Industry Odisha Bureau, Oct 2: Data centre spending is racing ahead, and borrowed money is filling the gap. Credit rests on cash flow rising from $600 billion to $2 trillion by 2030. Whether customers deliver remains unclear.
Much of the AI boom is financed against cash not yet earned. The revenue to pay for it remains a forecast.
The $2 Trillion Test
Torsten Sløk, chief economist at Apollo Global Management, has highlighted the central assumption. Consensus forecasts see hyperscaler operating cash flow rising from $600 billion in 2025. By 2030, it must reach $2 trillion. That requires roughly 27% growth every year for five years. Operating cash flow comes from customers, not from boardroom decisions.
Spending Outruns Cash
On the same forecasts, hyperscaler capital expenditure keeps climbing through the decade. Combined free cash flow, the cash left after investment, turns negative this year and next. It recovers only from 2028. ECB staff project total hyperscaler capital spending above $1 trillion by 2028. Negative free cash flow does not mean distress.
Even Alphabet Borrows
Alphabet generated about $39 billion of cash in the April-June quarter. It spent nearly $45 billion on data centres and equipment. Quarterly free cash flow turned negative. Alphabet raised more than $50 billion through bonds in the first half. In February, it also sold a 100-year sterling bond. Lenders are committing money for a century to a fast-changing industry.
Lenders Widen the Circle
Apollo, with Blackstone, led a $35 billion financing for Broadcom chips running Anthropic models. Apollo called it the largest private financing ever executed. ECB staff say US big tech sells nearly one in 10 new non-financial euro bonds. Amazon’s euro deal this year was the market’s largest ever. Yet demand for hyperscaler euro bonds weakened from mid-year. Investors now demand a higher premium across maturities, but have not walked away.
Debt Off the Books
Some financing never appears on corporate balance sheets. Meta’s Hyperion campus in Louisiana is worth roughly $27 billion. A joint venture, mostly owned by Blue Owl Capital funds, pays for it. Meta holds a minority stake and will rent the campus. It has guaranteed part of its value for over a decade. Yet the venture’s debt sits off Meta’s balance sheet. Such structures are lawful and common. The Bank for International Settlements described them as shadow borrowing: debt-like obligations kept largely off-balance-sheet. The concern is visibility, not legality.
One Loop, Many Links
An AI lab raises money from a chipmaker or cloud firm, then buys its computing. One firm’s spending becomes another’s sales.
Customers Decide
Everything depends on end customers paying. Bain & Company has estimated an $800 billion yearly revenue shortfall for computing by 2030. Montgomery Investment Management says high-margin AI software sales must grow over tenfold in five years. It expects cash flow to fall well short of $2 trillion. ECB staff warn ratings may rest on uncertain revenue and leverage assumptions, raising mispricing risk.
Why 2008 Has Limits
Some observers compare this to 2008, when risk hid outside bank balance sheets. Pension funds and insurers, seeking long-dated quality assets, are big buyers of hyperscaler euro bonds. But 2008’s weakest borrowers were low-income households. Today’s central borrowers are hugely profitable, and most hold top-tier credit ratings.
Infrastructure Meets Its Customers
Borrowed money can build data centres. Only customers can repay it. Sløk warns a shortfall could mean costlier credit, cut spending and slower US growth. That cash may yet arrive. For now, the AI boom’s lenders are financing a forecast.
