The AI Financing Loop: Who Bears the Default?

D E E P D I V E

ISSUE 27 COMPANION · FRIDAY 4 SEPTEMBER 2026 · R. LAURITSEN

The AI Financing Loop: Who Bears the Default?

A five-question test for separating self-funded demand from supplier-supported demand – and what it says about Nvidia, Broadcom and the next leg of the AI buildout.

Last week we asked who could pass the memory tax on. Broadcom was the only name I refused to force into a category.

This week Broadcom answered the growth question, not the pricing question. AI semiconductor revenue reached $16.7bn, up 221%. Management says it has secured supply for roughly $115bn of AI semiconductor sales in fiscal 2027 and $230bn in 2028. Scale, supply and demand are clear. Whether already-won custom programmes can reprice when costs move is still not.

Then the call moved to financing. Broadcom discussed arranging funding for OpenAI and Anthropic. Its June filing already disclosed an AI-rack lease backstop with maximum exposure of $29bn. Nvidia’s latest filing is larger still: up to $105bn of guarantees supporting 4.25GW of OpenAI data-centre leases in Ohio, with Nvidia infrastructure effectively locked into the site.

The hidden input is no longer just memory. It is also credit.

This does not make the demand fake. Vendor financing is normal in capital-intensive industries. It means part of the underwriting has moved. If the supplier helps fund the customer, revenue quality can no longer be analysed separately from the balance sheet.

HOW BIG IS THE FINANCING LAYER?

Big enough that it now sits beside capex and supply commitments in the filings.

Nvidia has agreed to provide up to $105bn of credit support around OpenAI’s leases at SB Energy’s PORTS campus in Ohio. The guarantees phase in as data centres become operational, the first expected in fiscal 2029, and decline as OpenAI makes payments. Nvidia’s broader ecosystem position is already large: as of 26 July it reported $99bn of equity investments plus $25bn of equity-investment commitments.

Broadcom’s June filing disclosed a five-year backstop on AI-rack lease obligations with maximum exposure of $29bn. The filing did not name the customer. On this week’s earnings call, management explicitly described OpenAI and Anthropic as the two frontier labs for which it may help create financing sources, with future structures evaluated deal by deal.

Oracle, CoreWeave and Meta show the other side of the same buildout. Oracle had $260bn of future data-centre lease commitments at its May year-end. CoreWeave had $35.5bn of leases not yet commenced at June and added a $2.6bn delayed-draw loan in August. Meta is using a BlackRock-backed joint venture for a $14bn, 1GW Texas campus and will lease capacity rather than own the whole asset.

Different structures, same point: the AI boom is now a financing system as much as a technology system.

THE FIVE QUESTIONS

Each can be answered from a filing or earnings call. None requires a credit model.

1. Who would still buy if the vendor stopped supporting the financing?

An investment-grade hyperscaler buying from its own balance sheet is different from a frontier lab whose site requires supplier guarantees or a financing vehicle. Both may have genuine demand. Only one is independent of the seller’s balance sheet.

Look for: customer concentration, named counterparties, credit-support language, and management’s distinction between investment-grade customers and supported AI clouds.

2. What exactly is the support?

Equity, a loan, a lease guarantee, a residual-value backstop and a purchase commitment are not interchangeable. Equity loses value gradually. A guarantee can sit dormant for years, then become a cash obligation after default. A lease can be economically debt-like even when the asset sits elsewhere.

Look for: guarantee, backstop, lease, residual value, variable-interest entity, investment commitment and maximum exposure.

3. When does the cash risk become real?

Nvidia’s $105bn headline is a cap, not a cheque written today. It phases in as facilities become ready and declines as OpenAI pays. Broadcom’s disclosed backstop increases as racks are deployed and decreases with lease payments. Timing changes the risk completely.

Look for: effective date, draw schedule, commencement conditions, amortisation, expiry and whether exposure is gross or net of collateral.

4. What can be recovered if the customer defaults?

A rack full of current-generation accelerators can be resold, redeployed or leased to someone else. A bespoke site, power contract or half-built shell is less portable. The recovery value often matters more than the guarantee cap.

Look for: rights to assume leases, sell equipment, replace tenants, seize collateral or redirect capacity.

5. Is the financing tied directly to future revenue?

This is the question that turns support into strategy. Nvidia’s Ohio arrangement gives the site Nvidia infrastructure exclusivity, subject to limited exceptions. Broadcom’s backstop relates to AI racks built around its custom accelerators. If the guarantee secures years of sales, the right question is return on risk, not simply whether the risk exists.

Look for: exclusivity, preferred-supplier terms, purchase obligations, minimum volumes and the revenue the supported project could generate.

THE LEDGER

Positions, not a single score. The exposures below are not apples-to-apples, so adding them together would create fake precision.

In plain English: backstoppers use their balance sheets to strengthen demand or lock share. Tenants commit to the infrastructure they need. Borrowers finance the capacity and hope utilisation stays high. Software companies such as Snowflake are useful as the control group because the customer pays through usage, not a financed rack.

Two rows deserve more explanation.

Nvidia is the most important. The $105bn cap is enormous, but the structure is deliberately conditional. Exposure rises only as phases go live, and the obligation can terminate if OpenAI reaches a satisfactory credit rating. Nvidia also gains site-level infrastructure exclusivity. This is closer to underwriting a strategic customer than lending it $105bn today.

Broadcom is the cleaner test case. The June filing says maximum exposure on the current backstop is $29bn and gives Broadcom remedies including assuming the lease or selling the AI racks. The earnings call then made clear that financing support could recur for OpenAI and Anthropic. The open question is whether the disclosed exposure stays isolated or becomes a repeatable sales tool.

THREE SIGNALS TO WATCH

Broadcom’s next 10-Q – does maximum backstop exposure stay around the currently disclosed $29bn, or do additional OpenAI / Anthropic structures appear? The direction matters more than the absolute number.

Nvidia’s next filing – is the $105bn OpenAI guarantee a one-off tied to an exceptional campus, or the first template for financing multiple AI clouds? Also track the $99bn of equity investments and $25bn of commitments.

Any Anthropic IPO filing – the prospectus, if and when it arrives, will let investors compare revenue, cash generation and compute obligations against the financing being arranged around it. That will be the cleanest external credit check on one of the system’s largest buyers.

WHAT TO DO WITH THIS

Take your largest AI infrastructure holding and make one page with four lines: current customer concentration, funded ecosystem investments, maximum contingent guarantees, and trailing free cash flow. Then mark which obligations only trigger on default and what assets can be recovered.

The important ratio is not “guarantees divided by market cap.” It is “risk-adjusted support divided by the gross profit and strategic share that support can secure.” You will not get a perfect number. You will get a much better question.

If you cannot tell whether the customer would still build the same project without the supplier’s credit support, that is the finding.

Disclaimer: For information and education only, not financial advice. iPrompt Signals is not a registered investment advisor. Do your own research and consult a qualified financial professional.