How the AI compute boom is being financed, and why the plumbing looks circular

The spending is real, the revenue is mostly a promise

The companies building artificial intelligence data centres have a cash problem their revenue cannot yet solve. Dell'Oro Group estimated that data centre capital spending rose about 57 per cent in 2025 as AI deployments accelerated. Morgan Stanley has put global data centre spending at roughly 2.9 trillion dollars from 2025 to 2028, and reckoned about 1.5 trillion of that must come from outside the hyperscalers' own cash flow. That gap is why the financing is getting creative.

OpenAI shows how far commitments now outrun income. It has agreed to buy about 300 billion dollars of computing from Oracle over roughly five years from 2027, according to reporting on the contract, a figure that dwarfs its current revenue. Promises on that scale cannot come from sales alone, so the industry has turned to its own suppliers, to equity swaps, and to debt kept out of sight.

Chipmakers are turning into lenders to their own customers

The sharpest shift is that the firms selling AI chips now help pay for them. In September 2025 Nvidia said it would invest up to 100 billion dollars in OpenAI, staged as OpenAI commits to buy Nvidia systems. OpenAI pays Nvidia cash for chips; Nvidia takes a non-controlling stake in OpenAI. Two months later, in its November quarterly report, Nvidia told investors there was "no assurance" it would reach definitive agreements on the OpenAI opportunity, a reminder that the headline number is an intention, not a signed cheque.

AMD went further on structure. In October 2025 it granted OpenAI a warrant to buy up to 160 million AMD shares at one cent each, vesting as OpenAI takes 6 gigawatts of AMD chips and as AMD's share price climbs toward 600 dollars. Fully exercised, OpenAI could own close to 10 per cent of AMD.

Broadcom pushed the logic onto the balance sheet. According to Anthropic's initial public offering prospectus, reported by Reuters in October 2026, Broadcom will lend Anthropic up to 42 billion dollars through convertible notes to fund chip leases, about a third of the 125.2 billion dollars Anthropic has committed over five years for tensor processing unit capacity. Anthropic's own filing says Broadcom's double role, as hardware supplier and financier, creates "potential conflicts of interest". When the chipmaker lends you the money to lease the chipmaker's chips, the sale and the loan are the same event seen twice.

The debt that does not appear on the balance sheet

Where money is borrowed rather than swapped, much of it is routed around the borrower's accounts. In 2025 Meta financed its Hyperion campus in Louisiana through a special purpose vehicle rather than on its own books. Meta's disclosures and reporting on the deal describe a vehicle majority owned by the private credit firm Blue Owl, with Meta holding a minority stake, raising roughly 27 billion dollars of senior secured notes due 2049. The vehicle owns the site and leases it back to Meta, so Meta records a lease and an equity stake rather than tens of billions in debt. Meta also gave a residual value guarantee, worth up to about 28 billion dollars, that compensates investors if the asset loses value, a commitment disclosed in the footnotes.

This is a pattern, not a one-off. The Bank for International Settlements, in its March 2026 Quarterly Review, described hyperscalers increasingly using special purpose entities and joint ventures that raise private debt while the tech firm keeps a minority stake and signs long-term leases or capacity offtake agreements. The BIS called the result "shadow borrowing", obligations that behave like debt but sit largely outside corporate balance sheets. It also noted that hyperscaler bond issuance topped 100 billion dollars in 2025, most of it at maturities beyond five years, and that credit default swap spreads had risen for lower-rated issuers. The BIS flagged refinancing pressure at the vehicle level and the activation of guarantees as ways a problem could spread.

Why 'circular' is the word that keeps coming up

Put the pieces together and capital starts to move in loops. A chip designer invests in an AI lab, the lab uses the money to buy the designer's chips, and the sale flatters the designer's revenue, which supports its share price, which funds the next investment. UBS and other analysts have compared this to the vendor financing that inflated the telecoms boom around 2000, when equipment makers lent customers the cash to buy equipment and the revenue vanished when those customers failed.

The specific risks are concentration and reflexivity. A small group, Nvidia, AMD, Broadcom, Microsoft, Oracle, CoreWeave, OpenAI and Anthropic among them, increasingly finances and buys from one another, so trouble at one node travels fast. Morningstar's Brian Colello, commenting on Nvidia's stake in OpenAI, said that if an AI bubble were to pop, "this deal might be one of the early breadcrumbs". None of this means the demand is fake. It means a growing share of the reported demand is being underwritten by the sellers.

What to watch

The signal will not be one dramatic default. It will be smaller tells: credit default swap spreads on lower-rated hyperscalers, which the BIS noted had already risen; whether warrant and convertible structures turn into real cash sales or quietly lapse, as Nvidia's own "no assurance" wording allows; and how much of the roughly 1.5 trillion dollars of outside capital Morgan Stanley expects actually arrives on the terms now being marketed. The buildout is real and large. The question for markets is how much of it rests on the same few firms financing each other, and what happens to the loop if demand grows more slowly than the contracts assume.