Alphabet

On 27 September we published Alphabet at $343.92 under the verdict Accumulate — value it on what it earns. We told readers to buy steadily and to add with conviction near $300. The shares closed at $324.89 on 28 October and changed hands at $292.50 at the end of that evening's trading, after the results. Anyone who followed us has lost 15% in a month. Here is the account — the parts we got right, the part we got wrong, and the parts we never saw.

Our September report said buybacks had fallen by a quarter in 2025 and that debt had risen to $59bn. Both were true and both were stale. We did not tell readers that in June 2026 Alphabet sold $49.6 billion of new securities — Class A and Class C shares at about $352–355 and a mandatory convertible preferred stock paying 6.25% — and opened a programme to sell up to $40 billion more shares into the market. By 30 September it had sold another $12.4bn that way. One of the buyers in June was Berkshire Hathaway, which took $10bn of the new shares directly from the company at about $350. It had bought back none of its own shares in 2026, against $39.8bn in the first nine months of 2025, and its long-term debt had reached $125.1bn. All of that was public in July. We read the income statement closely and the financing section of the cash-flow statement not at all.

★ The lesson is the one we wrote down after Adobe, and failed to apply: the number you choose to judge a business by decides what you see. We judged Alphabet on earnings, stripped of paper gains — a sound adjustment that pointed us at the wrong statement. For a company turning its cash flow into data centres, the statement that matters is the one that shows where the money comes from. This report starts there.

Analysts expected $3.01 a share. Alphabet reported $2.12. The release itself explains the biggest piece: a $8.7bn loss on equity securities — mostly SpaceX, whose shares fell from $170.86 on 30 June to $150.86 on 30 September — reduced net income by $6.7bn and earnings per share by $0.55. Put it back and Alphabet earned $2.67. That is still 34 cents short of the estimate, and the 34 cents are the part worth explaining.

The fair comparison is with the same quarter a year ago, on the same basis. In the third quarter of 2025 Alphabet reported $2.87; gains on its holdings added $0.68, and a $3.5bn European Commission fine took away about $0.29. Remove both and the year-earlier figure is about $2.48. So, on the same basis:

Read the right-hand column from top to bottom. Revenue grew 25%, operating income 17%, earnings per share 8%. The leak between the first and the second line is cost: depreciation of property and equipment up 49% to $8.4bn, research and development up 36% to $20.6bn, and the company-wide AI costs Alphabet does not charge to its segments up 71% to $7.7bn. The leak between the second and the third is new: a share count rising for the first time in a decade, $298m of dividends on the new preferred stock, interest expense of $967m against $143m a year ago, and less incidental income.

None of this is a disaster. An 8% rise in underlying earnings per share, in the quarter a company is building the largest computing fleet in history, is respectable. But it is not what a 25% revenue headline suggests, and it is the first quarter in which the cost of the building — depreciation, interest, dilution — showed up on every line at once. That, not SpaceX, is what the market was reacting to.

Roughly $400 billion of market value disappeared in four hours (our arithmetic: 12.3 billion shares × $32.39). Prices move for reasons and for moods; the honest exercise is to separate them. Here are the six things investors and analysts pointed to on the evening of 28 October, and our judgement of each.

Four of the six reasons are real; one is noise; one cuts the other way. Our answer to the question everyone asked that evening — was the fall justified? — is yes, mostly, but not for the reason in the headlines. If Alphabet had reported $3.01 with the same search growth, the same capex outlook and the same cash flow, the shares should still have fallen. If it had reported $2.12 with search growing 17% and capex flattening, they should have risen.

One more thing is worth saying about the mood. GOOGL had already fallen 10.5% between 13 October ($353.78) and 26 October ($316.80), a slide no single piece of news explains — some brokers trimmed targets, and investors were already uneasy about AI agents answering questions outside Google. The after-hours price of $292.50 is close to the 52-week low of $267.67 set last winter. Mr. Market had been gloomy for a fortnight; the results gave him a reason. That does not make him wrong.

In June we wrote that everything turned on one question: do AI answers earn as much as the ads beside ten blue links? We said nobody could yet know. In September the second quarter's 17% growth looked like a first answer, and we said so. In our closing letter we added a test: if search growth slows sharply while AI answers spread, pay close attention. Search growth has gone from 17% to 12% in one quarter, while AI Mode — Google's chat-style search — grew fast. We are paying attention.

Two things were said on the call, and they point in opposite directions. Sundar Pichai said search query growth was the highest in five years, and that questions typed into AI Mode are about three times longer than ordinary searches. People are not leaving Google; they are using it more. Philipp Schindler, the chief business officer, then explained part of the slowdown in plain words: “the strong growth in AI Mode queries had an impact because of the lower monetization we were seeing in AI Mode.” More questions, each worth less. Ads inside AI Mode are still being tested and are not yet sold outside the United States.

★ This is the bear case of June, in management's own voice — but only half of it. Lower monetisation per query is what an early product with immature ad formats looks like; it is also what a permanently thinner toll looks like. Google's own history favours the first reading — mobile search monetised worse than desktop for years, then better — but we will not know for several quarters. What we do know is that the gap is now visible in the revenue line, and that a business priced for 15%+ search growth has just delivered 12%.

The quarterly report on Form 10-Q, due in the next few days, will show paid clicks and cost per click; in the second quarter they rose 13% and 3%. If clicks keep growing and price per click falls, the story is the young-ad-format one. If clicks themselves slow, it is the other one.

Google Cloud is now a $123-billion-a-year business growing at more than 100%, with an operating margin of 36% — three years ago it was barely profitable. Its backlog — revenue customers have contracted to pay — reached $561bn, and management expects a little over half of it to be billed within two years. On any ordinary measure this is one of the most successful businesses on earth this year.

It has also changed shape. Alphabet now describes Cloud as generating “product revenues primarily from the sale of TPU systems” — its own AI chips, designed with Broadcom, sold as racks to other companies. The customers named on the call include Anthropic, Meta, Apple and Oracle; nearly a hundred companies use TPUs. That is good news for the moat — Google is one of only two or three suppliers of AI computing at scale that does not depend on NVIDIA — and it explains a number that appeared for the first time: $13.3bn of inventory, up from $2.4bn in December. Machines built but not yet sold are cash spent and not yet returned. The chief financial officer, Anat Ashkenazi, said as much: the TPU build will “put pressure on free cash flows”.

Three cautions belong next to the applause. Concentration: the largest buyers of AI computing are a handful of laboratories, and one of them, Anthropic, is also a company Alphabet owns about 14% of; the backlog's quality depends on their ability to keep raising money. Hardware economics: selling chip systems is a different business from renting them — lumpier, lower-margin over time, and exposed to the next generation of someone else's chip. Capital: most of the $51.4bn spent in the quarter went into servers and data centres for this business. Cloud's margin is struck after depreciation; the depreciation of what is being built now has barely begun.

In June we said Google's moat was fed by five springs and scored it 8 out of 10 — vast, but contested at the front door for the first time. Four months and two quarters later, here is each claim against the evidence.

Add it up and the moat score stays at 8, for the reason a Buffett-style analysis must always keep in mind: a moat measures how hard it is for a competitor to take the castle, not how much the castle earns this quarter. Nobody has taken anything from Google. People ask it more questions than ever, its cloud has become one of three or four places in the world to buy AI computing, and its distribution is protected by habit as much as by contract. What has changed is the toll. If AI answers settle at a lower price per question than the old results page, the moat survives and the economics behind it shrink — a castle with a cheaper toll.

Our own platform still scores the moat as wide, with pricing power “strong” because the gross margin has held near 57% for a decade. That judgement rests on history. The third quarter is the first piece of evidence against it, and it is one quarter.

The most instructive comparison of the year happened on a single evening. Microsoft and Alphabet reported within minutes of each other on 28 October. Both are spending at a pace no company has ever spent; both have cloud businesses growing fast on AI demand. Microsoft's shares rose about 5%; Alphabet's fell 10%. The difference was not the cloud — Google Cloud grew 103% to Azure's 48%. It was everything around it: Microsoft's free cash flow, though down 27%, was still positive at $18.6bn, while Alphabet's was negative; and Microsoft's core software is sold by subscription, while Alphabet's core is an advertising toll that slowed.

The second comparison is with Meta, whose advertising grew 26% in the same quarter against Google's 10.6%. Advertisers are spending more than ever; more of the increase is going to feeds and AI-recommended video, and less to the search box. That is not a new trend, but it is a wider gap than at any time in our three reports.

And the rival everyone watches is still small in money. ChatGPT's advertising reached a billion-dollar run-rate in about 200 days — a remarkable start, and roughly what Google Search takes in every day and a half. OpenAI's threat is not to this year's revenue. It is to the habit: the place a young person goes first when they want to know something. On that measure Google's query growth, the highest in five years, says the habit is holding. The price it can charge for each answer is the open question of Part IV.

For twenty years Alphabet was the closest thing in public markets to a perfect toll booth. A search cost almost nothing to serve; the advertiser paid for the click; and the cash piled up faster than the company could use it. In 2016 it spent 11% of its revenue on property and equipment. From 2019 to 2025 it bought back more than $300bn of its own shares and, from 2024, began paying a dividend. Warren Buffett has a name for that kind of business: one that grows without needing much capital, so that most of what it earns can be handed to its owners.

That company is changing into a different kind. In the third quarter Alphabet spent 40% of its revenue on equipment — data centres, servers, its own chips. Over the last twelve months capital spending was $159.8bn against operating cash flow of $179.6bn, leaving free cash flow of $19.7bn, down from $73bn in 2025. In 2026 it will spend $195–205bn, and it has said 2027 will be “significantly” more. To pay for it, it has done three things it had never done together:

Two details tell the story better than the totals. In February Alphabet sold a 100-year sterling bond — £1bn at 6.125%, due in 2126 — the kind of security railways and governments issue when they are building something meant to last a century. And the coupon it paid on 40-year dollar money rose from 5.75% in February to 6.50% in August: the market is charging even the best credit in technology more for each new tranche.

★ Buffett's warning about capital is famous in one form: “The worst sort of business is one that grows rapidly, requires significant capital to engender the growth, and then earns little or no money. Think airlines.” (2007 letter.) Alphabet is nowhere near that: it earns a great deal. But he drew a line between businesses that need little capital to grow and those, like his railway and his utilities, that must keep pouring money in to earn decent returns. He owns both kinds happily — and pays very different prices for them. Alphabet has moved from the first kind toward the second. Its returns on the new money may still be excellent; Cloud's margins suggest they are. What has changed for certain is that its owners now receive less of each year's earnings, and must wait longer for the rest.

The depreciation wave. Equipment bought this year starts to depreciate when it goes into service, over about six years for servers and much longer for buildings. Depreciation of property and equipment was $8.4bn in the quarter, up 49%; net property and equipment has grown from $247bn at the end of 2025 to $373bn in nine months and total assets passed $1 trillion. Even if capital spending stopped growing tomorrow, depreciation would keep rising for years — and every dollar of it comes out of reported earnings. Analysts' forecast of $15.35 a share for 2027 was made before the company said 2027 spending would rise “significantly”.

We have no doubts about the integrity or the ability of the people running Alphabet. Their operating record this year is outstanding: Cloud's growth and margins, Gemini at a billion users, and the search franchise defended in two antitrust cases without a break-up. The question we must ask is a narrower one, and it is the one Buffett asks of every manager: what are they doing with the owners' money, and at what price?

Here the record has a blemish. Between 2022 and 2025 Alphabet spent about $229bn buying back its shares at prices mostly well below today's. In June 2026 it sold new shares at about $348–355 — $20.5bn through banks and $10bn directly to Berkshire Hathaway — and over the summer at prices implied around $338–358, close to the top of its history. Buying low and selling high is what one wants a company to do with its own stock, and in that narrow sense the timing was good. But the reason for selling matters more than the price: the shares were sold not because they were dear, but because the company needed the money to keep up with Microsoft, Amazon, Meta and OpenAI. That is the institutional imperative — Buffett's term for the tendency of companies to imitate rivals' spending whatever the return — and nobody can yet tell whether this race will reward the winners.

There is an irony here that Berkshire's shareholders will appreciate. In 2017 Buffett told his annual meeting that GEICO had been paying Google ten or eleven dollars a click years earlier, and that Berkshire had missed what that meant; two years later Charlie Munger said he felt like “a horse's ass” for not seeing Google sooner. Berkshire finally bought in 2025 — and in June 2026 it wrote Alphabet a ten-billion-dollar cheque for new shares, helping to pay for the very building programme that has since cost those shares about 16%. It is a bet on the railroad, made by the people who know railroads best. We take that seriously; it is not a reason to pay any price.

Stock-based compensation, meanwhile, ran at $8.6bn in the quarter, up 36%, and is no longer offset by buybacks. Owners are now paying staff partly in new shares and paying for data centres partly in new shares too. That is why we lower the Management & Capital dial from 8 to 6: not a judgement of character, but of what the owners get back.

The dividend is safe and irrelevant to the investment case: at $292.50 it yields about 0.3%. What matters about it this year is what it reveals — a company paying a cash dividend while borrowing and selling shares is, for the moment, moving money from new investors to old ones. That is acceptable for a year or two of building. It would not be acceptable as a permanent arrangement, and Alphabet has not suggested it will be.

In September we valued Alphabet on its operating earnings, about $10 a share, and called $300 — twenty times 2027 forecasts — a price to buy with conviction. That method assumes the earnings are available to the owners. For most of Alphabet's history they were. For the next two or three years they will not be: the cash will go into data centres, and the owners will be asked for more. So this time we value the company the way Buffett defines value — the cash that can be taken out of the business over its life, discounted — year by year to 2031, then on the owner earnings of a normal year after the building slows.

What does $292.50 assume? On our arithmetic, something between our central and high cases: roughly our central business valued at about 22 times its 2031 owner earnings, or operating margins three points higher than we assume and a multiple near 20. That is not absurd. It requires Cloud to keep converting its backlog, search to resume mid-teens growth as AI ads mature, and capital spending to start falling as a share of revenue from 2028. All three are possible. None is yet visible, and the company has just told us that the spending will rise further before it falls.

The model is most sensitive not to the discount rate — 9% or 11% moves the central value only to about $249 or $229 — but to two things: the size of the 2027 capital budget, which Alphabet will announce in February, and the margin it earns once the depreciation of this building arrives. Raise the 2031 margin by three points and the central value is about $265; let capital spending flatten in 2027 instead of rising, and it is about $248.