What You Have to Believe for the AI Build-Out to Pay

Across the four largest US spenders (Amazon, Alphabet, Meta and Microsoft) the quarter's bill came to $201.9B, against $175.5B of operating cash flow. That only works if five specific things turn out to be true. This piece names them, shows what the filings say about each, and gives you a way to tell which way it's going.

Every figure here comes from a company filing, a company's own results release, or an official body such as the ILO or the IMF, and links to its source. Where we couldn't find a primary source, the number is out. The biggest gap is OpenAI. It hasn't filed, so what we can say about it comes from what its partners disclose.

Three of the four spent more on property and equipment in Q3 2026 than their operations produced. Amazon bought $77.3B against $42.4B of operating cash flow, 1.82 times. Alphabet bought $51.4B against $42.3B (1.21 times) and Meta $33.9B against $32.8B (1.04 times). Microsoft is the exception: $39.4B against $58.0B, or 0.68 times, in its fiscal first quarter (July to September). The figures come from Amazon, Alphabet, Meta and Microsoft.

These are gross purchases and leave out assets acquired under finance leases, so the real bill is larger at Microsoft and Meta. Amazon's own free cash flow definition gives negative $42.4B over the twelve months to September.

The cash side of the bill also trails the orders. Amazon's 10-Q shows property and equipment acquired but not yet paid for rose $21.5B in nine months, and $159.6B of further purchase commitments sit off its balance sheet.

How much of it is switched on? The filings show a large construction pipeline. Alphabet carried $145.3B of assets not yet in service at 30 September, up from $78.6B in December (10-Q). Meta's construction in progress, mostly data centres, servers and network gear, was $89.5B against $50.5B (10-Q). Amazon doesn't split it, but its AWS property and equipment grew from $190.1B to $313.0B in nine months, and it expects cash capital expenditure to "meaningfully increase in 2027". None of the filings says how many installed chips sit unused, so the claim that hardware is waiting in warehouses can't be tested from them. Alphabet does say its cloud revenue has been affected by "changes in customer usage, demand, and supply availability".

The commitments behind the spend are larger again. Alphabet reports $802.1B of contractual commitments, and Meta $454.5B of non-cancelable commitments, mostly for third-party cloud capacity. Amazon's total is $883.6B. The companies define these differently, so they shouldn't be added.

This belief fails if spending stays above operating cash flow for several more quarters. One quarter doesn't make a trend. Q4 results are the next test.

When spending outruns cash, someone else's money fills the gap. Amazon's 10-Q shows how it did it. It raised $48.1B of long-term debt in Q3 and $115.1B in the first nine months of 2026, and its cash fell from $86.8B at December to $55.8B at 30 September. Debt at face value went from $68.8B to $180.9B. That includes a $37.0B bond in March, a $25.0B bond in July and a $17.5B term loan drawn in September, plus euro, franc, Canadian dollar and sterling issues. A supplier financing facility adds up to about $80B for technology infrastructure. Amazon made no share repurchases in the period.

The others are doing the same. Meta issued $25.0B of senior notes in May, and its long-term debt rose from $58.7B to $80.9B (10-Q). Alphabet issued senior notes with $28.7B of net proceeds in Q3 and runs a $40B at-the-market equity programme that began on 1 June. It had sold $12.4B by 30 September, with $27.6B left, and the filing says the proceeds are primarily to meet tax obligations on employee equity grants (10-Q). Microsoft priced a $25B senior notes offering on 29 October, closing 4 November.

The smaller builders are more stretched. Oracle's 10-Q for the quarter to 31 August (its fiscal Q1 2027) shows capital expenditure of $28.5B against operating cash flow of $23.1B, free cash flow of negative $5.4B, and $664B of remaining performance obligations. It used up a $20B at-the-market equity programme, raising $19.9B net. CoreWeave, which borrows to buy chips and rent them out, had $35.6B of debt at 30 June (10-Q). Effective rates run from 7% to 15% across its facilities and about 10% on its senior notes, and interest expense on debt was $558M in the quarter against $256M a year earlier.

The cost shows up in depreciation with a lag. Meta's depreciation on servers and network assets was $5.34B in Q3 against $3.60B a year earlier, and Alphabet's nine-month depreciation of property and equipment rose from $15.1B to $22.0B. How many years a server is assumed to last is each company's own choice, and a longer life spreads the cost over more years. Alphabet depreciates servers and network equipment over six years. Microsoft's policy is two to six years. Meta extended most servers and network assets to 5.5 years from 1 January 2025, which cut its 2025 depreciation by $2.92B and raised net income by $2.59B. Amazon moved servers from five to six years in 2024, then cut a subset of servers and networking equipment back to five in 2025, adding $1.4B of depreciation and lowering net income by $1.0B.

This belief fails if debt and equity get harder or dearer to raise. For you, the point is that the vendors you rely on are now financed more like utilities, and their pricing power depends on the capital markets staying open.

For the first time one lab's numbers can be read from a filing. Anthropic filed a preliminary S-1 on 26 October 2026, with some figures still blank.

The growth is steep. Revenue was $4.6B in 2025 and $16.3B in the first half of 2026. By quarter it went from $2.0B in Q4 2025 to $4.7B in Q1 and $11.6B in Q2.

The commitments are larger still. Anthropic has agreed to buy at least $110B of compute from Amazon (to 2036), about $114.5B from Google (to 2033) and $44.9B from Microsoft (to 2033), whether or not it uses them. That is at least $269B across the three clouds. A separate five-year lease with Broadcom adds rent of about $125.2B, non-cancelable except on default, which takes the named total to at least $394B. The cloud commitments run for seven to ten years, so setting them against half a year of revenue isn't like for like. It does show how much growth is already assumed.

Two reported lines show the cost side. Compute and infrastructure expense was $7.3B in 2025 against $4.6B of revenue, and $8.9B in the first half of 2026 against $16.3B. The filing doesn't report a gross margin.

On who buys, the S-1 counts more than 11,900 customers paying over $100,000 a year and about 250 paying over $10 million. It also says a substantial portion of revenue and receivables comes from a limited number of customers, and gives no percentages.

The cash is better than the headline loss suggests. Operating cash flow was positive $3.9B in the first half, against negative $1.5B a year earlier, and cash and short-term investments were $131.7B at 30 June. Operating loss for the half was $0.7B, yet net loss was $80.7B. The difference is a $72.4B loss on the change in fair value of convertible notes, derivatives and SAFEs, which the filing says was primarily driven by a higher fair value of its convertible notes. It's a revaluation of what investors hold, and no cash goes out.

About 48% of first-half revenue was sold through cloud-provider marketplaces, and some of those providers also hold Anthropic's convertible notes and preferred stock. The S-1 says Anthropic contracts directly with those customers and pays each cloud a fee, so its revenue may not be measured the way another lab's is. Compare revenue across labs with care.

Cloud growth is the other side of those commitments. Google Cloud revenue rose from $17.7B in Q4 2025 (8-K) to $20.0B in Q1 (8-K), $24.8B in Q2 (8-K) and $30.7B in Q3 (8-K).

AWS sales rose 39% to $46.0B in Q3 2026 from $33.0B a year earlier, which Amazon's 10-Q attributes primarily to increased customer usage, partly offset by pricing changes from long-term customer contracts.

This belief fails if cloud growth slows while the commitments stay fixed. OpenAI's equivalents aren't public. Microsoft's annual report says it is a major investor in OpenAI and that net gains from those investments added $5.0B to its fiscal 2026 net income.

The same dollars can be counted more than once. A cloud provider invests in a lab, the lab buys compute from the cloud provider, and the cloud provider buys chips. Each step shows up as revenue somewhere.

Nvidia's revenue was $215.9B in fiscal 2026 (February 2025 to January 2026). Its 10-K says one direct customer was 22% of revenue and another 14%, and that some indirect customers are individually 10% or more. It names none of them. The hyperscalers' spending in Belief 1 is the kind that ends up as chip revenue, but the link remains an inference.

Amazon's 10-Q puts numbers on the loop. It invested $8.0B in Anthropic convertible notes between 2023 and 2025, $5.0B in Anthropic Series G preferred in Q2 2026, and $50.0B in OpenAI Series C preferred in 2026 ($15.0B, $13.7B and $21.3B across Q1 to Q3). In the same year it expanded its AWS commitments with OpenAI by $100.0B over eight years and with Anthropic by more than $100.0B over ten.

The notes it holds in Anthropic were carried at a fair value of $146.2B at 30 September, against $45.8B at December. Amazon booked $62.9B of upward adjustments on its Anthropic preferred stock in other income in the nine months. That gain isn't cash. Anthropic's $72.4B loss, in Belief 3, is the same revaluation seen from the other side.

Google sits in a similar position. Anthropic's S-1 shows up to $20B of Google notes available and an option on up to $10B of shares, against about $114.5B of compute commitments. Amazon had $6.2B of principal outstanding on Anthropic notes at 30 June (its own 10-Q says it has invested $8.0B in notes in total, part since converted to preferred stock), with up to $15B more available. The cloud providers are suppliers and financiers to the same labs.

OpenAI has no filing to read, but one of its biggest backers does publish. SoftBank says its total investment in OpenAI reached $64.6B on 1 October, for an ownership interest of about 13%. It paid for the final $10B tranche with $11.1B of senior notes priced on 24 September, rated BB+, with US dollar coupons of 8.625% to 9.75%. The stake is funded with debt rated below investment grade.

You can test this yourself. Follow one dollar from a business paying for an AI subscription back to the chip, and count how many balance sheets it touches. This belief fails if end-customer spending lags the spending between vendors.

Revenue forecasts only make sense if AI captures a meaningful part of labour income. No official body publishes a forecast of AI revenue as a share of that pool, so we don't offer one. What the primary sources do give is the size of the pool and how exposed it is.

The ILO reports that the global labour income share fell 0.6 percentage points from 2019 to 2022 and has been flat since. Had it stayed at its 2004 level, labour income would have been $2.4 trillion larger in 2024 alone (ILO).

The World Economic Forum's Future of Jobs Report 2025 surveys more than 1,000 employers covering 14 million workers in 55 economies. They expect 170 million jobs created (14% of today's employment) and 92 million displaced (8%) by 2030, a net 78 million.

The IMF's staff discussion note says almost 40 percent of global employment is exposed to AI. The figure is about 60 percent in advanced economies, 40 percent in emerging markets and 26 percent in low-income countries. In advanced economies about half of high-exposure jobs may be negatively affected.

This belief fails if labour's income share doesn't move. For you, it means the pace at which vendors need you to pay more is set by this assumption, whatever your own usage.

What the price of your AI vendors rests on is now visible, and that's enough to act on. Whether AI is a bubble can stay an open question.

1. 01

   Don't assume today's AI prices hold. Avoid contracts that lock them in for five years.

2. 02

   Ask each vendor which of the five beliefs its own plan depends on.

3. 03

   Keep switching costs low. Prefer portable workflows over deep lock-in to a single lab.

4. 04

   Watch four numbers each quarter: capex against operating cash flow, new debt and equity issuance, lab revenue disclosures, and labour income share.

5. 05

   Spend where the benefit shows up in your own numbers within a year, whatever happens to the market.

All figures were read from the primary documents linked in the text, as at 2 November 2026. Gross purchases of property and equipment exclude finance leases. Microsoft's quarter is its fiscal Q1 FY27 (July to September 2026). Anthropic's S-1 is preliminary and will change. OpenAI has not filed, so its figures are out.