AI circular financing: when the same dollar goes round and comes back as revenue

A question comes back every quarter: is the demand carrying AI valuations real, or partly manufactured by the players themselves? The suspicion has a name, circular financing. A chipmaker invests in an AI lab, which commits to leasing compute from cloud providers, which buy the maker’s chips. The same dollar goes round and comes back as revenue, which can make demand look organic. Without settling the bubble debate, here is the loop laid flat, quantified, and the argument of those who consider it healthy.

The core of the arrangement reads in three moves. Nvidia has committed to invest up to $100 billion in OpenAI as part of a capacity-deployment partnership. OpenAI, for its part, has piled up colossal compute commitments with cloud providers: around $300 billion over five years with Oracle, $250 billion with Microsoft, $22.4 billion with CoreWeave, $38 billion with Amazon Web Services. Yet these providers fit out their data centers with Nvidia chips. The capital injected by the maker at the top of the chain therefore comes back to it as orders at the bottom.

The loop, in plain terms

The clearest case is CoreWeave. Nvidia owns more than 5% of its equity, and agreed in September 2025 to buy $6.3 billion of cloud services from it, committing to pay for the compute time CoreWeave failed to sell to others. With that backstop, CoreWeave can order more Nvidia chips with confidence. The same pattern recurs, to varying degrees, between the maker, the lab and the compute landlords. None of these deals is illegal or abnormal in itself, vendor financing has existed for a long time, but their accumulation among a very small number of interconnected players blurs any read on real demand.

The numbers, and the gap

It is the scale that stops you. Per the publicly disclosed commitments, OpenAI has amassed close to $1.4 trillion of compute commitments over the decade, spread across a handful of providers, including roughly $350 billion with Broadcom and $90 billion with AMD on top of the amounts already cited. Against that, its revenue was on the order of $13 billion in early 2026, growing fast but on no common scale with its commitments, and the company was reportedly losing around $14 billion across 2026 per press estimates. This gap between spending promises worthy of a state and still-modest revenue is the knot of the debate.

The echo of the dot-com bubble

Industry veterans see a whiff of déjà vu. In the late 1990s, telecom equipment makers financed their own customers, the carriers, through loans and facilities, to support the build-out of fiber networks. Some carriers even swapped capacity rights between themselves, booking them as sales, while the transactions largely cancelled out. When demand disappointed, the model broke, over-leveraged carriers went bankrupt, and much of the capacity sat unused for years. The fear is that massive cross-commitments now play the same amplifier role, on the way up as on the way down.

The counter-argument

Against this charge there is a serious defence, which must be laid out honestly. Vendor financing is not fraud: it helped build the railways, the telecoms and the first waves of computing, by bootstrapping real markets. Underlying AI demand is not only circular, businesses, developers and individuals pay for genuinely real uses, outside the loop. And the scale of the cross-commitments, though considerable, stays measured relative to the businesses: according to UBS, the OpenAI-Nvidia deal would represent up to 13% of Nvidia’s expected 2026 revenue, around $272 billion, far from making up the bulk of its income. The risk is real, but equating it straight away with accounting fraud would be excessive.

What to watch

Strain, if it builds, will read first in the balance sheets of the infrastructure providers: rising debt, swelling lease commitments, widening credit-default-swap spreads. A warning light already flashed in early 2026, when the press reported that Nvidia’s investment in OpenAI was stalling, triggering a bout of nervousness across three giant market caps before a denial. That is the structural vulnerability of any loop: one link only has to hesitate for confidence to wobble across the whole. The role of a data journal is not to proclaim the bubble, but to map precisely who funds whom, and how much, and to make legible the share of demand that runs in a circle. The rest is a matter of judgement, and judgement needs numbers. For the wider picture, see also the debt behind AI and the BIS warning on the boom’s financial fragility.

---

Primary sources: company statements and disclosures (OpenAI, Nvidia, Oracle, Microsoft, CoreWeave, Amazon Web Services, AMD, Broadcom); Reuters and Bloomberg for the amounts and timeline of the deals (Oracle roughly $300 billion over five years, Nvidia up to $100 billion of investment, CoreWeave $22.4 billion, AWS $38 billion, the $6.3 billion Nvidia-CoreWeave backstop); UBS Chief Investment Office (share of the OpenAI-Nvidia deal in Nvidia’s 2026 revenue, estimated around $272 billion); press estimates for the roughly $13 billion of revenue and OpenAI’s expected 2026 loss. Commitments and dates verified one by one; the amounts are announced commitments, not recorded spending.