Microsoft Corporation

On 23 June we published our first report on Microsoft at $367.34. It was the first analysis on this site, and it is the one we have pointed most readers to since. It said the business was wonderful, the price was fair and the margin of safety thin, and it concluded "Keep watching" — begin buying in earnest in the low $300s, with a price alert at $310.

The share price today is $497.93. Before we say anything about Microsoft, we owe you an account of ourselves.

Four right, two half-right, and two wrong — and the two wrong ones are the ones that cost a reader money. We understood the business. We misjudged what to do about the price. That is a more useful failure to study than most, and it gets its own part before we go on.

The dials move accordingly: Management 9 → 8 (the capital-spending bet has grown and the dividend slowed), Financial strength 10 → 9 (net cash has become net debt once leases are counted), and Valuation 6 → 4 on a price 36% higher. The global score goes from 8.3 to 7.6 — the business did not get worse; the price got fuller.

Read the June scorecard again. Its headline was "A wonderful business — now at a fair, not foolish, price." Its valuation section put the honest intrinsic range at $300–$370. And then its verdict asked readers to wait for the low $300s, and set an alert at $310 — below the bottom of the range we had just called fair.

Those three sentences cannot all be acted on. If the price is inside your estimate of value for a business you would like to own forever, the instruction to wait for fifteen per cent less is not caution. It is a second, stricter test that nobody set out to apply, and it was applied to the one company on this board we were most certain about.

Why it was the mistake, not bad luck. Prices fall after publication all the time; that is not what went wrong. Three things did.

First, the second test. We had already done the work that matters — is the business wonderful, and is the price inside what it is worth? — and answered yes to both. Asking for fifteen per cent more on top was not a further margin of safety on an uncertain estimate. It was a margin of safety on a margin of safety, applied to the business we understood best.

Second, the arithmetic of waiting for a compounder. Microsoft grew earnings 22% last year. If a business's value grows at roughly 20% a year, a 15.6% discount is about eleven months of its growth: wait that long and the "bargain" price you asked for is simply what the business was worth when you started waiting. If the price merely tracks the value, the alert never triggers. With a wonderful business the risk of waiting is not that you pay too much; it is that you never own it.

Third, the anchor. Our "conservative" floor came from our data feed's discounted-cash-flow model, which said $301 in June. We have since learned, across dozens of reports, that this model breaks in predictable ways — on Microsoft it projects a capital-spending peak forever — and we now print it as a fault rather than use it. It says $270 today, after a quarter in which the business got better. We were measuring the building programme and calling it value.

★ "It's far better to buy a wonderful company at a fair price than a fair company at a wonderful price." We quoted the spirit of that line in June and then did the opposite. Buffett has told the story of stopping his purchases of Walmart in the 1990s because the price rose slightly — a piece of thumb-sucking he later reckoned cost Berkshire around $10 billion. Ours cost less and was cheaper to make: a few words in a verdict. But it is the same mistake, and it is the more common one for careful investors, because it feels like discipline.

What we should have written. The June analysis had the ingredients for a better call: a business scored 9, 9, 10 and 9 on four of five dials, a price inside our own range, and a thin but real margin of safety at the bottom of it. The right verdict was "Begin buying; add on weakness" — a starting position at a fair price, with the weakness as an opportunity to add rather than a condition for starting. At $349.20 two days later, that reader would be up 43%.

That rule applies to this report too, and it is why the zone we name at the end is a fair price, not a bargain one. It also carries a risk of its own, and we would rather name it than discover it: a rule that says "start" at a fair price will, some of the time, have readers buying just before a fall. We accept that trade for businesses of this quality, because a fair price paid for a compounder tends to be repaired by time, and a missed compounder is never repaired. When the rule costs a reader money, we will grade that in public too, the same way.

Microsoft's fiscal year ended on 30 June. It reported on 29 July: revenue $331.8bn (+18%), operating income $155.2bn (+21%, a 46.8% margin), net income $128.8bn excluding OpenAI (+22%), $17.28 a share. In the fourth quarter alone revenue was $90.0bn, Microsoft Cloud $59.3bn (+27%) and Azure +43%; management guided Azure to about +45% in the current quarter. The shares rose about 8% the next day. By product line, fiscal 2026:

★ The company has become, more than ever, one engine. Server and cloud is now 39% of revenue and nearly all of the growth; the consumer business — Windows, Xbox, devices — shrank. That is not a problem in itself: the engine is the best one in enterprise computing, and the Office franchise under it is growing 16% on price and Copilot. But it means the whole valuation now rests on a single question, which is whether the enormous spending that powers that engine earns its keep. Part IV.

That table is the whole debate. Microsoft's operating cash flow rose 34% last year — the business is throwing off more cash than ever. Every extra dollar, and more, went into data centres. Cash capital spending quadrupled in three years, and the figure understates it: the fourth quarter's reported capital spending of about $41bn includes finance leases that our cash-flow data ($35.8bn) does not. For calendar 2026 Microsoft has pointed to roughly $175bn of spending, and says it will grow again in fiscal 2027.

The bill arrives through depreciation, and it has only started arriving. Depreciation and amortisation has nearly tripled since fiscal 2023, and most of the $337bn of property now on the balance sheet was bought in the last two years. The fourth-quarter gross margin was 67%, down on the year, and Amy Hood guided the operating margin to decline slightly in fiscal 2027.

⚠️ An accounting change worth knowing about. From the start of fiscal 2027 Microsoft extended the estimated useful life of its data-centre buildings and offices from 15 to 25 years. Hood called the effect on fiscal 2027 operating income "minimal", and servers and network gear are unchanged at two to six years. We take her at her word. But it is the second time in four years that Microsoft has lengthened the life of its infrastructure — servers went from four years to six in 2022 — and both changes move reported profit in the same direction. An owner should notice the direction even when each step is small.

★ In June the gap between the lenses was the question. Now it is the answer. We argued that free cash flow understated Microsoft because most of the capex was growth, not maintenance — and that if it earned a good return, 30 times owner earnings would look cheap in hindsight. The return is showing up: Azure +43%, a backlog up 84%. But the spending grew faster than the return. A business in which capital spending rises 80% so that revenue can rise 18% is investing heavily ahead of demand; that may be wise, and Microsoft's demand signals suggest it is. It is not, yet, a business whose owners are being paid for it in cash.

Two things are true at once. The first is that Microsoft made one of the great investments of the decade: a stake worth tens of billions, a customer committed to hundreds of billions of Azure, and seven years of access to the leading models. The second is that the most-cited number for Microsoft's future growth — the order book — is now heavily dependent on the finances of a single company that has never made a profit, that is spending on computing at a rate unprecedented in private enterprise, and that since April is free to take new work elsewhere.

A backlog is only as good as the customer's ability to pay it. Microsoft's customers are, overwhelmingly, the most creditworthy organisations on earth. One of them is not like the others. Growth excluding OpenAI of 25% is still excellent, and it is the number we would build a valuation on; the rest is an option that could be worth a great deal or have to be renegotiated.

We said in June that gross margins are the first place a moat's erosion shows. The fourth-quarter gross margin of 67% was down on the year — but because of depreciation on new capacity, not because customers are paying less. That is the cost of widening a moat, not the sign of one narrowing. We keep the score at 9 and keep watching the same number.

The headline number hides the competitive one. Azure grew 43% in the June quarter — and so did the market, which means Microsoft's share of cloud infrastructure was 20%, exactly where it was a year earlier. The share Amazon lost went to Google, whose cloud revenue grew 82% and whose backlog more than doubled in six months, to $514bn; Oracle's grew 121% on contracts with the AI labs. Microsoft has an answer, and it is a fair one: it is short of capacity, so its growth measures what it can build, not what customers want. That is why the capex in Part IV exists. But it also means that in the most important race in its industry, Microsoft is keeping pace, not pulling away — and two of the runners are growing faster.

Two things are worth setting against that. Azure is still a third larger than Google in cloud infrastructure, on Synergy's measure, and earns its money on a broader base — the same enterprises that already run Office, Windows Server and SQL on Microsoft contracts, which is the switching-cost moat of Part VI. And Amazon, the leader, is growing slowest of the three.

Microsoft's strategy is the oldest one it has: put the new thing inside the suite and raise the price of the suite. Google is doing the same with Workspace. On distribution Microsoft wins easily — 30 million paid seats in under three years, sold through contracts nobody else holds. On usage the picture is less flattering: in LayerX's browser data, more enterprise users work in ChatGPT than in Copilot, and nearly half of all workplace AI conversations happen in personal accounts the employer never bought. The fight is not Word against Google Docs, where Microsoft's switching costs hold. It is over who owns the assistant that sits on top of all of them — and there Microsoft's most dangerous rival is the company it helped build.

The price of a given level of AI capability has been falling faster than almost any price in economic history: by roughly ten times a year on a16z's estimate, and by between 9 and 900 times a year depending on the task on Epoch AI's. The open models drive much of it. China's DeepSeek V4, released with open weights on 24 April 2026, charges about $3.48 per million output tokens at the top end, against roughly $25–30 for the leading American models. Meta, once the champion of open weights with Llama, made its newest model proprietary in April; France's Mistral raised €3bn in September at a valuation above €21bn.

For Microsoft this cuts both ways, and we think more one way than the other. Cheap models weaken the pricing power of any single model maker, including the one Microsoft owns a quarter of, and put a ceiling on what anyone can charge for a chat assistant. But Microsoft sells the computing, not just the model. Its Foundry service hosts more than 11,000 models — DeepSeek, Llama and Mistral among them, alongside OpenAI, Anthropic and its own — for 100,000 customers, and its revenue more than doubled on the year; in July Microsoft signed a multibillion-dollar European partnership with Mistral. When intelligence gets cheaper, more of it gets used, and every token still runs on somebody's data centre. Microsoft has positioned itself to be the landlord rather than a tenant.

In 2023 OpenAI was Microsoft's moat in AI. In 2026 it is Microsoft's largest customer, a quarter-owned holding, and a rival for the same enterprise budgets — selling its models on Amazon's cloud, taking Amazon's money, and building its own agent platform for the companies Copilot is trying to sell to. Microsoft has hedged sensibly: it uses Anthropic's models in Copilot, it is building its own, and it has signed Anthropic to Azure. The relationship is still enormously valuable. It is no longer exclusive in either direction, and the moat table in Part VI marks it, rightly, as narrowing.

Nothing in this part changes our verdict today, and we want to be plain about why. The competition has intensified on every front we can measure: Google is growing twice as fast in cloud, OpenAI is selling around Microsoft as well as through it, and open models are pushing down the price of the product everyone is building. But none of that has yet shown up where it would matter — in Microsoft's share, its prices, or its customers leaving. Its share held, its prices went up, and its backlog excluding OpenAI grew 25%. A moat is tested when competitors arrive, and they have. The score stays 9; the signals above are the ones that would move it, and we will report on each in our next follow-up, after the October results — a new report, not an edit to this one.

Integrity: no veto. Capital allocation: this is where we mark it down a point. Nothing Microsoft is doing is reckless — the spending is funded from operations, and the demand appears real. But the scale of the bet has grown faster than its disclosure: the single largest customer is not named in the backlog, the OpenAI economics have changed twice in a year, and the year the capex nearly doubled is the year the dividend grew by the least since 2017 and an accounting estimate was lengthened. None of that is wrong. All of it points the same way, and we would rather notice than not.

★ On 15 September 2026 Microsoft's board declared a quarterly dividend of $0.98, up from $0.91 — an increase of 7.7%. Payable on 10 December to holders of record on 19 November. Our data pull still showed $0.91. The annual rate is $3.92, a yield of 0.79% at $497.93.

It is the same signal we read at Johnson & Johnson yesterday, where the dividend rise slowed to 3.1% while the board kept cash for acquisitions — and the opposite of Rexford, where a slowing dividend preceded bad news. At Microsoft the business is accelerating and the dividend is decelerating, because the company believes it has better uses for the money. That is a legitimate choice. It is also a bet, and a sub-1% yield means nobody should own Microsoft for its income.

The question for an owner is not whether AI demand is real — today it plainly is — but what happens to the cash if it grows more slowly than the building. The 10-K answers half of it. On 30 June Microsoft had $169.0bn of purchase commitments, mostly for data centres, and $29.8bn of construction commitments falling due within a year: $198.8bn, against $115.9bn of cash capital spending in the whole of fiscal 2026. Beyond those it had signed $329.1bn of further leases, "primarily for datacenters", that have not yet begun; they start between fiscal 2027 and 2033, with terms of up to twenty years. For the first quarter of fiscal 2027 management guided capital spending to over $50bn — in a single quarter, including finance leases. A slowdown would not stop this spending quickly. Much of it is already signed.

The grid does the other half with plain arithmetic. If demand holds and operating cash flow keeps growing in the twenties, Microsoft can absorb $175–200bn a year of capital spending and still produce $30–55bn of free cash flow. If demand slows and operating cash flow stalls at last year's $183bn while the contracted build continues, free cash flow goes to roughly zero or below — and the ~$29bn dividend we call safe in Part X would be paid from the balance sheet rather than from the year's cash. Microsoft can afford that for years: it is rated AAA, and the build can be slowed beyond the contracted horizon. But at 55 times free cash flow, the share price assumes the top row of the grid, not the bottom.

OpenAI is now Microsoft's largest single customer by a distance — a customer bigger than Windows — and also its partner, a holding worth tens of billions, and, increasingly, a competitor for the same enterprise budgets (Part VII). None of those relationships is bad on its own. Together they mean that one company's ability to keep raising money and paying its bills now touches Microsoft's revenue, its backlog, its receivables and its reported earnings at the same time.

How to read the docket. None of these cases threatens Microsoft's solvency, and no regulator has fined it or filed a case on any of them. What they share is a direction: the EU, the UK and the US are all looking at the same thing — how Microsoft's licensing and bundling keep customers on Azure and Office — which is to say, at the switching costs we score as the heart of its moat. A regulator cannot easily make Office worse. It can make leaving cheaper. Over a decade, that is the legal risk that matters, and it moves slowly; the Azure gatekeeper decision this autumn is the first one to watch.

Gravity, too. On 16 September the Federal Reserve raised rates for the first time since 2023; on 23 September the ten-year Treasury reached 5.12%, the highest since 2007. Microsoft's earnings yield at $497.93 is 3.5% and its free cash flow yield 1.8%. For the first time in many years, the government pays more than Microsoft earns on your purchase price — a comparison that matters most for exactly the kind of long-duration growth company Microsoft now is.

So what does 28.8 times assume? That earnings keep compounding above 20% for several years — consensus has fiscal 2028 at $23.48, +36% on fiscal 2026 — and that the capex converts into free cash flow on roughly the timetable Microsoft implies. If both happen, the shares are fairly priced today and cheap in three years. If the capex peak lasts longer, or OpenAI's third of the backlog has to be renegotiated, 29 times earnings for a company whose owners receive 1.8% in cash is a thin cushion, and a 5% Treasury makes it thinner.

★ Applying our own new rule. Microsoft still scores 9 on moat, growth and financial strength. At $497.93 it is above the fair value we can defend, so waiting is allowed — this is not the June situation. But the zone we name is where it becomes fair again, not where it becomes a bargain: about $430. If it gets there, a reader should begin buying rather than wait for less.