---
title: "Income or Promises - Where the Bill Went"

author: "Lida Liberopoulou"

date: "2026-10-31"

license: "CC BY-SA 4.0"

canonical: "https://threadbaire.com/blog/posts/income-or-promises-where-the-bill-went.html"

description: "Five predictions and one control, filed before Microsoft, Alphabet and Amazon reported, scored against their filings. Four held, one could not be tested. The more useful answer sits in the notes, where more of the bill now lives: guarantees, collateral, and companies these firms part-own."

related:
  - income-or-promises
  - follow-the-money-both-ways
---

*Microsoft, Alphabet and Amazon are spending more on AI than ever, and paying for it with borrowing, new shares and bills signed for years ahead. Before they reported last week, I wrote down what I expected their accounts to show, with a test for each prediction, on one question: if their cash looked better, would that be new income, or a cost moved somewhere harder to see? Four predictions held, though for one of them I missed something in a filing that I should have caught. One couldn’t be tested. The more useful answer came from the small print, where more of the bill now sits.*

------

The three biggest companies renting out computing power for AI are spending money faster than their businesses bring it in. This year, Alphabet’s free cash flow, the cash left after paying for new equipment, turned negative for the first time since it went public. Amazon’s is negative over the past twelve months. To keep building, they have stopped buying back their own shares, borrowed heavily, sold new shares and signed leases for data centres that won’t open for years.

That makes their results harder to read. Good news can come from the business earning more, but it can also come from moving a cost somewhere the headline figures don’t show: paying suppliers later, renting instead of buying, or promising to pay in future instead of now.

So on 27 October, before the results, I [wrote down five predictions](https://threadbaire.com/blog/posts/income-or-promises.html) about which of these would show up, plus one check on the accounting itself. I did the same with six predictions in July, and all six held. This is how the new ones did, and then what the filings showed that the predictions didn’t ask about.

To go straight to what the predictions missed, [skip to where the bill went](#where-did-the-bill-go).

### The scorecard

Here are the original predictions with their final score.

| Claim | Result |
|---|---|
| 1. If Alphabet's cash gets better, it comes from timing | Not tested: cash got worse |
| 2. A big part of profit comes from investments, for at least one of the three | Pass, narrowly |
| 3. Nobody says how much of the backlog comes from the AI labs | Pass, with a research miss |
| 4. Microsoft's leases signed but not started keep growing | Pass |
| 5. Microsoft's leases counted as purchases fall as a share | Pass |
| Control: nobody changes how long their servers last | Held |

### 1. Did Alphabet's cash get better?

**Not tested**

No. Free cash flow, the cash left after paying for new equipment, was negative $5.9 billion in the quarter to June. In the quarter to September it was negative $9.1 billion. Alphabet brought in $42.3 billion from its business and spent $51.4 billion on equipment.

The claim was about where an improvement would come from. With no improvement, there was nothing to test, so I score it as not tested.

Two things are worth knowing. First, part of the drop is a one-off payment: Alphabet paid a $5.2 billion fine to the European Commission in July, in cash. Without it, cash flow would have been about negative $3.9 billion, an improvement. Second, one of the three things I said could make cash look better did happen: customers paid more in advance. That money rose from $10.1 billion in June to $12.6 billion in September, well above the $1 billion in my test. If cash had improved, it would have passed on that alone. I am not rescoring it. The rule said what it said. Customers paid more up front, and cash still got worse.

To be fair to Alphabet, part of the cash went into stock it plans to sell. Alphabet now sells its TPU chips to other companies, and it is buying them ahead of those sales: it holds $13.3 billion of them and other stock, against $2.4 billion at the end of 2025. Like a bakery buying flour for a large order, that cash goes out before the customer pays. It counts against this year’s cash flow, about $11 billion so far, even though it isn’t spent on Alphabet’s own data centres. When the chips are sold, most of that cash should come back. Alphabet itself named this stock build as a pressure on its free cash flow.

### 2. Did a big part of the profit come from investments?

**Pass, narrowly**

Yes, but only just, and only at Alphabet.

The test: for at least one of the three, income or loss from outside the business is larger than a quarter of profit before tax, counting losses by their size.

- **Alphabet:** a loss of $8.2 billion outside the business, against $32.4 billion of profit before tax. That is 25.4%, over the line by about $130 million. Most of it is SpaceX. At the end of June, Alphabet's SpaceX shares were worth $94 billion. At the end of September, $75.2 billion. Its finance chief said the loss was mainly from SpaceX. Interest on its new borrowing pulled the other way: $967 million this quarter, against $143 million a year ago.
- **Microsoft:** $38 million, against $46.0 billion of profit before tax. Almost nothing.
- **Amazon:** a loss of $703 million, against $26.9 billion. About 2.6%.

So the claim passed, narrowly, and on a loss rather than a gain.

But Amazon is the more interesting case, because of what that figure leaves out. Amazon owns its stake in Anthropic in two forms. Some of it is shares: when their value changes, the change goes into Amazon’s profit. This quarter that was almost nothing, $39 million. The rest is loans Amazon made to Anthropic that turn into shares later. Because they are loans, changes in their value don’t go into profit at all. They go into a separate account on the balance sheet. At the end of 2025, those loans were worth $45.8 billion. At the end of September, $146.2 billion. In this quarter alone, that account grew by about $36 billion after tax, most of it from those loans. None of it showed up in profit.

That is ordinary accounting, applied as the rules require. But it means the largest change in the value of Amazon's Anthropic stake this quarter is not in the profit figure.

### 3. Did anyone say how much of the backlog comes from the AI labs they fund?

**Pass, with a research miss**

I expected none of the three to give the share of its backlog that comes from Anthropic or OpenAI, in dollars or as a percentage. None did. But that is not the whole story.

In its quarterly report, right after its backlog figure ($578 billion at 30 September), Amazon names both labs. OpenAI’s existing $38 billion commitment was expanded by $100 billion over eight years. Anthropic’s was expanded by more than $100 billion over ten years. These are the sizes of the contracts, not the part of the backlog still to be paid, so they don’t meet the test I set. In the first piece, I treated Microsoft’s figures the same way: a share you can work out is not a share the company gives.

But the same two sentences were already in Amazon’s June report, and I should have found them. They belonged in the first piece as an example: Amazon names both labs and the size of their contracts, but not how much of its backlog comes from them. That would have made the prediction sharper, and more honest about what was already public. I checked Microsoft’s and Alphabet’s filings closely, and Amazon’s not closely enough.

Microsoft and Alphabet still give no share of their backlogs from the labs. Microsoft now reports what OpenAI paid it: $10.0 billion this quarter, about one dollar in nine of its revenue. That is a disclosure it did not make a year ago, and it deserves credit. But it is revenue, not backlog. Alphabet names one related company in its backlog: $16.8 billion from Crux AI, a joint venture it part-owns with Blackstone. More on that below. It names no AI lab.

What I am changing: a claim that something is *not* disclosed now needs a search of each company’s latest filings before it goes on record. Not having seen something is not the same as it not being there.

### 4. Did Microsoft's bills for later keep growing?

**Pass**

Yes. Leases Microsoft has signed but not yet started rose from $329.1 billion in June to $364.6 billion in September. The claim passes.

The pace slowed sharply. The total grew by $35.5 billion this quarter, against $132.5 billion the quarter before. Counting the leases that started during the quarter, Microsoft signed roughly $47 billion of new ones, about a third of the previous quarter. The leases now start between mid-2026 and 2034, a year later than in June, and the filing still says some depend on "certain contractual conditions being met".

For scale: the leases Microsoft already carries on its books come to about $98.5 billion. The ones signed but not started are about 3.7 times that.

On the call, an analyst asked how Microsoft would raise money for the build. Its finance chief, Amy Hood, answered: “There’s a form of financing called leasing, and we’ve been using that one as a vehicle.” Microsoft could also borrow, she said, but “we haven’t yet, I would say, tapped that.” The next day, Microsoft confirmed in a filing that it was selling bonds, its first US bond sale in more than nine years. Reuters, citing people familiar with the deal, put the size at about $25 billion.

### 5. Did Microsoft's spending look smaller while the building didn't slow?

**Pass**

Yes. The claim passes, and widely.

In the same quarter last year, leases counted as purchases (the kind that counts in the spending figure) made up 86.6% of the new leases Microsoft started using. This quarter, 44.3%. The total barely moved, from $10.6 billion to $11.7 billion. What changed is the split. At last year's split, roughly $5 billion more would have counted in the headline spending figure.

Microsoft reported $45.2 billion of spending this quarter, below the "over $50 billion" it said it expected in July. It gave its own reason: some deliveries of rented capacity it expected were delayed. I am not attributing the gap to the lease change. Both things happened in the same quarter.

There is a second route outside the headline figure. Microsoft's cash flow statement shows $12.9 billion under "Other" in its investing cash, against $6.2 billion a year ago, spent "primarily to facilitate the purchase of components". This is not new. It has been growing since last year. But it means the cash going into the build this quarter was up to about $52 billion, not $45 billion. I write "up to" because the filing does not say what the components are for, and it is not capital spending until it is counted as such.

Why did the split change? On the July call, Microsoft said that from 1 July it would count its data centres and offices as lasting 25 years instead of 15. Leases that cover most of a building’s life count as purchases; shorter ones count as rent. With buildings now expected to last longer, its finance chief said, “more of our future datacenter leases will shift from finance leases to operating leases.” But I could not find the change in the quarterly report: a search for “useful life”, “25 years” and “buildings” turns up only the standard wording. Companies have to disclose a change in an estimate like this when its effect is material, so its absence suggests Microsoft does not consider it material yet. The lease table does show what you would expect if the change were taking effect: the average remaining term on Microsoft’s ordinary leases went from six years to eight in one quarter, and those leases grew by 25% while the ones counted as purchases grew by 7%.

### Control: Did anyone change how long their servers last?

**Held**

No. Microsoft still says two to six years, Alphabet six, Amazon five to six. The check held.

### How did each company justify the spending?

The first piece named three ways, from strongest to weakest: a figure for what AI brings in today, a number about the future, or confidence with no figure attached.

**Microsoft** reported Azure growth of 48%. That is cloud revenue, not AI revenue. In April, Microsoft said its AI business had passed $37 billion in annual revenue. It has not repeated the figure since. Microsoft also stopped reporting how many people pay for Copilot. It gave the figure every quarter this year: 15 million paid seats in January, over 20 million in April, over 30 million in July, when it said new paid seats had more than doubled. This quarter it gave none. It reported that daily users rose by about half, a figure it had previously given alongside the paid count, not instead of it. And it said its E7 bundle, which includes Copilot, has “millions of paid seats”, without saying how many of those pay for Copilot itself. When an analyst asked for the number, the finance chief said Microsoft had not disclosed an update. GitHub followed a similar path: 4.7 million paying subscribers in January, then usage-based pricing from June, and a count of 50 million users in July.

**Alphabet** said Cloud had brought in “tens of billions of dollars in AI revenues” this year, and that it expects its return on the investment to stay “meaningfully above” its cost of capital. Neither came with a figure. Cloud revenue now includes sales of TPU chip systems to other companies. That is real revenue, but it is hardware sold once, not rented capacity.

**Amazon** leaned on the future. Andy Jassy said AWS would grow more than 40% in both 2027 and 2028, and that customers take new capacity almost as fast as it is installed. He said margins on AI were improving quickly, with no figure. His finance chief added that AI services currently earn lower margins than the rest of AWS. Jassy also called the $578 billion backlog "well diversified". The quarterly report filed the same week names two customers, OpenAI and Anthropic, with commitments of $100 billion or more each.

None of the three gave a current, specific figure for what AI brings in.

### How is the spending paid for?

**Alphabet** bought back no shares this quarter, as in the first half. It borrowed about $29 billion in new bonds during the quarter and sold $12.4 billion of new shares through its sale programme. It also paid its first dividend on the preferred shares it sold in June. Its borrowing now stands at $125.1 billion, against $46.5 billion at the end of 2025.

**Microsoft** brought in $18.6 billion of free cash flow and paid out about $11.1 billion in cash to shareholders in dividends and buybacks. It also repaid debt and spent the $12.9 billion on components. To cover the gap, it drew on its investments: it sold $10.5 billion and let $5.7 billion mature, while buying only $5.4 billion. A year earlier it bought $17.7 billion. Cash and short-term investments fell by $12.3 billion in the quarter, to $64.5 billion. That is ordinary money management at Microsoft’s size, but the direction is new. Its finance chief said free cash flow should be positive for the year to June 2027, and that Microsoft would look for other funding “when we believe it is attractive”. The next day, it launched its first US bond sale in more than nine years (see above). Microsoft also did not lean on its suppliers: what it owed for equipment already delivered was flat over the quarter, at about $26.5 billion.

**Amazon** bought back no shares. Its long-term borrowing rose from $65.6 billion at the end of 2025 to $173.1 billion at the end of September, including a $17.5 billion bank loan it drew in September. On Amazon's own definition, free cash flow over the twelve months to September comes to negative $42 billion. In June it was negative $7.6 billion. In this quarter alone, Amazon spent about $76 billion on equipment, against $42.4 billion brought in from the business.

### How much is built but not yet working?

At **Alphabet**, equipment and buildings not yet in use rose from $78.6 billion at the end of 2025 to $145.3 billion at the end of September, an 85% rise in nine months. What is in use grew by 28% over the same period. More of the money is going into things that do not earn anything yet.

Microsoft and Amazon do not break this out in their quarterly reports.

### Did anyone mention delays?

Microsoft did, in its opening statement, before anyone asked. It said it would make up for some rented GPU capacity delayed to the first half of 2027 by getting more out of the hardware it already has: finding unused capacity, and running its software faster. It gave no figure for how much these gains add. Its finance chief said hitting next quarter’s forecast depends on finding more of them: “We need to find some more, which we’re counting on.”

Alphabet said it is renting computing power from another company to fill the gap until its own data centres are ready. That capacity started only at the very end of September, so it barely affected margins in the quarter, but Alphabet expects it to weigh on margins in the last three months of the year, because renting costs more than running its own machines. It did not name the provider. But in June, Google described its deal with SpaceX in the same words: bridge capacity, to meet higher demand than it expected. Under that deal, Google pays SpaceX $920 million a month for computing power from October 2026 to June 2029. SpaceX is also the company whose falling share price drove Alphabet’s loss in claim 2.

Amazon said the opposite: capacity is used almost as soon as it comes online, and spending will be “meaningfully more” next year.

The first piece asked whether anyone would mention leases being delayed, renegotiated or cancelled. Nobody mentioned cancellations. Only Microsoft used the word delayed.

### Where did the bill go?

The predictions tested three ways a company could move the cost: paying suppliers later, getting customers to pay earlier, and renting instead of buying. The filings showed the cost moving in other ways too. Nothing here is hidden. It is all set out in the notes at the back of the filings. But none of it shows up in the two figures that usually tell you how much a company is spending: what it pays for equipment, and its free cash flow.

**Alphabet guarantees other companies' data-centre rent.** It has agreed to cover the lease and energy payments of certain data-centre operators if they cannot pay. The total it could owe under these guarantees rose from $16.9 billion at the end of 2025 to $48.0 billion at the end of September, with terms of up to 15 years. If an operator defaults, Alphabet can take over the lease. This is the same kind of promise Nvidia made in August for OpenAI’s data-centre campus in Ohio, up to $105 billion, from the other side of the table: the company that rents the capacity is helping the landlord get financed. 

**Alphabet's backlog includes a company it part-owns.** Crux AI is a joint venture Alphabet formed with Blackstone in May to sell data-centre capacity. Alphabet holds a minority stake, has promised up to $10.0 billion of guarantees for the venture's data centres and energy, and has signed a multi-year deal with it. That deal accounts for $16.8 billion of Alphabet's backlog. Alphabet discloses all of this because it part-owns the company and rules require that it does so. It is a small case of what claim 3 was about: promised revenue from a company the seller helped set up.

**Alphabet has a $20 billion promise to an unnamed company.** It has agreed to fund a private company with up to $20.0 billion through 2030, in stages, if the company meets certain operating and financial targets. The filing does not name it.

**A $6 billion lease kept off the balance sheet.** In June, Alphabet signed a lease lasting twelve months or less, and amended it in September. It commits about $6.0 billion that cannot be cancelled. Because the term is short, the rules say no asset or debt is recorded for it. The filing does not say what it is for.

**Amazon can borrow against the equipment it buys.** Amazon has arranged a loan facility of up to about $80 billion to buy technology equipment from suppliers, available from early 2027 to late 2028. Each loan runs five years, and Amazon will pledge the purchased equipment as security. It ties directly to the [Compute Watch](https://threadbaire.com/blog/posts/compute-watch.html) question: what the equipment is worth, and for how long, now matters to Amazon's lenders as well as to Amazon.

**Amazon lends to Anthropic as it delivers compute.** Amazon has a $20 billion facility for Anthropic, released in stages as AWS reaches delivery milestones for the computing capacity Anthropic has bought. Anthropic can then draw it as new notes or shares. Amazon has already invested $5 billion in Anthropic shares under the arrangement, which reduced what is left of the facility to $15 billion.

**Amazon has $165 billion of leases signed but not started.** 

Amazon’s leases signed but not yet started rose from $137.2 billion in June to $164.9 billion in September, after rising by about the same amount the quarter before.

None of this breaks a rule, and none of it means anyone is in trouble. But it answers the question the first piece asked. When the headline figures improve, it is worth checking whether the cost has gone, or moved into a guarantee, a pledge, a short lease, or a company the seller part-owns. 

This quarter, more of it sat in those places than before.

### Meta, checked after the fact

Meta reported on 21 October, before the first piece, so it was never part of the predictions. Checked afterwards against the same tests, it looks most like Alphabet. Its cash got worse, not better: by Meta’s own measure, free cash flow went from $784 million in the quarter to June to negative $2.1 billion, its first negative quarter since it went public in 2012. It spent $34.9 billion on equipment against $32.8 billion brought in from the business, and bought back no shares, against $3.3 billion a year earlier. The tests on investment income and backlog don’t fit Meta: it holds no large stakes in AI labs and rents out no computing power. Meta did not change how long its servers last. Five days later, it began the sale of up to $35 billion of new shares described in the first piece.

### So: income or promises?

Some of the improvement this week was real income. Amazon earned its highest operating margin ever. Azure grew 48%. Alphabet’s Cloud revenue doubled. These are sales, counted when the service is delivered, whether or not the customer has paid yet. And some of the biggest customers are funded by the companies selling to them. OpenAI paid Microsoft about one dollar in nine of its revenue this quarter, and Microsoft is one of its largest investors. In the same quarter, Amazon put $21.3 billion into OpenAI. And part of Alphabet’s Cloud growth was chips sold once, not services rented over time.

But the question was whether any improvement in cash was actually there, and the cash did not improve. Alphabet's got worse. Amazon's got much worse. Microsoft’s stayed positive, but it sold investments to cover what it paid out after that. Meanwhile the bills for later kept growing: in leases not yet started, in guarantees for other companies' rent, in loans secured on servers, in backlog from companies the sellers fund or part-own.

Four claims held and one could not be tested. One of the four held on its wording while I missed a disclosure I should have found, and I have changed the method because of it. The rest held, and the choices behind them pointed the same way: the headline figures looked better, and the cost was recorded further back in the filing.

------

### Primary sources

- [Income or Promises - Five Predictions for Microsoft, Alphabet and Amazon - 27 October 2026](https://threadbaire.com/blog/posts/income-or-promises.html)
- [Alphabet Form 10-Q for the quarter ended 30 September 2026](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000084/goog-20260930.htm)
- [Amazon Form 10-Q for the quarter ended 30 September 2026](https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000040/amzn-20260930.htm)
- [Amazon Form 10-Q for the quarter ended 30 June 2026](https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm)
- [Amazon Q3 2026 earnings release - 29 October 2026](https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000038/amzn-20260930xex991.htm)
- [Microsoft Form 10-Q for the quarter ended 30 September 2026](https://www.sec.gov/Archives/edgar/data/789019/000119312526436133/msft-20260930.htm)
- [Microsoft FY27 Q1 earnings release - 28 October 2026](https://www.microsoft.com/en-us/investor/earnings/fy-2027-q1/press-release-webcast)
- [Microsoft FY27 Q1 earnings call transcript - 28 October 2026](https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1)
- [Microsoft FY26 Q4 earnings call transcript - 29 July 2026](https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4)
- [Reuters, “Microsoft expects to raise $25 billion in US bond sale, sources say” - 29 October 2026](https://www.reuters.com/business/finance/microsoft-expects-raise-25-billion-us-bond-sale-source-says-2026-10-29/)
- [CNBC, “Google to pay SpaceX $920 million a month for compute capacity at xAI data centers” - 5 June 2026](https://www.cnbc.com/2026/06/05/google-to-pay-spacex-920-million-a-month-for-xai-compute-capacity.html)
- [Meta Q3 2026 earnings release - 21 October 2026](https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm)
- [Alphabet Q3 2026 earnings release](https://s206.q4cdn.com/479360582/files/doc_financials/2026/q3/2026q3-alphabet-earnings-release.pdf) 
- [Alphabet Q3 2026 earnings call transcript](https://abc.xyz/investor/events/event-details/2026/2026-Q3-Earnings-Call-2026-Ujou16CTw9/default.aspx)
- [Follow the Money Both Ways - Four Predictions for Anthropic's S-1](https://threadbaire.com/blog/posts/follow-the-money-both-ways.html)