Capex against cash flow, leases against bonds, and what sits in the footnotes
Three terms carry this page, so they are worth stating plainly. Capital spending is money laid out on data centres and chips. Operating cash flow is the cash the business itself throws off in a year. A purchase commitment is a contract to spend money later: it is not borrowing, it never appears as debt, and it is disclosed in a footnote. A company can therefore look almost debt-free and still have promised to spend more than it earns for years.
Why a European investor should care: almost none of the companies here are European, but the money is. European semiconductor, equipment and power businesses sell into this build-out, European pension and insurance money funds the credit behind it, and a large share of any European index tracker is invested in the five buyers. If the spending slows, it reaches a European portfolio through revenue, not headlines.
Capital spending as a share of the cash each business generates, on a rolling twelve-month basis. Below the line, the build is funded from operations. Above it, the money comes from somewhere else.
The charts above ask the question company by company. Asked of the whole US large-cap market, the answer is that the build is now big enough to move the aggregate. Free cash flow is the cash left after the spending, and these are the four builders against the other three of the seven largest US companies — Apple, Nvidia and Tesla, which are not building data centres.
The four builders' free cash flow peaked at $234bn and is $150bn in the twelve months to 30 June 2026, while the other three rose to $262bn and passed them. Measured against 381 large US companies outside financials, the four have gone from 18% of all the free cash flow those companies generate to 10%. Their operating cash flow did not fall; it rose. The capital spending is the whole of the difference.
The ratio at the top of this page compresses two moving numbers into one, and a ratio cannot say whether it moved because the cash came down or the spending went up. Drawn separately, the answer is the same at all five builders: operating cash flow kept rising, and capital spending rose faster. Apple, Nvidia and Tesla are here for contrast — large US companies that are not building data centres. The shaded area is what is left over, and at Oracle and Amazon there is none.
Two kinds of promise sit outside the balance sheet. A purchase commitment is a contract to buy something later: chips, cloud capacity, power, construction. A lease that has been signed but has not started is a data centre the company is bound to pay for from the day it is handed over, and until that day it appears in no lease liability, no debt figure and no leverage ratio. Both are contracted, both are multi-year, and neither is due now.
These five have promised $2.63tn that sits outside the balance sheet, against $771bn of debt and lease liabilities on it. That is 3.4 times as much off the balance sheet as on it, of which $1.10tn is data centres that have been signed for and not yet occupied.
| Company | Signed, not started | Purchase commitments | Total off | On balance sheet | Off vs on | Years of free cash flow |
|---|---|---|---|---|---|---|
| Alphabet | 91 | 811 | 902 | 121 | 7.5x | 12.3y |
| Meta | 279 | 349 | 628 | 112 | 5.6x | 13.6y |
| Microsoft | 329 | 229 | 558 | 129 | 4.3x | 8.3y |
| Oracle | 260 | 13 | 273 | 167 | 1.6x | none generated |
| Amazon | 137 | 130 | 267 | 242 | 1.1x | none generated |
The size of a promise says nothing on its own. What matters is whether the company can pay it out of the cash it already generates. Free cash flow here is operating cash flow minus capital spending, and the last column is how many years of it the whole commitment would absorb. On the three that still generate any, it runs from 8 to 14 years. Amazon and Oracle currently generate no free cash flow at all, because capital spending already exceeds the cash coming in, so anything they have promised has to be funded from borrowing or from a recovery in cash generation.
Microsoft is the case worth pausing on. Rank these five on debt and it comes last, at $40bn. Its lease liabilities are already 2.2 times its bonds, and what it has signed and not yet started is larger again than both together.
Everything above is a stock: how much is owed. None of it is a schedule. A company with $130bn of borrowings spread evenly over thirty years is in a different position from one with the same $130bn due inside three, and the balance sheet line reads identically in both cases. Refinancing risk lives entirely in the difference.
Below is what each company has contracted to pay in each of the next five years: bond principal, finance lease payments and operating lease payments together. This is cash leaving the business, not debt. Lease payments carry their own interest, so the total is larger than the borrowings and is not comparable to a debt figure.
The far more revealing cut is what share of the whole schedule falls inside those five years rather than beyond them. A company whose obligations sit mostly in the distant tail refinances at a time of its choosing. One with two thirds of them inside five years has to return to a market that may not be open on the day it needs to.
CoreWeave is the outlier and it is not close. Two thirds of everything it has committed to pay falls within five years, against roughly a third for Oracle and little more than a quarter for Meta. Meta and Oracle have borrowed heavily and borrowed long, which is the cheap way to be highly indebted. CoreWeave has borrowed heavily and borrowed short, against assets whose useful life is itself the open question of this cycle.
Outlier against what, though. Two thirds is alarming if a typical large borrower sits at a third and unremarkable if a typical large borrower sits at seventy per cent, and nothing above says which. So here is the same schedule for 103 US filers carrying at least $10bn of bonds, on the narrower measure most of them actually disclose: bonds alone, no leases.
CoreWeave is at 80% on that measure, the 98th percentile. The median large borrower has 38% of its bonds falling due inside five years; three quarters are under 51%. Every company paying for the build-out from its own earnings sits below that median, with Meta lowest at 17%, the 3rd percentile. Amber marks the ones above the median.
The size floor is doing real work here and is worth stating plainly. Include every borrower down to a billion and the distribution runs to a hard 100%, but every name up there is tiny: rolling one facility out of cash is not the problem a bond stack poses, and averaging the two flatters whoever is being compared. Above $10bn, only 2 of the 103 refinance more of their debt inside five years than CoreWeave does, and the larger of those two carries $17bn against CoreWeave's $22bn.
The top right is nearly empty, and that is where CoreWeave sits. Short ladders belong to small borrowers and long ladders to large ones, which is the diagonal the cloud makes. Being at one end of either axis is ordinary. Being at the short end of one while carrying the debt of the other end is not, and neither axis on its own can show it.
Bonds only, and a different measure from the chart above it, which also counts finance and operating leases: quoting a three-ladder figure against a one-ladder population would compare unlike things, so each company's bond-only share is recomputed here. Financials and real estate are excluded because a bank's liability schedule is not a corporate bond ladder. Of the filers examined, 86 disclosed an incomplete ladder and 61 none at all; both are dropped rather than treated as zero, since a missing year is an unread filing and zero-filling would push every one of them towards the short end.
| Company | Bonds | Finance leases | Operating leases | Next year | Within 5y | Beyond 5y | Share within 5y | As of |
|---|---|---|---|---|---|---|---|---|
| CoreWeave | 21.6 | 0.3 | 13.6 | 7.9 | 23.7 | 11.8 | 67% | 2025-12-31 |
| Amazon | 68.8 | 14.9 | 106.9 | 20.0 | 92.8 | 97.9 | 49% | 2025-12-31 |
| Microsoft | 46.1 | 89.7 | 24.7 | 22.5 | 64.4 | 96.1 | 40% | 2026-06-30 |
| Alphabet | 49.1 | 2.9 | 18.3 | 5.8 | 27.3 | 42.9 | 39% | 2025-12-31 |
| Oracle | 130.1 | 11.5 | 41.9 | 11.6 | 61.3 | 122.2 | 33% | 2026-05-31 |
| Meta | 59.0 | 1.5 | 33.5 | 3.6 | 26.1 | 67.9 | 28% | 2025-12-31 |
Years are counted from each company's own balance sheet date, not from a common calendar year. Microsoft's financial year ends in June and Alphabet's in December, so year two is a different window for each, and forcing both onto a calendar axis would invent precision the filings do not contain. The full five-year ladder is disclosed once a year, so where a company has filed since, the near rungs are more recent than the date shown.
Everything above is drawn from the companies' own filings. This is not. A credit default swap is insurance against a company failing to pay, and its price is what professional lenders charge each other to carry that risk. Every such trade is reported the next morning, so the market's own view is public.
Over the 20 most recent trading days there were 29,970 new single-name trades. Ranked by how many, with the cost of protection beside each:
Oracle is the most actively traded corporate credit on the list, and CoreWeave is next. The two names this page identifies as most stretched on their own accounts are also the two the credit market is busiest pricing. Oracle protection costs 138 to 195 basis points, against 94 for the broad investment-grade index, so lenders charge roughly three times the going rate to insure it. CoreWeave trades with no running spread quoted at all, which is how the market prices credit it no longer treats as ordinary investment-grade risk.
What this does not say. The tape reports trades, not sides, so it cannot show whether anyone is buying protection or selling it, and activity is not the same as alarm: a widely held bond is a widely hedged one. Notional is capped by rule on 26% of trades, meaning a size is reported as "at least" rather than exactly, so trade counts are used here and dollar amounts are not. 8% of trades name only the bond being referenced and not the company behind it, across 386 identifiers, and are excluded from the ranking rather than shown as codes; they cannot be pooled into a single unknown bar because each code is a different company. 20 days is a snapshot, not a trend.
A fund that wants leveraged exposure to an AI company often does not buy the shares. It writes a swap, and a dealer takes the other side. The dealer is then carrying the position, hedged or not, and is owed money if the fund is wrong. That relationship is invisible in every holdings figure, because the fund does not own the shares and the dealer does not want them.
Registered funds have to name the counterparty on every derivative they hold. Below is every such swap on an AI company that shows up in those filings: the dealer on the left, the company the swap is written on at the right, and the width of each ribbon is the notional exposure, not the current value of the contract.
The dealer side is narrower than the borrower side and it is not the names most people would guess. Nomura and Clear Street sit alongside Goldman Sachs at the top of it, and Clear Street is not a bank at all. Cantor Fitzgerald appears too. On the right, Meta is the largest single exposure, which is unremarkable for a company of its size. What is less expected is that CoreWeave and Nebius, both small and both entirely dependent on the build-out continuing, carry more swap notional between them than Microsoft and Amazon combined.
$4.76bn of notional across the swaps disclosed by registered funds. That is a slice, not a total: it sees mutual funds and ETFs, and it does not see hedge funds, insurers or pensions, which is where most swap exposure of this kind actually sits. Read it as evidence of who intermediates, not of how much. The same limit applies to ownership: sweeping these filings for the whole AI complex identifies about $142bn of holdings against roughly $17.9tn of market value, so under one per cent of who owns these companies is visible this way. That is why no ownership ranking appears anywhere in this series. 3 entries are excluded because the counterparty field holds the reference company rather than a dealer name.
The companies above fund most of this from their own earnings. The specialist data-centre operators renting capacity to them do not. CoreWeave's assets and liabilities have grown together, and the distance between the two lines is all the equity standing between its lenders and a loss.
Institutional managers above a size threshold have to publish their complete list of US shareholdings every quarter. Counting the distinct managers that report each name gives a measure of how widely owned it is, and it lines up with the rest of this page in an uncomfortable way.
Mostly this measures size, and it is worth saying so before reading anything into it. Rank these companies by how many managers hold them and rank them by market value and you get almost the same list: the rank correlation is 0.93. Large companies sit in more index funds and more mandates, so a bar chart of holders is close to a bar chart of market capitalisation wearing a different label. CoreWeave looks thinly held here and is not: its 943 managers place it 501st of the 23,939 securities institutions reported holding at all.
Which is why the useful cut is the change. A difference cancels the size effect entirely, because a company's market value is in both quarters. Here is the move from 2026-03-31 to the latest filing quarter.
Institutions rotated, and not towards the companies doing the spending. Micron gained 27% more holders and AMD 23%, while Meta and Microsoft lost them. The names picking up institutional breadth are the ones selling into the build-out rather than paying for it.
The control matters here more than the finding. The number of managers filing fell from 8,794 to 8,697 between the two quarters, so these gains are not an artefact of more institutions reporting. A rising tide would have lifted every name on the chart, and it did not.
This is breadth, not ownership, and it is not adjusted for size. These filings cannot say what proportion of a company institutions own. Added up as reported they exceed some issuers' entire market value, because a manager and its sub-adviser both report the same shares; filtered hard enough to remove that, they fall to a fraction of the true figure. What survives either way is the count of managers, since a firm reporting twice is still one holder. Whether any of these names is unusually held for its size is a different question this chart does not answer, and cannot until every name here has a market value attached to it. Options are excluded, because a contract on shares is not shares. Nothing here covers debt, retail investors, insiders, or institutions outside these reporting rules.
What this page does not say. Nothing here is a forecast, and none of it says the spending is unwise. Large purchase commitments at a company generating enormous cash may be entirely affordable, which the coverage column above tests directly, and the commitments should not be read as debt. The figures are drawn from company filings and are levels, not predictions. The free cash flow comparison uses the most recent full year for each company, which for two of them is depressed by the very spending the commitments will continue, so it is a snapshot rather than a capacity estimate.