AI spending strains Big Tech balance sheets, Moody's warns

Moody’s Ratings has put a price on the AI buildout that Amazon, Meta, Alphabet and Microsoft have spent the past year calling strategic necessity. In Moody’s reading, that necessity is not just inflating capital expenditure; it is changing what the cloud business is. Hyperscalers that once looked like asset-light software machines are starting to behave more like infrastructure owners, with fatter lease books, more borrowing and less room for revenue to arrive late.

On the analyst’s screen, the numbers are stark. Moody’s estimates hyperscaler capital expenditure will reach $US785 billion, about $1.2 trillion, in 2026 and climb to roughly $US1 trillion, about $1.5 trillion, in 2027, according to its research note and figures reported by CNBC. Direct debt across the six companies Moody’s tracks has already risen to $US460 billion, about $704 billion, while lease commitments sit near $US1.2 trillion, about $1.8 trillion. The debate, then, is no longer whether AI requires huge spending. It is which pocket pays first, and for how long.

What nags from the sceptic’s side is that even those large figures may still understate the obligation. Moody’s says about $US820 billion, roughly $1.25 trillion, of lease commitments have not yet commenced, meaning a sizeable share of the exposure is still rolling towards reported debt rather than sitting cleanly inside it. A business can look composed on paper while the bills are still walking up the drive.

Moody’s itself framed the shift bluntly.

“Previously, these companies relied on asset-light structures centered on software, intellectual property, and scalable cloud services that required modest capital investment.”
Source: Moody’s, via CNBC

Cash flow is where the strain shows first

Optimists, fairly, can still point to scale. The biggest platforms are not distressed borrowers, and Moody’s does not present them that way. Its argument is narrower and more interesting: balance-sheet headroom still exists, but the return timetable now matters more because AI compute, cloud tooling and model training are consuming capital at a pace that software-era investors were not trained to treat as normal. This is not yet a downgrade story. It is a cash-flow story.

Already, that pressure is visible in the line investors punish fastest. FactSet’s analysis argues that AI capex is now outrunning internally generated cash for several hyperscalers, while BBC reporting on Alphabet showed the company’s latest quarter tipping into negative free cash flow as infrastructure spending climbed. Buybacks have to compete with racks. So do dividends, land banks, power contracts and the next wave of data-centre commitments.

Set beside the old cloud narrative, that is a structural break. For years, scale in enterprise tech meant software margins widening faster than fixed costs. AI reverses that logic, at least for now: the more serious a platform becomes about winning, the more it starts to look like a builder of expensive physical networks, only with a Silicon Valley valuation attached. Flexibility goes first. Solvency, if it ever comes under pressure, is later.

The hidden leverage is sitting in leases

Look past the bond tally and the financing architecture becomes harder to ignore. The Bank for International Settlements describes an AI funding chain built not just on traditional debt, but on special-purpose vehicles, private-credit structures and long-term data-centre leases that secure capacity without always placing the full economic burden in one simple line item on day one. That is the sceptic’s answer to the question of where the real obligations sit: not nowhere, just somewhere messier.

Here the BIS is blunter than most equity research.

“These arrangements amount to "shadow borrowing".”
Source: Bank for International Settlements, Quarterly Review

From a policy angle, that phrase broadens the story. If AI demand disappoints, or if pricing fails to catch up with capex, the risk does not stop with one issuer’s bond spread. It can travel through landlords, structured vehicles, private-credit funds and the banks writing backup lines. The Financial Stability Board has already been mapping those non-bank channels more broadly, and Data Center Dynamics has reported on investor concerns that current lease disclosure still leaves too much hidden. The regulator-policy concern, then, is not that Big Tech is on the brink. It is that the web around Big Tech is growing faster than transparency.

That also gives a partial answer to one of the sceptic’s core questions. How much of the AI bill is still parked off to the side rather than fully reflected in reported debt? At minimum, Moody’s says $US820 billion, about $1.25 trillion, of lease commitments have yet to start. Not trivial. Not secondary. On some readings, it may be the most important financing number in the whole buildout.

Oracle is the stress test peers would rather avoid

Then there is Oracle, the outlier that makes the abstract version of this story easier to see. Moody’s treats it as a lower-rated exception among the larger names, and FactSet notes that S&P has already downgraded the company as capex and cash-flow pressure mounted. Oracle is not a clean proxy for Amazon Web Services or Microsoft’s broader cloud machine, but it does show what happens when AI infrastructure ambition meets a balance sheet with less slack and a customer mix that offers fewer ways to absorb mistakes.

Seen that way, Moody’s warning is comparative rather than apocalyptic. Cash-rich groups such as Alphabet and Meta may be able to spend through a long monetisation gap because they can issue debt more cheaply, sell equity from a position of strength and sign enormous leases without immediately losing market trust. Others will not get the same patience. The Financial Times reported this week that Morgan Stanley has become Wall Street’s leading arranger of AI debt deals, which is another way of saying capital markets are already sorting the strong from the merely enthusiastic.

For enterprise tech, that is the part worth watching. The AI race is turning into an infrastructure race, and infrastructure races are decided as much by financing tolerance as by model quality or launch cadence. If hyperscalers keep evolving from asset-light software platforms into asset-heavy builders, enterprise customers should expect capacity discipline, pricing discipline and return discipline to matter more than the industry’s product rhetoric suggests. Moody’s is not calling time on the boom. It is marking the moment when the boom stopped being mainly a story about software and started becoming a story about balance sheets.