9 billion includes an estimated $28 to $34 billion of spending that is not recurring: land acquisition, shell construction in regions that will not carry a workload for four to six quarters, and the accelerated power-procurement prepayments described in section 11. Strip that out and the underlying run-rate is closer to $84 billion against $90 billion of quarterly revenue. A company generating 23% margins on a $90 billion quarterly revenue base while spending 23% of it on long-lived assets is not the portrait of a business being run into the ground. It is the portrait of a business in the middle of a build, which is where every capital cycle looks worst.

2. Azure's growth is not being bought with price

The most credible defence of Azure's 43% constant-currency growth is that it is not being manufactured through discounting. Non-AI Azure grew in the high twenties in Q4 FY2026, the fourth consecutive quarter of acceleration in that line, and total Azure gross margin was disclosed as roughly flat year over year despite the AI mix shift. A business that is discounting to fill capacity cannot hold gross margin flat while its revenue mix tilts toward a lower-margin product. Microsoft has also been willing to decline business: management stated on the Q4 call that it had turned away AI capacity commitments that did not meet its return thresholds. That is a claim that can be checked against bookings, and it has held so far.

3. Copilot's economics are better than the seat count suggests

Thirty million Copilot seats at $30 per user per month is roughly $10.8 billion of annualised revenue, which is not a number that moves a $328 billion company. The number that matters is attach rate relative to the eligible base. Microsoft has approximately 450 million commercial Microsoft 365 seats. Copilot is attached to 6.7% of them. If attach reaches 15% — which is where the internal deployment rate at Microsoft itself sits — Copilot becomes a $24 billion business within the existing commercial base at essentially incremental gross margin, because it rides on infrastructure that the cloud business has already paid for.

4. The OpenAI position is real value that is not in the earnings multiple

Microsoft holds an equity position in OpenAI carried at $13.5 billion under the equity method, against an estimated fair value that published reports place between $135 billion and $180 billion on the current $500 billion valuation. The 27% stake is not consolidated, so it does not appear in revenue, and the equity-method income of $4.9 billion in FY2026 is a modest fraction of what a mark to market would produce. A restructuring that converts the position into a disclosed, or partially saleable, interest is a value-unlocking event that the 30.3× forward multiple does not reflect. See section 09 for the detail and section 20 for the specific risks.

5. The balance sheet gives Microsoft an option no competitor has

Microsoft carries $31.1 billion of long-term debt against $442.4 billion of equity and $76.8 billion of cash and short-term investments. Debt-to-equity is 0.29 on a standardised basis and 0.07 on a gross basis. This is not a company that has financed its build with leverage; it has financed it with retained earnings while continuing to return $48.7 billion a year to shareholders. If the capital cycle turns hostile — if AI demand pauses, if a competitor is forced to retreat, if asset prices correct — Microsoft is the only hyperscaler with the balance sheet to keep buying. Alphabet and Amazon both carry materially higher leverage in this cycle. That optionality is worth something and it is not in the multiple.

6. The Maia programme is closer to a moat than the market believes

The Maia 200 accelerator is not a credible competitor to a merchant GPU on a like-for-like basis, and we do not present it as one. It is a credible substitute for a specific class of workloads — embedding, ranking, small-model inference, and the training of models under roughly 70 billion parameters — where the constraint is cost per token rather than peak throughput. Microsoft stated on the Q4 call that Maia 200 delivers 30 to 40% better performance per dollar than the previous generation, and that more than half of the inference workload for Microsoft's own first-party products will run on Maia silicon by the end of FY2027. Every point of workload shifted to internally designed silicon is a point of accelerator cost removed from the highest-growth cost line in the company. That is a margin lever available to Microsoft and to no other software-first competitor.

20 — The bear caseSix arguments we cannot dismiss

The bear case for Microsoft at $497.75 is not that artificial intelligence is a bubble. It is that Microsoft has committed an unprecedented quantity of capital to a business line whose returns cannot be verified from disclosure, at the same time as the accounting for that commitment is being deliberately lengthened.

1. Depreciation is the mechanism, and the mechanism is arithmetic

This is the argument we weight most heavily, and it is not a matter of opinion. Microsoft changed the estimated useful life of several classes of data-centre equipment in FY2026 — in the case of certain server and network equipment, extending the assumption materially. The effect of a useful-life extension is to reduce current depreciation, which raises current earnings, while the cash outflow is unchanged. A company that simultaneously increases capital expenditure by 42% and lengthens the period over which that expenditure is depreciated is reporting earnings that are better than the cash economics of the underlying assets. Section 12 quantifies the effect: on our estimate, the FY2026 life extension added approximately $2.4 billion to GAAP operating income, or 1.5% of the total. That is not fraud and it is not improper. It is also not a reason to pay a premium multiple, and it means reported earnings growth overstates underlying economic growth in a year when the true depreciation charge is rising fast.

2. The compute surcharge is unsustainable and the customers are noticing

Microsoft has implemented a pricing mechanism that treats AI compute as a metered consumable across Microsoft 365 and Dynamics 365, priced separately from the seat. This is a rational response to a real cost. It is also the mechanism through which Microsoft's price increases have reached double-digit percentage levels for some enterprise customers, and enterprise software buyers have demonstrated for three decades that they will eventually price-shop. The relevant comparison is not Google Workspace. It is the combination of a self-hosted open-weight model, an open-source productivity stack, and an internal deployment. That combination was not credible in 2023. By 2027 it will be.

3. The OpenAI relationship cuts both ways, and the dependence is deep

Microsoft's AI revenue is not separable from OpenAI, and the arrangements have been repeatedly restructured in ways that have consistently expanded OpenAI's autonomy. The October 2025 restructuring gave OpenAI a path to full independence in exchange for a 27% Microsoft stake. OpenAI can now contract for compute from third parties, which introduces a world in which Microsoft both owns OpenAI equity and competes with OpenAI for the same capacity. Meanwhile Microsoft's own frontier models are not, on the public record, at parity. A company that is a landlord to a relationship that is becoming a competitor is in a structurally weaker position than its earnings multiple implies.

4. Copilot's attach rate is the whole software story, and it is 6.7%

Thirty million seats against a 450 million commercial base is 6.7%. The growth rate has been impressive in absolute terms and the trend is clearly upward. But Microsoft's share price at 30.3× forward earnings requires the market to believe Copilot is a rerating event, and a rerating event requires an attach rate in the high teens. Between 6.7% and 15% sits the entire question of whether enterprises will pay $30 per user per month for a productivity assistant. Microsoft's own usage disclosure — which has migrated from seat counts to "monthly active usage by knowledge workers" — is a signal that management understands the difference between a licence sold and a product used. A licence sold that is not used is a renewal risk, not a revenue line.

5. Regulation has moved from theoretical to active in four jurisdictions

The European Commission has opened a formal proceeding into the bundling of Teams and Copilot into enterprise agreements. The Federal Trade Commission has a non-public investigation into the acquisition of Inflection AI and the licensing arrangements that accompanied it. The UK Competition and Markets Authority has designated Microsoft as having strategic market status in cloud infrastructure and is examining licensing practices. And the Securities and Exchange Commission has requested information about certain AI-related capital commitments and the corresponding disclosures. Any one of these is manageable. Together they represent a coordinated narrowing of the strategic degrees of freedom that made Microsoft's bundling strategy work.

6. The multiple assumes the mix does not change, and the mix is changing

Microsoft is being valued as a software company. In FY2026, 39.2% of its total assets were property and equipment and 36% of its capital expenditure was funded from operating cash flow that was not available to shareholders. The trend across every disclosed metric points one way, and the market has not repriced the multiple to reflect it. When Microsoft's revenue mix crosses the point where infrastructure exceeds software in operating income — on our base case that happens in FY2029 — the comparable set changes, and infrastructure companies do not trade at 30× earnings.

21 — Risk matrixRanked by probability times impact

The table below ranks risks by our assessed probability of occurrence over a twenty-four-month horizon multiplied by the impact on our base-case value of $512. Probabilities are Farstar judgements, not company disclosures. Impacts are expressed as the approximate effect on intrinsic value per share against the base case.

Value effects are not additive. They are individually estimated against the base case and several would occur together, in which case their combined effect is smaller than the sum because the bear-case multiple is already applied. The probability-weighted expected value of this register, treating each risk independently, is approximately −$34 per share against the base case, which is contained within our bear-to-base range of $355 to $512.

Catalysts

The dates below are the points at which the questions raised in this report become answerable. We order them by information value rather than by proximity.

22 — ConclusionConstructive at $512, on a franchise paying for its own furnace

Microsoft at $497.75 is a genuinely balanced proposition, and the balance is unusual enough that it is worth stating plainly. The company has the best distribution in enterprise software, the second-largest cloud infrastructure business, a $678 billion commercial backlog that is growing 27% year over year, an equity stake in the most valuable private company in the world, and a balance sheet that is close to unlevered. It also has a capital programme that will consume more cash in FY2027 than the entire company generated in operating cash flow five years ago, an accounting policy that has been adjusted in the direction of higher reported earnings, and a regulator in Brussels that has opened a formal proceeding on exactly the mechanism that makes its software strategy work.

The case for owning Microsoft here rests on three things that are not seriously contested. First, Azure is growing at 43% in constant currency with flat gross margins and accelerating non-AI growth — that is a franchise gaining share, not defending it. Second, Copilot's 6.7% attach rate on a 450-million-seat base means the product's economics are still almost entirely ahead of it: every point of attach is roughly $1.6 billion of annualised revenue at software margins on infrastructure that is already paid for. Third, the balance sheet means Microsoft can continue to invest through a downturn that would force a leveraged competitor to retreat, and the value of that option rises with the length of the cycle.

The case against rests on three things that are equally uncontested. First, depreciation is arithmetic and the arithmetic is unfavourable: the transition from $43.8 billion of annual depreciation to something above $88 billion by FY2029 has to be absorbed by pricing, mix, or operating leverage, and none of the three is guaranteed. Second, the useful-life extension and the finance-to-operating lease reclassification both move reported earnings in the same favourable direction at the same moment the cash economics are deteriorating — a coincidence that deserves more scrutiny than it has received. Third, Microsoft's AI revenue is not separable from OpenAI, and the relationship has been restructured three times in a direction that consistently favours OpenAI's independence.

Our target of $512 sits 2.9% above the market and 10.7% below consensus. We are not making a call on the direction of the shares over the next quarter, and we are explicitly not arguing that Microsoft is cheap. We are arguing that the market has priced a single scenario — continued mid-thirties Azure growth with no depreciation consequence — and that a 25/50/25 weighting of $355, $512 and $690 is a more honest reflection of what can actually be verified from the company's disclosure than either the consensus $573 or the bear-case $340 that a capital-cycle-agnostic reading would produce.

23 — AppendixFinancial summary tables

A. Annual income statement summary

FY2027E figures for revenue and EPS are S&P Global consensus as of 17 September 2026; the segment, margin and tax detail within FY2027E is a Farstar construction derived from disclosed segment growth assumptions and is not consensus. FY2026 diluted EPS reflects the reduction in share count through repurchases; FY2025 EPS benefits from the change in the treatment of the OpenAI equity-method loss discussed in section 10.

B. Cash flow and capital allocation

FY2027E cash flow lines are Farstar constructions consistent with the capex framework discussed in section 11 and the consensus revenue estimate. Free cash flow in FY2027E falls to $41.0 billion on our assumptions, which is 64% below FY2025 despite revenue growing 37% over the same two-year period. That compression is the clearest single expression of the capital cycle in Microsoft's financials, and it is why we describe the FY2027 free cash flow yield as 1.11% on the current market capitalisation rather than the 2.50% that a consensus free cash flow figure would imply.

C. Segment detail, FY2026

Segment totals exceed consolidated operating income because segment results exclude certain corporate items, including the OpenAI equity-method income and the amortisation of acquired intangibles. Azure revenue is a Farstar estimate derived from the disclosed constant-currency growth rate applied to the prior-year base and adjusted for currency; Microsoft does not disclose Azure revenue in dollars. Intelligent Cloud operating income includes the AMD and other capacity prepayments discussed in section 11, which are amortised through cost of revenue over the life of the associated agreements.

D. Quarterly record, four most recent quarters

Commercial backlog is the disclosed remaining performance obligation. Quarterly capital expenditure is calculated from the cash flow statement and excludes finance leases, which are disclosed separately. Free cash flow is operating cash flow less payments for property and equipment on a standardised basis and differs from the company's own presentation, which adjusts for certain items. Source: Microsoft quarterly earnings releases and Forms 10-Q.

E. Key metrics at a glance

24 — SourcesPrimary references and where each figure comes from

The financial figures in this report are drawn from the following primary sources, except where explicitly labelled as an estimate. Where a figure comes from a secondary compilation, the compilation is named.

Company filings and releases

Microsoft's quarterly earnings releases and Forms 10-Q for the quarters ended 30 September 2025, 31 December 2025, 31 March 2026 and 30 June 2026, including the segment tables, the commercial backlog disclosure, and the constant-currency reconciliations; Microsoft's fourth-quarter and full-year FY2026 earnings release dated 30 July 2026 and the accompanying earnings-call transcript; Microsoft's FY2025 and FY2026 Forms 10-K, including the property and equipment note, the useful-life disclosures, the lease note, the income taxes note and the commitments and contingencies note; Microsoft's Forms 8-K relating to the OpenAI restructuring dated 28 October 2025; Microsoft's investor-relations disclosures on the Maia 200 programme and the FY2027 capital expenditure framework; the Microsoft annual report to shareholders for FY2026.

Regulatory and third-party sources

The European Commission's formal proceeding into the bundling of Teams and Copilot within Microsoft 365 enterprise agreements; the UK Competition and Markets Authority's designation and investigation of Microsoft in cloud infrastructure services; publicly available statements of the US Federal Trade Commission regarding the Inflection AI transaction; OpenAI's corporate announcements and the published terms of the October 2025 restructuring; NVIDIA, AMD and Broadcom quarterly disclosures for accelerator revenue and supply commentary; Alphabet and Amazon quarterly cloud disclosures used for relative growth and margin comparison; S&P Global Market Intelligence for consensus estimates, standardised financial statements, daily closing prices and peer multiples; and analyst target compilation data as of 17 September 2026.

Estimates and limitations

Several figures in this report are Farstar estimates rather than company disclosures, and they are labelled as such where they appear. Specifically: Azure revenue in dollars is derived from the disclosed constant-currency growth rate because Microsoft does not disclose the figure; the split of FY2026 capital expenditure between recurring and catch-up categories is a Farstar estimate; the quantified effect of the FY2026 useful-life extension on operating income is a Farstar estimate derived from the disclosed change in estimated lives applied to the gross asset base; the allocation of the $678 billion commercial backlog between AI and non-AI components is a Farstar estimate based on management's qualitative disclosure; the fair value of the OpenAI stake is drawn from published reports of the most recent valuation and is not a company disclosure; the Copilot attach rate is calculated from the disclosed seat count against the commercial Microsoft 365 base, which is itself estimated; and all scenario values are Farstar constructions based on stated assumptions.

Microsoft does not disclose the return on capital of the Azure business separately, so the returns analysis in this report is applied at the consolidated level and adjusted for the estimated infrastructure capital base. Microsoft does not disclose depreciation by asset class, so the useful-life analysis relies on the disclosed estimated ranges and the aggregate depreciation charge. Microsoft does not disclose Copilot revenue, gross margin or churn, so the Copilot economics in section 07 are modelled rather than measured. Where a figure could not be sourced, it is marked "—" or omitted rather than estimated.