← Blog

Income or Promises - Five Predictions for Microsoft, Alphabet and Amazon

Lida Liberopoulou ·27 October 2026 · CC BY-SA 4.0

Raw .md ↗ Share on LinkedIn ↗ Share on X ↗

Independent research on AI economics and infrastructure · Commissioned research

Microsoft, Alphabet and Amazon report their results for July to September on 28 and 29 October. They are spending more on AI than ever, in some cases more than their businesses bring in, and covering it with borrowing, new shares and bills signed for years ahead, as Reuters and others have reported. If their numbers look better this week, is that new income, or has the cost just moved somewhere harder to see? Five predictions, filed before the results, with the tests to check them.

In the last few months, two of the biggest tech companies, Meta and Alphabet, reported record sales while their free cash flow, the cash left after paying for new equipment, turned negative. For Alphabet, it was the first time since it went public in 2004. For Meta, the first time in years: the closest it had come before was in 2022, when its free cash flow fell to $173 million, still just above zero.

That is unusual, and a little worrying, but it doesn’t necessarily mean they are in trouble. Keeping up with a new technology takes heavy investment, and every technology boom has needed it. But how these two companies ended up here, and how they plan to get out, says a lot about whether the spending is working. These companies, and a few others like them, pay for most of the AI you use. Their shares sit in most pension funds and index funds. If the AI build-out pays off, it shows up here first. If it doesn’t, the first signs show up here too, usually in the small print.

This week is a chance to see whether the biggest players in AI will show how they are spending this money and what they are getting for it, or whether they keep that to themselves. Microsoft and Alphabet report on Wednesday, Amazon on Thursday. Before they do, I’m writing down what I expect their accounts to show, and how to tell whether any good news is real. In July, I did the same with six claims, and all six held. Let’s see what happens this time.

To go straight to the predictions, skip to them here. The sections before explain what they test.

How can a company sell more than ever and still run short of cash?

The answer is in how the cost of equipment is counted.

Say a bakery buys a €60,000 oven that is expected to last six years. It doesn’t count the whole price as a cost straight away. It counts €10,000 a year, spreading the cost over the oven’s life. So in the year it buys the oven, the bakery can report a healthy profit while its bank balance takes a dive. The profit counts one slice of the oven. The cash paid for all of it in one go.

The largest technology companies are now buying their “ovens” on a scale never seen before: data centres, chips, power. In the same quarter two years ago, Meta spent $8.3 billion on equipment. Last quarter, it spent $33.9 billion, while the slices it counted as cost that quarter, for all its equipment old and new, came to $6.8 billion. That’s how a company can report record sales and still spend more cash than it brings in.

The bakery also decides how long the oven will last, and so do the tech companies for their servers: two to six years at Microsoft, six at Alphabet, five to six at Amazon. Most servers still work after six years. The question is whether they still earn their keep. A customer picks whichever chip does the job for less, and a new generation of AI chips arrives every year or two, each doing more work for the same electricity. An old chip still works, but it has to get cheaper to stay rented, and it earns less every year. If it really pays its way for three years but is counted over six, each year’s profit looks better than it should, until the company admits it and the cost lands all at once. Amazon has already done this once: in 2025, it cut the stated life of part of its servers from six years to five, pointing to how fast AI was moving.

That’s why the predictions also include a check on whether anyone changes their estimate, and why I keep a monthly record of what older chips actually rent for.

Haven't they always spent like this?

For most of the last decade, these companies had the opposite problem: more cash than they could use. When a company has money left over that it can’t put to better use, it often buys back its own shares, which pushes the share price up and rewards the people who own them, including most pension funds. Alphabet spent $28.3 billion on buybacks in the first half of 2025. Meta spent $26.2 billion buying back its own shares in the first nine months of the same year.

In 2026, both stopped. Alphabet bought back no shares in the first half of the year, Meta none in the first nine months. The spare cash went into data centres instead. Microsoft is the exception: it still returned $10.2 billion to shareholders in the quarter to June.

But keeping the spare cash wasn’t enough, so they started raising new money to pay for the building. They did it in two main ways: borrowing, which has to be paid back with interest, and selling new shares, which leaves every existing owner with a slightly smaller piece of the company. The borrowing itself wasn’t new. They have always borrowed some money, often simply because it was cheap. What changed in 2026 is how much, and why. Selling shares was new. For a decade they had been buying them back, and the reversal came at a scale nobody had seen from them before.

Between the end of 2025 and the end of June 2026, Alphabet’s borrowing more than doubled, from $46.5 billion to $98.2 billion. In June, it sold $49.6 billion of new shares, and it has set up a way to sell up to $40 billion more over time, which it says is mainly to cover taxes on staff share awards. Meta’s borrowing stood at $80.9 billion at the end of September, and on 21 October it created a new type of share that its board says can be used to raise money. On 26 October it went further, launching a sale of up to $35 billion in new shares: $12.5 billion in shares that turn into ordinary shares in 2029, and $22.5 billion sold now through banks, which Meta can collect at any time until November 2027. Among those who said they might buy were Nvidia, AMD and Broadcom, the companies that sell Meta its chips, for up to $5 billion.

Amazon, which reports on Thursday, is heading the same way. It reports its cash over twelve months rather than one quarter, and over the twelve months to June it spent $7.6 billion more than its business brought in. A year earlier, it had brought in $18.2 billion more than it spent. Between the end of 2025 and the end of June 2026, its borrowing nearly doubled, from $65.6 billion to $128.9 billion.

So the pattern is not just heavy spending. It is a shift from handing cash back to raising it.

But they have money, so why does it matter?

This is the fair objection, and it deserves a straight answer. None of these companies faces an immediate cash crisis. Each holds tens of billions of dollars in cash and investments.

But look at where the money is coming from now. Part of the money Alphabet has in the bank today came from borrowing and selling shares this year, not from its business. Using borrowed money changes the risk. If a company spends money it already had on something and the spending doesn't pay off, the money is gone, but nothing more is owed. If it borrowed the money, the repayments still come due on schedule, whether the spending paid off or not.

Either way, someone has to pay for all this. Lenders must be repaid, on schedule. And every new share means existing owners, including pension funds, get a smaller slice of every future profit. If the spending makes the whole company earn more, that’s a fair trade: a smaller slice of a bigger pie. If it doesn’t, owners are left with a smaller slice of a pie that didn’t grow, and the lenders still get paid first. So the question is whether the spending turns into income before the bills arrive.

Where does the rest of the bill go?

So far, this is all money already spent. But there is a third pile: bills agreed now and paid later. If the bakery signs a ten-year lease on a new shop, nothing leaves its bank account this year, but it has promised years of rent. The cash figures don't show that promise. These promises sit in the notes at the back of the filings, and they are growing fast.

Meta has $454.5 billion of commitments it cannot cancel, mostly for computing power rented from other companies and for servers. Of that, $45.9 billion falls due before the end of this year, and $121.4 billion next year.

At the end of June 2025, Microsoft had signed $92.7 billion of leases, mostly for data centres, that had not started yet. By March 2026, it was $196.6 billion. By June, $329.1 billion: $132.5 billion added in a single quarter. The leases start over roughly the next seven years, and a new phrase appeared in the filing: some are "subject to certain contractual conditions being met." In March, the latest start date was 2031. Three months later, it was 2033.

There is even a second kind of commitment: equipment already delivered but not yet paid for. At the end of June, Alphabet owed suppliers $29.1 billion for equipment it had already received, against $10.6 billion a year earlier. At Amazon, the figure rose by $29.3 billion over twelve months.

Promises run the other way too. Microsoft, Alphabet and Amazon rent out computing power and sell software to other businesses, often on contracts lasting years. Think of a café that signs up to buy the bakery's cookies for the next five years: the bakery hasn't been paid yet, but it has a promise. These promises are called the backlog. Microsoft's total backlog nearly doubled in a year, to $684 billion. But less of it is due soon: about 30% within the next twelve months, down from about 40% a year earlier. The part due later grew three times as fast as the part due soon. Alphabet's is $519.5 billion, just over half of it due within two years. The promises got bigger, but a growing share of them sits years away.

And some of the biggest customers behind these promises are the AI labs Anthropic and OpenAI, companies these same firms have invested in. Part of the money those customers spend came from the sellers in the first place, as investment in the lab.

So each quarter, more of the money goes into things that have not happened yet: bills to be paid later, and revenue to be received later. The question is whether that turns into income, or stays a promise.

The predictions

The short version, if you skipped ahead: these companies are spending more cash than their businesses bring in. They are borrowing heavily and selling new shares to keep going, and signing bills for years ahead. In return, they hold growing promises of future revenue, some from AI labs they have invested in. The claims below test whether any good news this week is real, or whether the cost has just moved somewhere harder to see.

These are five things, written as the questions a reader would ask, each with a test underneath. They are about where the money comes from and where it goes. Claims 2, 3 and 4 are the likelier ones; 1 and 5 are closer calls.

1. If Alphabet's cash gets better, who is carrying the bill? If Alphabet's cash looks better this quarter, I expect the improvement to come from timing, not from the business earning more: paying suppliers later, customers paying earlier, or renting instead of buying. Each makes this quarter look better by moving money between now and later.

Applies if: Alphabet’s free cash flow is better than last quarter’s negative $5.9 billion.

Pass: at least one of these holds alongside the improvement:

the amount Alphabet reports for equipment received but not yet paid ("property and equipment included in accrued liabilities and accounts payable") is above $39.1 billion in its September report (it was $29.1 billion in June);
money customers paid in advance, mostly for cloud capacity and subscriptions (deferred revenue, $10.1 billion at 30 June), rises by more than $1 billion;
equipment spending falls below $44.9 billion while leases signed but not yet started rise above $85.2 billion (their level at 30 June).

Fail: cash improves with none of these. That would mean the improvement didn't come through any of the routes I named. The likeliest reading is that it is real: more money came in, or spending fell without new bills signed for later.

Not tested: cash stays the same or gets worse. This claim is about where an improvement comes from, so with no improvement there is nothing to test. I will report the result anyway, and a worse figure would say something on its own.

2. Will a big part of the profit come from investments, not from the business? Alphabet holds shares in SpaceX and Anthropic. Amazon holds a stake in Anthropic. When those companies are revalued, the change shows up in their profit, up or down, even though no actual cash changes hands. Microsoft works slightly differently: it books a share of OpenAI’s own profits and losses, and gains when OpenAI raises money at a higher value. In Alphabet’s last quarter, gains on its stakes came to $99.0 billion before tax. Its whole net income was $112.1 billion. In Amazon's last quarter, investment income came to $53.4 billion, against $27.5 billion of profit from running the whole business. And SpaceX has traded on the stock market since June: at the end of June, Alphabet's SpaceX shares were worth $94 billion, and by the end of September SpaceX's share price had fallen by about 12%. Part of that is likely to show up in Alphabet's profit this quarter, as a loss.

Pass: for at least one of the three, income or loss from outside the business is larger than a quarter of profit before tax, both counted by their size, whether positive or negative. The accounts call this “other income”. It also includes smaller items such as interest, but at present it is driven mainly by these investments. A loss counts too: a markdown moves profit as much as a mark-up.

Fail: for all three, income or loss from outside the business stays below a quarter of profit before tax, counted the same way. That would mean their profits this quarter came mainly from the business itself, not from investments going up or down.

3. Will anyone say how much of the backlog comes from the AI labs they fund? Some of the biggest customers behind their backlogs (revenue customers have signed up to pay later) are Anthropic and OpenAI, the same labs these companies have invested in. If a large share of the promised revenue comes from them, part of it would be paid with money these same companies invested in the labs in the first place. Microsoft has said how fast its business backlog grew without OpenAI: 84% overall, to $678 billion, but 25% leaving OpenAI out. Since it stood at $368 billion a year earlier, that means OpenAI accounts for at least $218 billion of it, about a third. That is my calculation: Microsoft has never given the figure directly. In July it also said that nearly 90% of its cloud revenue came from customers other than "frontier model companies", and that all of the last quarter's backlog growth came from outside them. That is revenue and growth, not a share of the backlog. I expect none of the three to say it this time either.

Pass: none of the three gives a dollar figure or a percentage of its backlog that comes from Anthropic or OpenAI, in the release, the earnings call or the quarterly filing. A growth rate without OpenAI, like Microsoft’s, doesn’t count, and neither do figures in Anthropic’s own filing: the test is what the three companies choose to show.

Fail: any of them gives the share of its backlog that comes from Anthropic or OpenAI, in dollars or as a percentage, either named or grouped together as "frontier model companies" or similar.

The 23 October piece runs the same test from Anthropic’s side.

4. Will Microsoft’s bills for later keep growing? Microsoft rents many of its data centres rather than owning them, and signs the leases years before the buildings are ready. Those leases signed but not yet started don’t appear in its spending or its cash flow until they begin, but the payments are already promised. No money has changed hands on most of them yet, but the developers building the data centres often borrow against these leases, so real money is already being spent on the strength of them. They went from $92.7 billion in June 2025 to $329.1 billion in June 2026, and $132.5 billion of that was added in the last three months alone. I expect them to keep growing.

Pass: Microsoft’s leases signed but not yet started are above $329.1 billion at 30 September.

Fail: they are at or below $329.1 billion, meaning Microsoft signed fewer new leases than the ones that started or were cancelled.

Whichever way it goes, I will report how much it changed. A smaller jump than last quarter's would be news too. And a rise doesn't mean every deal is healthy: a lease in trouble stays in this figure until it is formally ended.

5. Will Microsoft’s spending look smaller while the building doesn’t slow? In July, Microsoft said that from 1 July it would treat its data centres and offices as lasting 25 years instead of 15. That changes how some new leases are counted. A lease that covers most of a building’s life is treated like buying it with a loan: if the bakery leases an oven for almost all the years it will work, it has in effect bought it on credit. Such a lease counts in the spending figure Microsoft reports. A lease that covers a smaller part of it counts as ordinary rent, which is left out. With buildings now expected to last longer, more new leases fall into the second group. The total cost to Microsoft is the same either way. What changes is the headline spending figure that most coverage watches. It can slow down without less being built. At a time when investors are nervous about how much is being spent, a slower number reads as restraint, whatever the reason for the change. Microsoft has said this openly: its finance chief told investors the change would move more leases out of the spending figure. The question is how much, and whether the coverage notices.

Pass: in the quarter to 30 September, leases counted as purchases (finance leases) make up a smaller share of the leases Microsoft started using than in the same quarter last year. The figures come from the lease table in its quarterly report ("right-of-use assets obtained in exchange for lease obligations"): finance leases divided by finance plus operating leases, for the three months.

Fail: that share stays the same or rises.

A pass would show the share fell, not on its own why. I will report whether Microsoft links it to the change in building life.

And one control: will anyone change how long their servers last? Back to the bakery: if it decides the oven will last ten years instead of six, its profit looks better, though nothing about the oven or the cash has changed. The same goes for servers. Their latest annual reports say two to six years for Microsoft, six for Alphabet, and five to six for Amazon. I expect none of the three to change that this quarter. I call this a control because it checks that any change in profit isn’t coming from the estimate itself. Microsoft’s change to its buildings (claim 5) doesn’t count here; this is about servers.

Pass: none of the three changes its stated server life.

Fail: any of them does, in either direction. Lengthening would flatter profit. Shortening, as Amazon did in 2025, would mean admitting chips wear out faster.

Four things I will note without pass or fail

These are harder to turn into a clean yes or no, but they say a lot about how the spending is going.

How each company justifies the spending. There are roughly three ways. The strongest is a figure for what AI already brings in today, such as annual AI sales. (Not profit: none of them gives that.) The next is a number about the future: the backlog, a big customer's commitment, or a sales target for a coming year. The weakest is confidence with no figure attached (“we see strong demand”). Meta's call last week leaned mostly on this one. Which one a company uses tells you how much it can actually show. None of the three reports AI revenue as a figure of its own. Much of what they call AI is sold bundled into products people already pay for, like Gemini inside Google One or Copilot inside Microsoft 365. So even the strongest answer, an AI sales figure, is the company's own calculation, and can't be checked against anything. Microsoft also raised the prices of its business Microsoft 365 plans from 1 July for new and renewing customers, citing AI among the reasons, so this is the first quarter in which part of that business pays the higher prices. I will note whether Microsoft separates the effect of price from new customers or AI use.
How the spending is paid for. Whether buybacks stay paused at Alphabet, whether Microsoft keeps paying its shareholders through dividends and buybacks, and whether any of the three announces new borrowing or share sales.
How much is built but not yet working. Companies report the value of data centres still under construction separately from equipment already in use. If the first grows faster than the second, more money is going into things that don’t earn anything yet.
Whether anyone mentions data-centre leases being delayed, renegotiated or cancelled since the end of September. Those wouldn’t show in this quarter’s figures yet, but they could come up on the call.

How this will be scored

Each claim is scored from the earliest company source that settles it: the results release, the investor slides, the earnings call or the detailed quarterly report (the 10-Q). Claim 2 can be scored from the release. Claim 1 starts there too, because the release gives the cash figure, but the tests under it need the quarterly report. Claims 4, 5 and the control also need figures from its notes, and claim 3 needs all the sources, since showing something is missing means checking everywhere it could appear. The quarterly reports come out the same day or within a few days. The results will go out in a follow-up piece, hits and misses alike.

In July, six claims about Microsoft, Apple and Amazon all held. On 23 October, four more went on record about Anthropic’s listing filing. That filing went public on 26 October, the day before this piece. The claims here don't depend on it, and claim 3 leaves it out on purpose. Two of the four Anthropic claims can already be checked, and both held; the full scoring comes on 2 November. A past result is no guarantee. It is the reason to do it the same way.

Why Meta is not in the claims. Meta reported on 21 October, six days before this piece. Any claim about it now would be written with the answer in hand. Meta is also built differently: it doesn't rent out computing power or hold large stakes in AI labs, so two of the claims wouldn't fit it anyway. It carries the bills without the promises on the other side. Two of the observations above, how the spending is justified and how it is paid for, were sharpened after reading Meta’s results and call, and I say so for that reason. The fourth, on delayed or cancelled leases, comes from news since the end of September. The follow-up will run Meta through the same tests, marked as checked after the fact.