AI’s Circular Financing: What Capacity Buyers Must Check

Before you sign the next 15-year colocation lease or build-to-suit power agreement, ask one question that rarely makes it into the term sheet: how much of the demand behind this deal is real end-user revenue, and how much is a vendor financing loop that only works if the music keeps playing? That question has moved from Twitter threads to earnings calls in 2026, as Nvidia, OpenAI, Oracle, and CoreWeave have built a web of investments, supply contracts, and cloud-purchase commitments worth, by some tallies, north of $745 billion — with each company showing up on more than one side of the ledger. For anyone actually buying, leasing, or financing data center capacity right now, this isn’t an abstract Wall Street argument. It’s the counterparty risk sitting underneath a growing share of the tenants knocking on your door.

The deal web, in plain numbers

Start with the pieces, because the dollar figures alone explain why analysts started using the word “circular.” Nvidia has committed to invest up to $100 billion in OpenAI, with the money tied to OpenAI deploying roughly 10 gigawatts of Nvidia systems. OpenAI, in turn, committed to buy around $300 billion of cloud computing capacity from Oracle over roughly five years — compute that Oracle will largely build using Nvidia chips. Separately, OpenAI expanded its compute agreement with CoreWeave by up to $6.5 billion in September 2025, pushing that single relationship’s cumulative value past $22 billion, according to CoreWeave’s own disclosures. Meta has reportedly discussed a cloud purchase from Oracle worth roughly $20 billion on top of that. Oracle’s own backlog of contracted-but-not-yet-recognized revenue (its “remaining performance obligations”) swelled past $500 billion during 2026, a figure the company has pointed to repeatedly as evidence of real, locked-in demand.

Lay those commitments end to end and you get a loop: Nvidia funds OpenAI, OpenAI buys from Oracle and CoreWeave, Oracle and CoreWeave buy chips from Nvidia, and Nvidia’s reported revenue growth from those chip sales helps justify the valuation that lets it keep funding OpenAI. None of this is hidden — it’s disclosed in SEC filings and press releases — but it means a meaningful slice of the “demand” driving today’s capacity buildout traces back to a small number of counterparties financing each other, not to a broad base of independent paying customers.

Why “circular” doesn’t automatically mean fake

It’s worth resisting the simplest version of the bear case. Vendor financing is not new or inherently fraudulent — auto manufacturers have financed dealers for a century, and telecom equipment makers financed their own customers’ network buildouts in the late 1990s. Oracle’s backlog, for what it’s worth, has started converting into real, recognized cloud revenue, not just promises, and that revenue is what ultimately pays for the data center capacity being built against it. Some of this skepticism is also already working itself out in public: Fortune reported in August 2026 that the actual Nvidia-OpenAI data center financing structure landed roughly $145 billion lower than earlier reporting had suggested, a sign that these deals are being renegotiated and tested in the open rather than rubber-stamped. Our earlier look at REIT earnings found landlords like Digital Realty and Equinix still reporting genuine backlog growth and tightening lease terms — evidence that real-economy demand for capacity hasn’t disappeared, even if one layer of it is more concentrated and interlinked than headlines suggest.

The nuance that matters for a buyer isn’t “is this a bubble, yes or no.” It’s that an unusually large share of near-term AI infrastructure demand is now concentrated among a handful of counterparties whose financial health is mutually dependent, which is a structurally different risk than demand coming from a thousand unrelated enterprise customers.

The stress test nobody asked for

That concentration became a live issue the moment interest rates moved against it. Analysts tracking the combined Nvidia-OpenAI-Oracle financing structure have described a Fed rate move in 2026 as the “first stress test” of the roughly $745 billion web, because higher borrowing costs raise the bar every link in the chain has to clear to keep refinancing its share of the buildout. Michael Burry’s Scion Asset Management publicly flagged Nvidia’s equipment depreciation schedules and the circular nature of its OpenAI-adjacent deals as a specific risk in mid-2026, prompting a public back-and-forth with Nvidia’s own leadership. Whatever you think of the short thesis, the exchange is a useful marker: skepticism about this financing structure is no longer fringe, it’s showing up in public filings commentary and triggering real responses from the companies involved.

None of this means the lights go off tomorrow. It means the debt and equity backing a growing share of hyperscale and neocloud capacity is more sensitive to rate moves and sentiment than it was two years ago, which is exactly the kind of fragility a buyer locking in a long-term contract needs to price in. Our prior analysis of counterparty risk in 2026 underwriting covered the balance-sheet side of this; the circular financing web is the demand-side mirror of it — it’s not just whether your tenant can pay, it’s whether the demand that justified their lease in the first place is as independent as it looks.

What a capacity buyer should actually check

If you’re negotiating a lease, a power purchase agreement, or a build-to-suit deal with any AI-native tenant — a foundation model lab, a neocloud, a GPU-as-a-service provider — this changes your due diligence list beyond the standard credit check. First, ask what share of the tenant’s revenue comes from entities that are also their investors or suppliers, and get that disclosed in writing, not inferred from press releases. Second, look past backlog headlines to conversion rates: a $500 billion remaining-performance-obligation figure means little if the realistic timeline to convert it into cash is uncertain. Third, push for parent-company guarantees or letters of credit rather than relying on the operating entity’s standalone balance sheet, especially for newer neoclouds whose debt is already pricing wider than investment-grade hyperscalers. Fourth, consider shorter initial lease terms with renewal options instead of a full 15-year commitment at today’s capacity premium — you give up some rate certainty but you cap your exposure to a single counterparty’s financing structure unwinding. Finally, if you’re a multi-tenant colocation operator, actively diversify tenant mix across enterprise, government, and AI-native workloads rather than let one or two circular-financing-dependent tenants dominate a facility’s revenue.

Where this nets out for 2026 deals

The circular financing debate isn’t a reason to sit out the AI infrastructure buildout — the power, land, and interconnection constraints covered elsewhere on this site make clear that real, durable demand for capacity exists independent of any single lab’s balance sheet. But it is a reason to underwrite the specific counterparty in front of you with more rigor than “they’re an AI company, so the demand must be real.” The companies at the center of this web are disclosing enough for outsiders to map the loop; the diligence failure would be not using that disclosure before you sign.

Before locking in your next capacity commitment, run the counterparty’s revenue concentration and backlog-conversion numbers the same way a lender would — our Data Center Buyers’ Toolkit has the checklists to help you do exactly that.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →