Markets · Credit · Featured Research

When AI Capital Spending Becomes a Credit Question

Capital spending across Alphabet, Amazon, Meta, Microsoft and Oracle absorbed 82 per cent of the cash their operations generated in the year to June 2026, against 40 per cent two years earlier, and net borrowing turned from minus 27 billion dollars to plus 195. The interest-bearing debt and lease obligations total 0.83 times a single year of that cash flow.

Issue date
Last revised
Data through
30 June 2026

In the four quarters to June 2026, Alphabet, Amazon, Meta, Microsoft and Oracle spent 566 billion dollars on property and equipment. Their operations generated 692 billion over the same period. Two years earlier the same pair of numbers was 164 and 411: investment then absorbed 40 per cent of the cash the businesses produced, and it now absorbs 82 per cent.

Over those same two years the five moved from repaying about 27 billion dollars of debt a year, on net, to raising about 195 billion. Together those changes mark a financing transition: capital spending is absorbing far more internally generated cash just as net debt issuance has accelerated.

Section 01The funding gap that can be observed

Plant spending ultimately draws on cash from operations, cash already held, asset-sale proceeds or external capital. But capital expenditure minus operating cash flow is not a closed financing identity: dividends, buybacks, acquisitions and other investing flows also use cash, while disposals and working-capital movements affect what is available. The measured gap therefore describes pressure on internal cash generation, not a traced source of funds. As capital expenditure approaches operating cash flow, less internally generated cash remains for other uses and funding choices become more relevant.

The transition can therefore be dated, which is a narrower exercise than judging whether the investment is justified. What follows measures the transition at the five US-listed companies whose capital expenditure is dominated by data-centre and compute infrastructure and whose accounts are on one comparable basis. Nvidia is excluded: it sells the equipment rather than installing it, so its capital expenditure is a different object. The private builders are excluded because no comparable primary filing exists to read.

The filings do not identify AI capital expenditure separately. A cash flow statement records cash paid for property and equipment without breaking it down by purpose, so every figure below is total capital expenditure: at Amazon that includes fulfilment and transportation, and at all five it includes ordinary buildings. Attributing the increase to AI infrastructure follows the companies' own characterisation and the Bank of England's.

Section 02What the filings show

Figure 1 is built from as-filed XBRL facts. Cash-flow items are tagged year to date, so a fourth quarter never appears on its own; single quarters are recovered by differencing consecutive year-to-date facts, and reported quarters are used wherever they exist.

Hyperscaler capital expenditure against operating cash flow, 2022 to 2026Upper panel: quarterly capital expenditure, drawn as bars, and quarterly net cash from operating activities, drawn as a line, summed across Alphabet, Amazon, Meta, Microsoft and Oracle from the first quarter of 2022 to the second quarter of 2026, in US dollar billions. Capital expenditure rises from about 38 billion a quarter to about 182 billion and closes almost all of the gap to operating cash flow. Lower panel: trailing four-quarter capital expenditure as a percentage of trailing four-quarter operating cash flow, one line per firm. Every line falls to a trough in late 2023 or early 2024 and then rises; at the latest reading Oracle is above 174 per cent, Amazon above 107 per cent and the other three are between 63 and 71 per cent.Capital spending against the cash flow that funds itAlphabet, Amazon, Meta, Microsoft, Oracle050100150200USD BILLIONS PER QUARTER20222023202420252026CALENDAR QUARTER IN WHICH THE REPORTED PERIOD ENDSQuarterly capital expenditure (bars) and operating cash flow (line)operating cash flowcapital expenditure0%100%200%CAPEX AS % OF OPERATING CASH FLOW2023202420252026TRAILING FOUR QUARTERSTrailing four-quarter capex as a share of operating cash flowOracleAmazonAlphabetMetaMicrosoftcapex equals operating cash flowMeasured from as-filed SEC XBRL facts. Nothing here is an estimate or a forecast.Oracle's fiscal quarters end one month before the calendar quarter they are placed in.
Hyperscaler capital expenditure against operating cash flow, 2022 to 2026Upper panel: quarterly capital expenditure, drawn as bars, and quarterly net cash from operating activities, drawn as a line, summed across Alphabet, Amazon, Meta, Microsoft and Oracle from the first quarter of 2022 to the second quarter of 2026, in US dollar billions. Capital expenditure rises from about 38 billion a quarter to about 182 billion and closes almost all of the gap to operating cash flow. Lower panel: trailing four-quarter capital expenditure as a percentage of trailing four-quarter operating cash flow, one line per firm. Every line falls to a trough in late 2023 or early 2024 and then rises; at the latest reading Oracle is above 174 per cent, Amazon above 107 per cent and the other three are between 63 and 71 per cent.Five hyperscalersUSD bn050100150200USD BILLIONS PER QUARTER20222023202420252026QUARTER ENDINGCapex (bars) vs operating cash flow (line)operating cash flowcapital expenditure0%100%200%CAPEX AS % OF OPERATING CASH FLOW2023202420252026TRAILING 4QTrailing 4Q capex / operating cash flowAlphabetAmazonMetaMicrosoftOraclecapex equals operating cash flowMeasured from as-filed SEC XBRL facts.Oracle's quarters end one month early.
Figure 1 · The gap that used to fund everything else Capital expenditure at these five firms absorbed 82 per cent of the cash their operations generated in the four quarters to June 2026, against 40 per cent at the trough a little over two years earlier, in the four quarters to March 2024. Operating cash flow did not fall over that period: it rose 68 per cent. Capital expenditure rose 246 per cent. The change in the ratio is the result of the numerator, and it is present at all five firms rather than at one. Source: U.S. Securities and Exchange Commission, XBRL company facts, as filed, for Alphabet (CIK 1652044), Amazon (1018724), Meta (1326801), Microsoft (789019) and Oracle (1341439). Notes: Capital expenditure is cash paid for property and equipment from the investing section of the cash flow statement, gross of disposal proceeds. Operating cash flow is net cash provided by operating activities as reported. Cash-flow items are tagged year to date, so single quarters are recovered by differencing consecutive year-to-date facts; reported quarters are used where they exist. Each period is placed in the calendar quarter its reporting period ends in, so Oracle's February, May, August and November quarters sit one month early. Neither measure captures data-centre capacity financed through leases, supplier arrangements or unconsolidated vehicles. Data through: Latest reported quarter, ending 30 June 2026 (31 May 2026 for Oracle).

The lower panel is the measurement. Capital expenditure as a share of operating cash flow fell steadily through 2023, reaching 40 per cent in the four quarters to March 2024, and has risen in every quarter since to 82 per cent. The trough matters for reading the size of the move: at the start of the sample, in the year to September 2022, the ratio was 54 per cent, so 2024 was a low point rather than a normal level. The move is therefore from roughly one-half to roughly four-fifths, off a base that had itself been falling.

The rise is not the denominator giving way. Trailing four-quarter operating cash flow rose from 411 to 692 billion dollars over the two years, a gain of about two thirds. Trailing four-quarter capital expenditure rose from 164 to 566 billion, which is roughly three and a half times. The ratio moved because the numerator moved.

Nor is it one firm. Four of the five had a ratio in the thirties in the year to March 2024, and Amazon was at 54 per cent; at the latest reading Microsoft is at 63 per cent, Meta 69, Alphabet 71, Amazon 107 and Oracle 174. Removing Oracle, which is much the most extreme, leaves the aggregate at 77 per cent against 40 at the trough, so the result survives dropping the observation that drives it hardest.

Measured against revenue rather than cash flow, capital expenditure has gone from 12 per cent to 31 per cent of sales. A group of companies reinvesting close to a third of revenue in physical plant is a different kind of business from one reinvesting an eighth, whatever the plant is for.

Section 03From surplus to issuance

The left panel of Figure 2 puts two financing indicators on one axis. The internal surplus, defined here as operating cash flow minus capital expenditure and nothing else, peaked at 247 billion dollars in the four quarters to March 2024 and is now 126 billion. Net borrowing, proceeds minus repayments, was minus 27 billion in the year to June 2024 and is now plus 195 billion. Gross proceeds over the last four quarters were 250 billion.

Internal funding surplus, net borrowing, and a capex-to-cash-flow scenarioLeft panel, measured: the trailing four-quarter internal surplus, defined as operating cash flow minus capital expenditure, falls from a peak of about 247 billion dollars in the four quarters to March 2024 to about 126 billion at the latest reading, while net borrowing, defined as debt proceeds minus repayments, rises from about minus 27 billion to about plus 195 billion, so the two lines cross. Right panel, scenario: cumulative capital expenditure minus operating cash flow over the following eight quarters, in US dollar billions, plotted against the assumed annual growth rate of capital expenditure from zero to ninety per cent, with one line for each of three assumed operating cash flow growth rates of zero, ten and twenty per cent. A marked point shows the requirement implied by the growth rates actually recorded over the last year.Internal surplus, external borrowing, and what a scenario needstrailing four quarters, USD bn0100200USD BILLIONSJun 2023Jun 2024Jun 2025Jun 2026TRAILING FOUR QUARTERS ENDINGMeasured: internal surplus and net borrowingoperating cash flow minus capexborrowing minus repayments-50005001000USD BILLIONS OVER EIGHT QUARTERS0%30%60%90%ANNUAL GROWTH IN CAPITAL EXPENDITURE+0%+10%+20%last twelve months:capex +81%, cash flow +35%Scenario: cumulative capex / cash-flow gaplines are annual growth in operating cash flowLeft panel measured. Right panel arithmetic under stated growth rates, not a forecast.No borrowing is attributed to a project; the gap is not traced to a funding source.
Internal funding surplus, net borrowing, and a capex-to-cash-flow scenarioLeft panel, measured: the trailing four-quarter internal surplus, defined as operating cash flow minus capital expenditure, falls from a peak of about 247 billion dollars in the four quarters to March 2024 to about 126 billion at the latest reading, while net borrowing, defined as debt proceeds minus repayments, rises from about minus 27 billion to about plus 195 billion, so the two lines cross. Right panel, scenario: cumulative capital expenditure minus operating cash flow over the following eight quarters, in US dollar billions, plotted against the assumed annual growth rate of capital expenditure from zero to ninety per cent, with one line for each of three assumed operating cash flow growth rates of zero, ten and twenty per cent. A marked point shows the requirement implied by the growth rates actually recorded over the last year.Financing capacityUSD bn0100200USD BILLIONS2023202420252026TRAILING FOUR QUARTERS TO JUNEMeasured (trailing 4Q)operating cash flow minus capexborrowing minus repayments-50005001000USD BILLIONS OVER EIGHT QUARTERS0%30%60%90%ANNUAL GROWTH IN CAPITAL EXPENDITURE+0%+10%+20%last twelve months:capex +81%, cash flow +35%Scenario: cumulative capex / cash-flow gaplines are annual growth in operating cash flowLeft: measured. Right: arithmetic, not a forecast.No borrowing is attributed to any project.
Figure 2 · Internal surplus, borrowing, and the gap under a scenario The left panel is measurement. The internal surplus these firms generated after paying for their own investment has fallen by about 122 billion dollars from its peak, and net borrowing has moved from repaying debt to raising it. The right panel compounds that measured base at stated growth rates and reports the arithmetic shortfall, showing how quickly the capex-to-cash-flow gap scales with the growth assumption. Source: Left panel: U.S. Securities and Exchange Commission, XBRL company facts, as filed. Right panel: the author's arithmetic on the measured base. Notes: Internal surplus is operating cash flow minus capital expenditure, before any payment to shareholders or lenders. Net borrowing is proceeds from borrowing minus repayments, taken from each registrant's own financing-section tags, which are not identical across the five: Alphabet reports proceeds net of issuance costs and includes commercial paper, Amazon separates short and long term, Microsoft reports debt maturing in more than three months, and Oracle reports senior notes. The aggregate is therefore comparable in sign and order of magnitude rather than to the dollar. The scenario grows the trailing four-quarter base at constant annual rates compounded quarterly and accumulates capital expenditure minus operating cash flow over eight quarters. Existing cash, asset sales and other cash uses are omitted. Reported operating cash flow already includes interest paid; no incremental financing cost is modelled. Data through: Latest reported quarter, ending 30 June 2026 (31 May 2026 for Oracle).

The crossing of those lines is the empirical result: internal surplus fell as net debt issuance rose. The filings do not identify the use of debt proceeds. Cash is fungible, and these firms also pay dividends, repurchase shares, buy other assets and hold large securities portfolios, so nothing in the panel attributes a dollar of borrowing to a data centre rather than to any of them.

Borrowing is less standardised across the five filings than capital expenditure or operating cash flow. Each registrant tags its financing section differently: Alphabet reports proceeds net of issuance costs and includes commercial paper, Amazon separates short-term from long-term, Microsoft reports debt maturing in more than three months, and Oracle reports senior notes. The aggregate is therefore reliable in sign and in order of magnitude rather than to the dollar.

Section 04The stock, and what it is not

Flows have turned; the stock has not. Adding disclosed borrowings to finance-lease liabilities gives 577 billion dollars of interest-bearing obligations across the five at 30 June 2026, 0.83 times a single year of their operating cash flow. That total aggregates each registrant's own balance-sheet classification: current and non-current long-term debt at four of the five, notes payable at Oracle, and each firm's disclosed finance-lease liability, Meta's carried from its latest annual tag. Gross of cash and before any coverage test, the multiple is a scale marker rather than a leverage measure, and at this level it is not the profile of a group already under credit strain.

The aggregate conceals a distribution, and the distribution is where the interest lies. Alphabet is at 0.55 times operating cash flow, Microsoft 0.58, Meta 0.65 and Amazon 0.90. Oracle is at 4.29, roughly five times the group figure, and it is also the firm whose capital expenditure most exceeds the cash its operations generate. The panel establishes that the two facts coincide, not why.

The Bank of England reached a similar reading in its July 2026 Financial Stability Report, judging that at the start of 2026 the stock of outstanding debt from AI companies was relatively modest, and that this had helped contain the immediate risk to financial stability, while noting that the pace of activity in credit markets during the first half of 2026 was causing potential risks to build rapidly.

funding capacity, per period and per firm:

    cash available for capex  =  operating cash flow
                              +  reduction in cash and securities
                              +  proceeds of asset sales
                              +  net external capital raised
                              -  other cash uses

    capex minus operating cash flow is an indicator of pressure
    on that capacity, not a traced funding identity

measured, five firms, trailing four quarters, USD bn:

                        to Mar 2024     to Jun 2026
    capital expenditure          164             566
    operating cash flow          411             692
    capex / cash flow           39.8%           81.8%
    internal surplus             247             126
    net borrowing                -18             195

    interest-bearing obligations at 30 Jun 2026:  577 bn
    as a multiple of one year of cash flow:       0.83x

Capital expenditure is cash paid for property and equipment from the investing section. Operating cash flow is net cash provided by operating activities, as reported, which includes a large non-cash add-back for stock-based compensation at all five.

One firm shows why cash capital expenditure is a floor rather than a total. Microsoft added 24.6 billion dollars of right-of-use assets in exchange for finance-lease liabilities in the year to June 2026, against 115.9 billion of cash capital expenditure. That is roughly a fifth again of infrastructure obtained with no investing outflow at all, because the lessor financed the building. It shows up on the balance sheet: 62 per cent of Microsoft's interest-bearing obligations are finance leases, and its lease liability of 66.6 billion is larger than its 40.3 billion of borrowings. A reader following only the capital-expenditure line would miss both the asset and the obligation.

A cash flow statement records the spending of a company. It does not record the spending done on its behalf.

The panel captures what sits on five balance sheets and no more. The Bank of England's report describes capital being raised for the same build-out through securitised data-centre structures, special purpose vehicles and other bespoke arrangements, and reports the OECD's estimate that the share of private credit financing AI investment rose from 9 per cent in 2024 to 34 per cent in 2025. It also records that AI issuers accounted for 41 per cent of non-refinancing US high-yield issuance in the year to the report, against 1 per cent of the high-yield index at the end of 2025. None of that flow appears in Figure 1 or Figure 2. The measurement is therefore a partial view of a financing transition whose least disclosed part sits outside these five consolidated balance sheets.

Section 05What the arithmetic then requires

The right panel of Figure 2 takes the measured trailing four-quarter base, grows capital expenditure and operating cash flow at constant annual rates, and accumulates capital expenditure minus operating cash flow over eight quarters. The panel is a sensitivity map, not a forecast.

Over the last year capital expenditure at these five grew 81 per cent and operating cash flow 35 per cent. Repeating both rates for two years produces a cumulative gap of about 378 billion dollars between the two series. That gap could be met from existing cash, asset sales or external funding; shareholder distributions and other cash uses could enlarge it. Hold cash-flow growth at 10 per cent instead, and the gap becomes positive once capital expenditure growth rises above about 30 per cent. Because it is the difference between two large and separately growing quantities, assumptions that sound like small revisions move it by hundreds of billions.

The other variable the scenario holds fixed is time. Debt raised to build a data centre has a stated maturity; the equipment inside has a useful life set by the pace of hardware improvement, and the Bank of England's report notes mixed evidence on how quickly AI chips depreciate, with current shortages arguing for longer lives and rapid efficiency gains for shorter ones. Whether that becomes a maturity mismatch depends on the terms of individual debt agreements and on how the risk is allocated between lender, borrower and tenant, none of which is visible from consolidated filings.

Two things would change the reading, and both are observable in the same quarterly data. If capital expenditure growth decelerates towards the growth rate of operating cash flow, the ratio stabilises wherever it happens to be, and the financing question recedes without any debt being repaid. If cash generation disappoints while spending is already committed, the capex-to-cash-flow gap widens and management must choose among cash reserves, asset sales, reduced other uses or outside capital. What the last two years establish is that the group has used up the margin which made that choice unimportant.

  • Capital expenditure here is total; no cash flow statement separates AI infrastructure from other property, and the attribution to AI follows the companies' own descriptions and the Bank of England's.
  • Operating cash flow is as reported and includes a large non-cash add-back for stock-based compensation. A reader who treats equity issued to employees as a real cost should read the ratio as flattering.
  • The comparison base is the trough, the most favourable base for showing a rise; Section 02 gives the value at the start of the sample.
  • The five financing-section tags differ, so the borrowing aggregate is comparable in sign and order of magnitude rather than to the dollar, and none of it is attributed to a project.
  • Only obligations on these five balance sheets are captured. The securitised structures, special purpose vehicles, supplier financing and private credit in the Bank of England's report are not, and that omission runs one way: including them would raise the measured financing.
  • Oracle's fiscal quarters end a month before the calendar quarters they are placed in, and Meta's finance-lease liability is an annual tag carried into the June 2026 column.
  • The scenario reports capital expenditure minus operating cash flow under stated growth rates, not an external-funding requirement, and models no incremental financing cost.

This research is analysis and commentary for general information. It is not investment advice, an offer, or a solicitation, and it contains no price forecasts and no view on any security. Figures 1 and 2 are the author's own calculations from the primary filings cited; the right-hand panel of Figure 2 is a scenario under stated assumptions.

References & notes

  1. U.S. Securities and Exchange Commission. XBRL company facts API. sec.gov. Every figure in Figures 1 and 2 is an as-filed fact drawn from this endpoint for Alphabet (CIK 1652044), Amazon (1018724), Meta (1326801), Microsoft (789019) and Oracle (1341439), and therefore from those registrants' own Forms 10-Q and 10-K.
  2. Bank of England (7 July 2026). Financial Stability Report, July 2026. bankofengland.co.uk. Source for the Financial Policy Committee's judgement on the stock of AI-related debt, for the description of securitised, special-purpose and private-credit financing of the build-out, for the OECD estimate quoted in Section 04, for the high-yield issuance share, and for the observation on chip depreciation in Section 05. Published after this issue's July date and incorporated in the September 2026 revision.
  3. Organisation for Economic Co-operation and Development, as reported in the Bank of England's July 2026 Financial Stability Report: the share of private credit financing AI investment rose from 9 per cent in 2024 to 34 per cent in 2025. Quoted at second hand and labelled as such, because the underlying series is not published in a form this journal can reproduce.
  4. Financial Accounting Standards Board. ASC 842, Leases. The basis on which a finance lease adds a right-of-use asset and a liability without an investing cash outflow, which is the mechanism discussed in Section 04. The amounts used here are the registrants' own tagged disclosures of right-of-use assets obtained in exchange for finance lease liabilities.
  5. The quarter ending 30 June 2026 was reported in filings made after this issue's July date. The issue was revised in September 2026 to use them, and the publication record above states the data cut-off separately from the issue date for that reason.
  6. The panel construction, the year-to-date differencing, the scenario and the derived series are reproduced by the script in research/2026-07/ in the journal's repository, which also writes the results file that supplies every number quoted above.

Return to the front page