Capital · Research note
The AI Build-Out Becomes a Credit Story
Incremental debt funded 9% of hyperscaler capex in 2024 and 32% by mid-2026. The risk has moved from shareholders who signed up for it to bondholders who did not.
For three years the artificial intelligence build-out was an equity story: four companies with enormous free cash flow chose to spend it on data centres, and shareholders decided whether that was a good use of the money. Nobody else carried the risk.
That has changed, and the change is measurable: incremental debt funded about 9% of hyperscaler capital expenditure in fiscal 2024 and roughly 32% on a trailing twelve month basis by mid-2026. This piece argues that the shift matters more than the spending level, because it moves the consequences of a wrong capacity forecast from investors who chose the exposure to creditors and insurers who were sold a different one.
The spending
Company guidance for 2026 sums to roughly $630 billion across the four largest spenders, against about $388 billion to $410 billion in 2025. Aggregate figures between $630 billion and $725 billion circulate in commentary because different sources use different definitions and some include Oracle. Figure 1 uses the narrower set, which is the one traceable to guidance.
The growth rate is the striking part rather than the level: Meta guided to $115 billion to $135 billion for 2026 against $72 billion in 2025, and Microsoft guided to $110 billion to $120 billion against about $90 billion. Increases of that size cannot be absorbed by a depreciation schedule set in prior years, and they cannot be funded indefinitely from operating cash flow.
The funding
Amazon, Alphabet, Meta and Oracle issued roughly $194 billion of bonds in 2026 through 7 July, against about $108 billion across the whole of 2025. Goldman Sachs expects issuance across five hyperscalers to reach approximately $250 billion this year and $400 billion in 2027.
Note that the two series in Figure 2 do not cover identical issuer sets, so the comparison indicates direction rather than a precise multiple; the direction itself is not in doubt.
Figure 3 is the central chart of this piece: a shift from 9% to 32% of capex funded by incremental debt in roughly two years is a change in the nature of the business rather than in its scale. Equity has also returned to the mix: Alphabet priced an $84.75 billion equity raise in June 2026. Hence the story is not that these companies became leveraged; it is that internally generated cash stopped covering the build-out, and both external markets are now being tapped to close the gap.
The demand side of this is already visible in credit markets. Hyperscaler issuance now accounts for something in the region of 16% to 23% of gross issuance across US investment grade, high yield and leveraged loans. J.P. Morgan estimates data centre construction requires around $1.5 trillion of investment-grade bonds over five years.
| Funding source | Position in 2024 | Position in mid-2026 |
|---|---|---|
| Operating cash flow | Covered essentially all capex | Covers roughly two thirds |
| Incremental debt | ~9% of capex | ~32% of capex |
| Equity issuance | Not used | Alphabet raised $84.75bn in June 2026 |
| Leasing and structured arrangements | Marginal | Growing, including private credit participation |
| Share of US IG, HY and loan gross issuance | Immaterial | Roughly 16% to 23% |
Sources: FactSet, Goldman Sachs, Morgan Stanley Research, J.P. Morgan, company filings.
The accounting question underneath it
Figure 4 sets the two lives against each other. Hyperscalers depreciate Nvidia-based data centre hardware over five to six years. Nvidia’s release cadence implies that the compute layer specifically has an economic life closer to two or three years, because a two-generation-old accelerator is not competitive for frontier training even when it still functions.
Be careful with this comparison, because it is frequently overstated: a data centre is not only GPUs, and buildings, substations, transformers and cooling plant have genuinely long lives while representing a large share of the capital cost. The mismatch applies to the fastest-depreciating component of a mixed asset base, which is why Figure 4 is presented as an estimate rather than a disclosure.
The mismatch still matters for a specific reason. If the accounts assume five years and the hardware is competitive for three, reported earnings in years one through three are overstated relative to the economics, and the correction arrives in years four and five as either an impairment or a replacement cycle funded by new capital. Depreciation and amortisation is already compounding at 30% to 40% a year at these companies, which is the mechanical consequence of the spending in Figure 1 arriving on the income statement.
There is a counterweight worth recording. Global AI revenue excluding China reached roughly $25 billion in the first quarter of 2026, against estimated depreciation of about $21 billion on data centre and chip investment, and revenue exceeded depreciation for the second consecutive quarter. The build-out is not yet obviously running ahead of the demand it serves.
Why the funding shift changes who bears the risk
An equity holder in these companies bought a claim on residual profit and accepted that management might spend badly; that is what equity is for.
A bond holder bought something different: investment grade credit is purchased by insurers against long-dated liabilities, by pension funds matching duration, and by index funds that hold whatever is issued. None of those buyers made a judgement about data centre utilisation in 2029. They bought a credit rating.
Furthermore, the tenor mismatch is the reverse of the one that usually causes trouble. These are typically long-dated bonds funding assets whose most valuable component turns over every few years, which means the debt outlives the equipment. Refinancing that structure requires continued revenue growth or continued market access. Market access for capital-intensive technology has historically been the first thing to close.
Markets have already shown they are watching: Alphabet’s results in late July triggered a sell-off on capex concerns, and Amazon, Meta and Microsoft faced visibly more sceptical investors in the same week.
Where this argument falls short
The strongest objection is that these are among the most creditworthy issuers in existence, with net cash positions, dominant market shares and revenue that is still growing faster than the depreciation charge. A 32% debt share of capex at a company with Alphabet’s balance sheet is not comparable to the same ratio anywhere else, and treating it as a warning sign risks confusing a change in financing preference with a change in credit quality. Issuing debt while it is cheaply available is ordinary corporate treasury behaviour, and reading it as distress is a category error.
Furthermore, the depreciation argument assumes older hardware becomes worthless rather than migrating to less demanding work. In practice inference, internal workloads and rental to smaller customers all extend useful life beyond the frontier training window, and the two-to-three year figure is the most aggressive reading available.
Finally, this piece has said nothing about whether the capacity is needed, which is the question that actually determines the outcome; if demand grows into the build-out, the funding mix is a footnote. The argument here is conditional: it says the consequences of being wrong now fall on a different group of people, not that the spending is wrong.
Conclusion
The AI build-out crossed from equity funding to credit funding somewhere in the last eighteen months, and the crossing was quiet because nothing broke while it happened. Roughly a third of capital expenditure now comes from incremental debt, issuance by these four companies is a meaningful share of the entire US corporate bond market, and J.P. Morgan’s estimate implies $1.5 trillion more to come.
That is the transfer worth watching: shareholders in these companies chose to fund an uncertain multi-year capacity bet and are compensated for it in the equity. Insurers and pension funds holding the paper made no such choice, and they are being paid an investment grade coupon for exposure to a capacity forecast, a depreciation assumption, and a chip cycle that turns over faster than the debt matures.