Vendor Financing Loops: What 1999 Telecom Tells Us About 2026 AI

On 22 September 2025, NVIDIA announced that it "intends to invest up to $100 billion in OpenAI" as OpenAI deploys at least 10 gigawatts of NVIDIA systems, with the capital released progressively as each gigawatt comes online.¹ The arrangement was disclosed as a non-binding letter of intent, and the structure is circular by construction: NVIDIA puts equity into OpenAI, OpenAI uses the capital to help build data centres filled with NVIDIA chips, and NVIDIA books the chip sales. In its own Q3 FY2026 filing the company cautioned that there is "no assurance" any such investment "will be completed on expected terms, if at all."²

A chip supplier becoming a major investor in its largest prospective customer is not a normal supplier-customer relationship. It is vendor financing, the practice of a manufacturer funding the purchase of its own products, and it has a history. The most-cited precedent is the telecom equipment bubble of 1999 to 2001, when Lucent and Nortel lent billions to the carriers that bought their gear, booked the resulting shipments as revenue, and then wrote much of it off when those carriers failed. The comparison is invoked often enough that it has become a reflex. This piece takes it seriously rather than reflexively: it traces the 2026 loop from primary filings, separates the part that is genuinely vendor-financed from the part that merely looks circular, sets the result against what the 1999 episode actually was, and asks where the rhyme holds and where it breaks. The conclusion is a lens, not a prediction.

Tracing the Circle

The loop has several legs, and each one is documented in a primary source dated before this is written.

The first leg is NVIDIA's capital into OpenAI: the up-to-$100 billion letter of intent of 22 September 2025, structured so that the money is released in tranches tied to gigawatts deployed, with the first gigawatt targeted for the second half of 2026 on the Vera Rubin platform.¹ As of the most recent disclosure available, no definitive agreement had been signed; NVIDIA's chief financial officer confirmed in early December 2025 that the deal "still" was not finalised and was not in the company's revenue guidance.³

The second leg is OpenAI's compute purchase commitments. OpenAI, Oracle and SoftBank announced the Stargate framework on 21 January 2025, a $500 billion buildout targeting 10 gigawatts of capacity, of which $100 billion was described as immediate.⁴ The Wall Street Journal reported on 10 September 2025 that OpenAI had agreed to purchase roughly $300 billion of computing power from Oracle over about five years, effective 2027; Oracle declined to confirm the figure, and the companies' own releases characterise the relationship as one that "exceeds $300 billion" and adds 4.5 gigawatts of Stargate capacity, rather than publishing an itemised contract.⁵⁶ In parallel, OpenAI contracted with CoreWeave, beginning with an $11.9 billion agreement on 10 March 2025 under which CoreWeave also issued OpenAI a $350 million equity stake, then expanded to a cumulative figure of roughly $22.4 billion by late September 2025.⁷⁸

The leg that closes the circle is NVIDIA underwriting that demand directly. On 9 September 2025 CoreWeave signed an order form, disclosed in an 8-K filed 15 September 2025, under which NVIDIA committed to purchase CoreWeave's unsold cloud capacity, an initial $6.3 billion obligation running through 13 April 2032.⁹ NVIDIA had already anchored CoreWeave's March 2025 IPO with an order of approximately $250 million and held a pre-IPO equity stake.¹⁰¹¹ The same template recurs at smaller scale: NVIDIA was reported to be investing up to roughly $2 billion in xAI's funding round in October 2025, and committed $2 billion to Nebius in March 2026.¹²¹³

By October 2025 the arrangement had a canonical visual. A Bloomberg map of "circular" AI deals, widely circulated from 8 October 2025, drew the web of investment, services and hardware flows linking NVIDIA, OpenAI, Oracle, CoreWeave, AMD and a dozen smaller names.¹⁴ The map's own taxonomy is the analytically useful part, because not every line in it is vendor financing. NVIDIA holds investment edges into OpenAI, CoreWeave, Nebius and xAI; it holds no investment edge into Oracle. Oracle is connected to the loop only commercially: it sells cloud capacity to OpenAI and buys chips from NVIDIA, and it funds those chip purchases from its own balance sheet, through the bond issuance whose rating consequences a later piece in this series takes up. The distinction matters for an article about vendor financing specifically. The vendor-financed core of the loop is narrow: NVIDIA's equity in OpenAI and its purchase backstop for CoreWeave. Oracle is the largest single node by dollar flow, but it is a self-financed conduit, not a vendor-financed one, and lumping it in overstates how much of the circle is actually NVIDIA's own capital returning to it.

The pattern is also not NVIDIA's alone. On 6 October 2025, AMD and OpenAI announced that OpenAI would deploy 6 gigawatts of AMD Instinct GPUs over several years, and AMD issued OpenAI a warrant for up to 160 million AMD shares, vesting in tranches tied to deployment and to AMD share-price targets.¹⁵ That is vendor financing in close to its purest modern form: equity handed to the customer in exchange for a commitment to buy the supplier's product. It is worth holding in view, because it shows the mechanism is structural to the present cycle rather than a quirk of one company.

Whether the vendor-financed core amounts to recycling or to ordinary ecosystem investment is contested, and the contest is worth reporting directly. Jeremy Kahn, writing in Fortune on 28 September 2025, framed the open question as "how much of the AI boom is just Nvidia's cash being recycled."¹⁶ Bernstein's Stacy Rasgon noted the arrangements "will clearly fuel 'circular' concerns." The dispute cannot be settled from filings alone, because NVIDIA does not name its customers: its 10-Q discloses that two direct customers accounted for 23 percent and 16 percent of revenue in the quarter ended 27 July 2025, but those "direct" customers are integrators and distributors, not the end buyers whose demand the loop is funding.¹⁷

What 1999 Actually Was

The telecom analogy is worth stating precisely, because the loose version of it misleads. Between roughly 1998 and 2001, the dominant equipment makers extended credit to the carriers buying their switches and fibre gear, disproportionately to the competitive local exchange carriers, the CLECs, that had little revenue and depended on continued capital-market access. Lucent committed on the order of $8.1 billion of vendor financing; Nortel committed roughly $3.1 billion, of which about $1.4 billion was drawn; Cisco promised around $2.4 billion in customer loans.¹⁸ McKinsey later put the combined vendor-financing exposure of nine suppliers at approximately $25.6 billion by the end of 2000.

The mechanism that made it dangerous was the booking, not the lending as such. A vendor loan let the supplier recognise equipment revenue at the point of shipment, while the corresponding risk sat on its balance sheet as a receivable. As outside capital fled the sector in early 2001, the vendors did not pull back; they accelerated financing to keep weakening customers buying, with Nortel reported to be financing as much as 130 percent of equipment cost on terms that were often unsecured and tied to future purchases. When the carriers failed, the booked revenue reversed into uncollectible receivables. Lucent took bad-debt provisions of roughly $2.2 billion in fiscal 2001 and a further $1.3 billion in fiscal 2002, about $3.5 billion of customer-loan losses, of which some $700 million was attributable to a single carrier, Winstar, after Lucent declined a final $90 million extension and Winstar filed for bankruptcy.¹⁸ The Securities and Exchange Commission separately charged Lucent with improperly recognising $1.148 billion of revenue in fiscal 2000, a matter settled with a $25 million penalty.¹⁹

Two features of that episode are the ones worth carrying forward. The failure mode was not gradual; it was a synchronised unwind, with roughly four dozen CLECs entering bankruptcy across 2000 to 2003 as the same funding source dried up for all of them at once. And the early-warning sign was visible before the collapse: it was the loosening of terms, the willingness to finance a higher share of cost on weaker security, to keep a deteriorating customer in the game.

Where the Rhyme Holds

The structural similarity is real and operates at three points. First, both arrangements pull demand forward rather than creating it: a customer that could not otherwise fund a purchase at this pace is enabled to make it now, which means some portion of present revenue is borrowed from a future that has to arrive on schedule. Second, both concentrate counterparty risk on the supplier's balance sheet, so the supplier's reported strength is partly a claim on its customers' survival. Third, both feature concentration severe enough that the failure of one or two nodes is not a diversifiable event: NVIDIA's two-customer 39 percent figure, while measured at the direct-buyer level, is a reminder that the demand base is narrow.¹⁷

The diagnostic from 1999 transfers cleanly. The thing to watch is not the existence of the loop, which is openly disclosed, but any loosening of its terms: an expansion of NVIDIA's purchase backstops, financing extended on weaker security or to weaker nodes, or capacity commitments that grow faster than the end-demand that is supposed to absorb them.

Where the Rhyme Breaks

Two differences are large enough that the analogy is a lens and not a forecast, and an honest treatment keeps them in view rather than waving them away.

The first is the instrument. Lucent and Nortel extended trade credit, which means they recognised revenue at the sale and carried a receivable that, when it went bad, reversed through the income statement as a charge against earnings. NVIDIA's outbound exposure is mostly equity and purchase commitments, not loans. Under the relevant accounting, an equity holder's loss is generally capped at the carrying amount of the investment, lands in non-operating income, and never inflates revenue in the first place, so there is no booked-revenue reversal to suffer; in a bankruptcy the equity is simply wiped out, ahead of any creditor. The revenue-recognition and collectibility tests that the SEC used against Lucent bite on receivables, not on minority equity stakes. The risk does not disappear, but it sits in a different place and behaves differently when it crystallises.

The second difference is the counterparty. The CLECs that received vendor financing in 1999 had, in most cases, almost no revenue and no demonstrated business model. The 2026 counterparties are loss-making but not revenue-less. OpenAI was reported in early March 2026 to be running at an annualised revenue rate of roughly $25 billion; CoreWeave reported $5.1 billion of revenue for fiscal 2025 against a $1.2 billion net loss and a $60.7 billion remaining-performance-obligation backlog; Oracle, the largest node, is profitable and reported $523 billion of remaining performance obligations in December 2025.²⁰²¹²² "Real but unprofitable" is a materially different starting point from "no revenue," and it changes the timing of any unwind even if it does not change its direction.

The Counter-Case

The strongest version of the opposite view deserves a full hearing, because it is more than a quibble.

The first and most serious objection is that the loop does not, in fact, fund most of its own demand. NVIDIA's up-to-$100 billion commitment to OpenAI is large in absolute terms but small against the total capital behind OpenAI's purchasing: Microsoft's multi-year investment, the SoftBank-led funding rounds, and OpenAI's own revenue run-rate of roughly $25 billion all sit alongside it. The economist Noah Smith argued on 22 October 2025 that the flow is one-directional, fully disclosed, and a form of diversification rather than the concealed round-tripping of the dot-com era.²³ The honest concession is that the bulk of the demand NVIDIA books is funded by third parties, not by NVIDIA. What survives the objection is narrower: a supplier that becomes a marginal, relatively price-insensitive funder of its own customer can still flatter the demand signal at the edge, even when it is not manufacturing the demand wholesale.

The second objection is that the magnitude cannot be quantified. Because NVIDIA does not name its end customers, the share of its revenue that is genuinely circular is unknown. The most-cited public estimate, from NewStreet Research, is that roughly $10 billion of NVIDIA investment is associated with about $35 billion of GPU revenue, on the order of a quarter of the prior year's total; but that is a single analyst's construction, not a figure that can be tied to a filing.²⁴ A piece that leaned on it as a hard number would be overstating what is knowable.

The third objection cuts against the instrument argument and may cut in the bulls' favour. If the exposure is equity rather than a receivable, the loop may be less systemically dangerous than 1999, not more: there is no booked-revenue reversal, no bad-debt charge cascading through the income statement, and no vendor receivable secured against a failing carrier. The loss is confined to the carrying value of NVIDIA's investment. The instrument difference, in other words, can be read as making the structure safer, not merely different.

The fourth objection is about the asset. The 1999 CLEC fibre was, in large part, dark and unused; the demand it was financed against never materialised. AI compute is being consumed now, by products with hundreds of millions of users. If the end-demand is real and productive, then "pulling it forward" is closer to financing a genuine buildout than to borrowing from a future that may not arrive.

What survives all four objections is modest but not nothing: the concentration is real, the headline figures are partly conditional (the OpenAI investment was an unsigned letter of intent, the Oracle figure press-reported rather than filed), and the speed-of-unwind property holds if end-demand disappoints. The lens is a way to monitor those risks, not a forecast that they will crystallise.

What to Watch

The 1999 episode does not predict 2026; it supplies a checklist. Three indicators will resolve the open question over the coming quarters. First, the terms: whether NVIDIA's and AMD's purchase backstops and equity commitments expand, and whether any new financing is extended on weaker security or to weaker counterparties, which is the 1999 tell. Second, the gap between committed capacity and realised end-demand: whether OpenAI's and the neoclouds' revenue converts on the schedule the compute commitments assume, or whether the commitments keep growing faster than the revenue behind them. Third, the conversion of the loop's headline figures into signed, funded reality: the NVIDIA-OpenAI investment was still a letter of intent at the latest disclosure, and the Oracle figure remains press-reported rather than filed, so the difference between announced and executed is itself a thing to track.

Personal View

The body above traced the loop. What follows is my own view of what to make of it.

The reflex I want to argue against is the one that reaches for "round-tripping" the moment a vendor's capital shows up in its customer's funding round. The 2026 loop is real and it is disclosed: every leg traced above is in a filing, a press release or a named report, which is precisely what the 1999 loop was not. Treating openly disclosed minority investment as concealed accounting fraud confuses the part of the analogy that survives with the part that does not.

What survives is also narrower than the canonical Bloomberg map made it look. The vendor-financed core is NVIDIA's equity in OpenAI and its purchase backstop for CoreWeave, with the AMD warrant in close proximity. Oracle is the largest dollar node, but Oracle is a self-financed conduit: it sells compute to OpenAI and buys chips from NVIDIA, and it pays for the chips out of its own bond issuance, not out of NVIDIA's. Lumping Oracle into the round-trip overstates how much of the circle is NVIDIA's own capital returning to it. The narrower picture is the right picture.

The risk I take seriously is not fraud and it is not concealment. It is the synchronisation property the 1999 episode actually showed: forty-odd CLECs failing in close sequence because the same funding source dried up for all of them at once. The 2026 counterparties are revenue-generating, not revenue-less, which changes the timing of any unwind but not its shape. What changes the shape is the instrument. Equity wipes out in non-operating income, ahead of any creditor; trade credit reverses through revenue and propagates through bad-debt charges. That difference probably makes the structure systemically less dangerous than 1999, and it also makes the damage less visible while it is happening.

So the indicator I would actually watch is not the existence of the deals, which is fully public, but their terms. The 1999 tell was not the size of the financing; it was the loosening of it: weaker counterparties, a higher financed share of equipment cost, financing extended to keep deteriorating buyers in the game. If those terms loosen visibly in 2026 or 2027, the lens stops being a lens.

This is part 2 of a four-part series on AI infrastructure financing. Part 1 examined the SPV structures financing the GPU clusters the loop's compute capacity is built on. Part 3 follows the obligation upstream into the quarter-trillion of compute commitments hidden in 10-K footnotes; Part 4 closes the series with the on-balance-sheet credit profile of the hyperscalers absorbing the demand.

END · 02