The most consequential question in the artificial intelligence boom is not whether the technology is transformative. It is whether the demand that the market has priced into the world's largest companies can be trusted as an independent fact. Those are different questions, and the distance between them is precisely where forensic accounting lives.
Over the past year, the firms at the center of the AI build-out have entered into a dense lattice of reciprocal arrangements. A chip designer announces its intent to invest as much as $100 billion in a model developer that has, in turn, committed to deploy that designer's hardware at the scale of small nations' power grids. A second chip designer grants that same model developer warrants to acquire roughly a tenth of its equity, with vesting tied directly to how many of its processors the developer agrees to buy. A cloud provider books a backlog of contracted revenue larger than the annual output of most economies, the majority of it owed by a single, unprofitable counterparty. A specialized data-center operator purchases its hardware from the very chip designer that owns a double-digit stake in it and has separately agreed to buy any capacity the operator cannot sell on its own. In each instance, capital supplied by the seller becomes, directly or indirectly, the purchasing power of the buyer.
The transactions are real. The cash moves. Revenue is recognized under generally accepted accounting principles. The question Buxton Helmsley believes every institutional allocator must now confront is narrower and considerably more uncomfortable: how much of the demand reported across this ecosystem would exist if the sellers were not, in one form or another, financing the buyers?
This is not a novel question. The market answered a version of it a quarter-century ago, at staggering cost, and then appears to have forgotten the answer.
The clearest way to understand the concern is to trace the capital, because the capital travels in a circle.
In September 2025, Nvidia Corporation announced its intent to invest up to $100 billion in OpenAI, tied to OpenAI's plan to deploy at least 10 gigawatts of data-center capacity built on Nvidia systems.¹ To its credit, Nvidia stated that the investment was intended to take the form of equity rather than vendor debt, and that the invested funds were not earmarked for direct purchases of Nvidia products.² That distinction is real and deserves to be acknowledged. It is also less protective than it first appears, because capital is fungible. An equity infusion that funds a data-center build-out relieves the recipient of the need to raise that capital elsewhere, with substantially the same net effect on its capacity to purchase the sponsor's hardware. By late 2025, the framework remained unfinalized and the headline figure was reported to be in flux, even as Nvidia separately committed up to $10 billion to a competing model developer.³ Whatever its final form, the arrangement was sufficiently striking that one prominent semiconductor analyst observed it would "clearly fuel" concerns about circularity, and another, who maintains a rare sell rating on the stock, compared it to a parent co-signing a child's first mortgage.⁴
The second leg of the structure is more explicit still. In October 2025, Advanced Micro Devices, Inc. agreed to supply OpenAI with up to 6 gigawatts of its Instinct graphics processors and, in the same transaction, issued OpenAI a warrant to purchase up to 160 million AMD shares at an exercise price of $0.01 per share.⁵ As disclosed in AMD's own filing with the Securities and Exchange Commission, the warrant vests in tranches keyed to OpenAI's purchases of AMD hardware, beginning with the initial 1 gigawatt deployed and reaching full vesting only upon the purchase of all 6 gigawatts, with vesting further conditioned on AMD's share price reaching escalating targets of up to $600 per share.⁶ If fully exercised, the warrant would convey approximately 10% of AMD to its largest prospective customer.⁷ AMD's chief financial officer described the arrangement as expected to be "highly accretive" to the company's non-GAAP earnings per share.⁸ The candor of that characterization is itself instructive: an equity instrument issued to a customer, vesting on the customer's purchases, was presented to the market as a driver of adjusted profitability. Even a competing chief executive remarked that he was surprised a company would "give away 10%" of itself "before they even built it," allowing only that the structure was "clever."⁹
The third leg shifts the circle from the supplier side to the customer side. In its fiscal first quarter of 2026, Oracle Corporation disclosed that its remaining performance obligations—the contracted revenue it has booked but not yet delivered—had risen 359% year over year, to approximately $455 billion.¹⁰ Subsequent reporting indicated that a single counterparty, OpenAI, accounted for at least $300 billion of that backlog under a five-year compute agreement reported to begin in 2027.¹¹ Oracle's management guided that remaining performance obligations would surpass $500 billion as further contracts were signed.¹² The reception was not uniformly enthusiastic. As the scale of Oracle's reliance on a single, cash-burning customer became clear, the stock surrendered a large portion of its gains, shedding tens of billions of dollars of market value in a single session, even as the company committed to maintaining an investment-grade credit rating it had funded in part through one of the largest corporate bond sales in recent technology history.¹³
The arithmetic beneath that backlog is the heart of the matter. OpenAI reported revenue on the order of $13 billion on an annualized basis in 2025, does not expect to reach profitability until approximately 2029, and has told investors it anticipates spending sums measured in the trillions on infrastructure.¹⁴ A backlog is not the same thing as collected revenue. Remaining performance obligations represent contracted but undelivered services; they can be delayed, renegotiated, or in some circumstances canceled, and they are not guaranteed to convert.¹⁵ When the obligor's committed spending vastly exceeds what its own operations can presently fund, the durability of that backlog becomes a function not of the seller's execution but of the buyer's continued access to outside capital.
The fourth leg closes the circle visibly. CoreWeave, Inc., a specialized provider of AI computing capacity, purchases its hardware from Nvidia, which has been a shareholder of CoreWeave since 2023 and increased its stake to approximately 11.5% through a $2 billion investment in early 2026.¹⁶ CoreWeave, in turn, leases that capacity primarily to a small number of large customers, among them Microsoft and OpenAI—the latter itself a recipient of Nvidia capital.¹⁷ In September 2025, the relationship acquired one further feature: Nvidia agreed to purchase up to $6.3 billion of CoreWeave's unsold cloud capacity through April 2032, a commitment commonly described as a backstop and disclosed only after the fact in a securities filing.¹⁸ The economic shape of the arrangement is therefore complete. The supplier invests in the operator, sells hardware to the operator, and guarantees the operator's revenue if the market will not, while the operator's largest tenants are themselves the supplier's funded counterparties. CoreWeave's own disclosures underscore the fragility this concentration can introduce: for the third quarter of 2025 it reported a contracted revenue backlog of $55.6 billion while relying on a small number of large customers, carrying billions of dollars of debt, and posting quarterly interest expense that had roughly tripled year over year, to $311 million, against a quarterly net loss of $110 million.¹⁹ The company's chief executive has characterized allegations of circular financing as "ridiculous," noting that Nvidia's stake remains modest relative to the total capital CoreWeave has raised.²⁰ That is a fair point, and it is also not a complete answer to the structural question.
It is tempting to file the foregoing under the heading of bubble psychology and move on. Buxton Helmsley regards that as a category error. The relevant issues are not matters of mood; they are matters squarely addressed by the accounting and disclosure framework, which exists precisely to let investors distinguish demand that is independent and collectible from demand that is sponsored and contingent.
Begin with revenue recognition itself. Under Accounting Standards Codification Topic 606, an entity may recognize revenue from a contract with a customer only when, among other criteria, it is probable that the entity will collect substantially all of the consideration to which it will be entitled in exchange for the goods or services transferred.²¹ Collectibility is not a footnote to the revenue model; it is a gating condition for the existence of a contract under the standard. When a seller has financed, invested in, or guaranteed the resources of the very party from which it expects to collect, the probability of collection ceases to be an arm's-length judgment and becomes, in part, a judgment about the seller's own continued willingness to fund the buyer. That is not necessarily improper. It is, however, exactly the kind of arrangement that the collectibility criterion was designed to force management to scrutinize and disclose, rather than to presume.
Turn next to the backlog disclosures that have so excited the market. Topic 606 requires entities to disclose the transaction price allocated to performance obligations that are unsatisfied at the reporting date.²² That disclosure is useful, but it is a measure of contracted intention, not of realized or even probable cash. A remaining-performance-obligation figure dominated by a single counterparty whose ability to pay depends on raising hundreds of billions of dollars it does not yet possess is a materially different asset from a backlog diversified across self-funding customers, even when the two carry identical headline values. The standard discloses the number; it does not, on its own, disclose the number's quality. That is the responsibility of management's broader disclosure and of the investor's diligence.
Consider, third, the question of related parties. Where one enterprise holds an equity interest, warrants, or other influence over another with which it transacts, Accounting Standards Codification Topic 850 requires disclosure of the nature of the relationship, a description of the transactions, and the dollar amounts involved.²³ A separate question is whether such a holding confers significant influence, which under the equity-method guidance is generally presumed at 20% of voting interests but can exist below that threshold depending on the facts.²⁴ Buxton Helmsley does not assert that any particular counterparty in the AI ecosystem is, as a matter of law, a related party of any other; that determination turns on facts the issuers themselves are best positioned to know and to disclose. We do assert that double-digit equity stakes, warrants convertible into roughly a tenth of a counterparty, and seats at the table of one another's capital raises are exactly the circumstances in which the related-party and significant-influence analyses become non-trivial, and in which investors are entitled to see management's reasoning rather than its silence.
Fourth, consider the backstops. An agreement by a supplier to purchase a customer's unsold output functions economically as a guarantee of that customer's revenue. Arrangements of that character may fall within the scope of Accounting Standards Codification Topic 460, which governs the recognition and disclosure of guarantees, including the nature of the guarantee, its term, and the maximum potential amount of future payments.²⁵ When such commitments are disclosed only after the fact, or are embedded in the narrative of a strategic partnership rather than quantified as the contingent obligations they are, the investor is left to reconstruct the firm's true risk profile from press releases.
Finally, and most concretely, consider customer concentration. Accounting Standards Codification Topic 280 requires an entity to disclose the existence and magnitude of revenues from any single external customer that amount to 10% or more of the entity's total revenues.²⁶ Topic 275 separately requires disclosure of concentrations that make an entity vulnerable to a near-term severe impact, including concentration in the volume of business transacted with a particular customer.²⁷ These are not obscure provisions, and the AI build-out has tested them in ways that should give allocators pause. In its annual report on Form 10-K for the fiscal year ended January 25, 2026, Nvidia disclosed that one direct customer accounted for 22% of its total revenue and another for 14%, both concentrated in a single segment, up sharply from the prior year.²⁸ On a quarterly basis the concentration ran higher still: by its fiscal third quarter, four direct customers each exceeding 10% of revenue together accounted for roughly 61% of sales, against 36% a year earlier.²⁹ Nvidia satisfies the letter of the standard. It also identifies these customers only as "Customer A," "Customer B," and so forth, declining to name them, even as it discloses elsewhere that it holds investments in, and commercial relationships with, several of the largest buyers and end-users of its hardware.³⁰ An investor attempting to determine how much of one of the most heavily weighted companies in the major equity indices depends on counterparties that the company itself has funded is therefore left to infer, from anonymized concentration figures and separately disclosed investments, a picture the disclosure framework gestures toward but does not assemble. The standards were written to surface dependency. They were not written to surface circularity, and that gap is the entire subject of this paper.
Briefings on markets, governance, and our investment philosophy—delivered directly.
None of this would feel unfamiliar to anyone who lived through the last great infrastructure boom financed by its own suppliers.
In the closing years of the 1990s, the dominant manufacturers of telecommunications equipment confronted a wave of newly formed carriers that wanted to buy their gear but could not always afford it. The manufacturers solved the problem by lending their customers the money. By the end of its fiscal year 2000, Lucent Technologies had entered into agreements to provide as much as $8.1 billion in customer financing and loan guarantees, of which roughly $2.1 billion was outstanding.³¹ Among the recipients was a competitive carrier to which Lucent extended approximately $2 billion in vendor financing; when the carrier faltered and Lucent declined to advance a final tranche, the carrier filed for bankruptcy and Lucent wrote off approximately $700 million.³² Across 2001 and 2002, Lucent recorded provisions for customer bad debts of $2.2 billion and $1.3 billion, respectively.³³ Separately, Lucent disclosed and reversed previously recognized revenue, and in 2004 the Securities and Exchange Commission charged the company and several individuals in connection with the improper recognition of roughly $1 billion in revenue, with Lucent agreeing to a $25 million penalty assessed in significant part for a failure to cooperate.³⁴ Its principal rival followed a similar path, extending well over $1 billion in vendor financing of its own before a multi-year sequence of restatements, investigations, and decline that ended, years later, in one of the largest technology bankruptcies on record.³⁵ A leading networking-equipment maker promised billions more in customer loans and, when the cycle turned, took a single inventory write-down of approximately $2.25 billion.³⁶
The downstream carnage was indiscriminate. Dozens of the new carriers that the equipment makers had financed went bankrupt between 2000 and 2003, and the funding that had appeared limitless in January 2001 had effectively vanished by that April.³⁷ The mechanism is worth stating plainly, because it is the mechanism that matters now. Vendor financing did not merely add risk at the margin. It manufactured the appearance of end demand. Equipment sold to customers who could pay only because the seller had financed them was recognized as revenue, validated growth narratives, and supported valuations—until the moment the sellers could no longer extend credit, at which point the revenue reversed, the receivables soured, and the write-offs cascaded through the very firms that had reported the strongest growth.
Intellectual honesty requires acknowledging the differences between that episode and this one, and they are not trivial. The telecom manufacturers extended debt; today's sponsors describe their commitments primarily as equity, which carries no contractual repayment and therefore cannot default in the same fashion. The earlier customers were thinly capitalized start-ups; some of today's funded counterparties are among the most richly valued private companies ever assembled. And the present demand for computing capacity is, by any measure, enormous and at least partly genuine in a way the speculative fiber overbuild was not. Buxton Helmsley does not contend that the AI build-out is a fiber-optic bubble in new clothing. We contend that the economic question the earlier episode posed has not changed merely because the instruments have: when a seller's own capital is what allows the buyer to buy, the reported demand is not a fully independent fact, and the magnitude and durability of that demand cannot be assessed without isolating the portion that the sellers are funding.
The participants and their defenders offer a coherent rebuttal, and it deserves to be engaged rather than dismissed.
The strongest version runs as follows. The demand for AI computing is not manufactured; it is real, constrained by physical supply, and growing faster than the industry can build. In a market defined by scarcity of the most advanced chips and the power to run them, buyers do not simply place orders; they secure supply by pairing long-term purchasing commitments with capital, and sellers reasonably take equity in the customers whose success will determine their own. On this view, the reciprocal arrangements are not a closed loop concealing weakness but, as one large asset manager has described them, a "virtuous circle" that aligns suppliers, builders, and end users to meet genuine and exploding demand.³⁸ Equity, the argument continues, is categorically different from the vendor debt that destroyed the telecom manufacturers: a minority equity stake cannot trigger a default spiral, and where a sponsor's investment is small relative to the capital the recipient has raised from unrelated parties, it cannot plausibly be said to be propping up demand.²⁰ And the underlying usage is observable: hundreds of millions of people use these products, and revenue, while dwarfed by commitments, is growing at rates rarely seen.
Each of these points has force, and a fair analysis concedes them. Demand for computing capacity is plainly real. Equity is plainly safer than debt. And a sponsor's stake that is small relative to a recipient's total capitalization is plainly less determinative of that recipient's behavior than a controlling one.
The forensic rejoinder is not that these points are wrong, but that they are insufficient, because they answer a question that is not the one in dispute. The issue is not whether AI demand exists; it is whether the reported magnitude and durability of that demand can be trusted as independent of the sponsors' own capital. Real demand and circular financing are not mutually exclusive; the latter can inflate the measured size and apparent permanence of the former. The relevant inquiry is therefore marginal and quantitative: of the revenue and backlog being reported across this ecosystem, how much would survive if the reciprocal capital were withdrawn? That is a number the participants are uniquely able to produce and have not produced. Even the most candid voice in the debate has conceded the premise. Asked how its largest prospective counterparty would actually pay for the hardware it had committed to buy, the chief executive of the leading chip designer answered simply that the counterparty did not yet have the money, and would need to raise it through revenue, equity, or debt.³⁹ That is not an indictment. It is, however, an acknowledgment that the demand has been booked in advance of the means to fund it, which is the precise condition under which reciprocal financing does its most flattering work. It is also worth recalling that the demand-skeptics are not confined to short-sellers; one widely cited study from the Massachusetts Institute of Technology reported that the overwhelming majority of enterprise AI pilots had failed to deliver measurable returns, a finding that, whatever its limitations, sits uneasily beside commitments measured in the trillions.⁴⁰
Buxton Helmsley does not believe the appropriate response to the foregoing is to predict a collapse or to short an asset class. We believe the appropriate response is to relocate the burden of proof to where it belongs—onto the issuers—and to insist, as a condition of underwriting, on disclosure sufficient to assess whether reported demand is independent, collectible, and complete. The following are the demands we believe every institutional allocator should now make, of the equity issuers in which it invests and, equally, of the managers to whom it allocates capital that is exposed to these names.
First, allocators should demand a gross-up of revenue and backlog attributable to counterparties in which the issuer holds equity, warrants, guarantee exposure, or other financial interest. A single line disclosing the dollar value of revenue and remaining performance obligations associated with funded or sponsored counterparties would do more to clarify the AI ecosystem than any quantity of management commentary, and its absence should itself be treated as informative.
Second, allocators should demand a standalone collectibility analysis for the largest backlog counterparties. Where a single obligor accounts for a material share of remaining performance obligations, the issuer should be asked to articulate the basis on which it concluded that collection is probable, addressing the counterparty's own capacity to fund the commitment from sources other than the issuer and its affiliates.
Third, allocators should demand explicit related-party and significant-influence determinations, together with the reasoning behind them. Where an issuer transacts with a counterparty in which it holds a double-digit stake or convertible instruments approaching one, investors are entitled to know whether management concluded that the counterparty is a related party, whether it concluded that significant influence exists, and on what facts.
Fourth, allocators should demand that backstops, capacity-purchase commitments, and similar revenue guarantees be quantified as the contingent obligations they are, with their maximum potential exposure and term disclosed prospectively rather than narrated retrospectively.
Fifth, allocators should demand that customer concentration be disclosed with sufficient specificity to be useful. Where materiality warrants, the anonymization of customers that an issuer is simultaneously financing deprives the concentration disclosure of its purpose, and investors should press for identification or, at minimum, for disclosure of the extent to which the issuer's 10% customers overlap with its investees and counterparties.
Sixth, and most simply, allocators should demand a sensitivity disclosure: a clear statement of what portion of reported revenue and growth would remain if the reciprocal capital were removed. No participant in this ecosystem has volunteered that figure. Until one does, the prudent assumption is that the figure is material, because participants disclose the numbers that help them and omit the numbers that do not.
For the limited partners and fiduciaries who allocate to managers rather than to securities directly, these demands translate into a single discipline. Before signing the next commitment to a strategy with meaningful exposure to the AI infrastructure complex, the allocator should require its manager to demonstrate that it has stress-tested the loop—that it has modeled the effect of a withdrawal of sponsor capital on the revenue, backlog, and creditworthiness of the names it holds—and has not simply underwritten the headline growth at face value. A manager who cannot describe the circularity in its own portfolio has not analyzed it.
Buxton Helmsley is not forecasting the failure of artificial intelligence, nor the collapse of the companies discussed in this paper. We are making a narrower and, we believe, more durable point. The accounting and disclosure framework was constructed over decades to allow investors to distinguish demand that is genuine, independent, and collectible from demand that is sponsored, circular, and contingent. The current build-out has produced a web of reciprocal arrangements that the framework, as presently applied, surfaces only in fragments—a concentration figure here, an after-the-fact backstop there, an anonymized customer somewhere else—and never assembles into the single picture an allocator needs. The remedy is not suspicion. The remedy is disclosure, demanded as a condition of capital.
The last time suppliers financed the demand for their own products at this scale, the market mistook the financing for the demand, and the correction, when it came, fell hardest on the firms that had reported the most. The lesson was expensive and it was clear. The most dangerous figure in any boom is not the one that looks too high. It is the one that everyone has quietly agreed not to question.
May 21, 2026
Read
May 18, 2026
Read
May 15, 2026
Read
Briefings on markets, governance, and our investment philosophy—delivered directly.
Referenced Sources:
Bloomberg, "Nvidia's Massive OpenAI Deal Fuels 'Circular' Financing Concerns," September 23, 2025, https://www.bloomberg.com/news/articles/2025-09-23/nvidia-s-massive-openai-deal-fuels-circular-financing-concerns.
Axios, "Nvidia's $100B OpenAI Investment Fuels AI Bubble Fears," September 25, 2025, https://www.axios.com/2025/09/25/nvidia-openai-investment-ai (reporting that Nvidia told the research firm Bernstein its OpenAI investment would not be used for any "direct purchases" of Nvidia products); Highline Wealth Partners, "Circular Investing in the AI Ecosystem," October 9, 2025, https://www.highlinewp.com/post/circular-investing-in-the-ai-ecosystem (noting Nvidia's stated position that the investment would likely take the form of equity rather than vendor financing or debt).
Reuters, "Nvidia Says $100 Billion OpenAI Investment Framework Not Yet Finalized," via Yahoo Finance, December 2, 2025, https://finance.yahoo.com/news/nvidia-says-100-billion-openai-183732361.html.
Business Standard (citing Bloomberg), "Nvidia-OpenAI Deal Sparks Concerns Over Circular Financing in AI Boom," September 24, 2025, https://www.business-standard.com/amp/companies/news/nvidia-openai-deal-sparks-concerns-over-circular-financing-in-ai-boom-125092401589_1.html (quoting Bernstein Research analyst Stacy Rasgon and Seaport Global Securities analyst Jay Goldberg).
CNBC, "OpenAI Looks to Take 10% Stake in AMD Through AI Chip Deal," October 6, 2025, https://www.cnbc.com/2025/10/06/openai-amd-chip-deal-ai.html.
Advanced Micro Devices, Inc., Current Report on Form 8-K, filed October 6, 2025, U.S. Securities and Exchange Commission, https://www.sec.gov/Archives/edgar/data/0000002488/000119312525230895/d28189d8k.htm (describing the warrant for up to 160 million shares at an exercise price of $0.01 per share, tranche vesting tied to purchases of 1 through 6 gigawatts of AMD Instinct GPU products, share-price targets escalating to $600 for the final tranche, and an expiration date of October 5, 2030).
CNBC, "OpenAI Looks to Take 10% Stake in AMD Through AI Chip Deal," October 6, 2025.
Advanced Micro Devices, Inc., "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs," press release, October 6, 2025, https://www.amd.com/en/newsroom/press-releases/2025-10-6-amd-and-openai-announce-strategic-partnership-to-d.html (quoting Jean Hu, Executive Vice President, Chief Financial Officer and Treasurer, AMD, that the agreement "is expected to be highly accretive to AMD's non-GAAP earnings-per-share"); see also OpenAI, "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs," October 6, 2025, https://openai.com/index/openai-amd-strategic-partnership/.
CNBC, "Nvidia's Huang Says He's Surprised AMD Offered OpenAI 10% of Company in 'Clever' Deal," October 8, 2025, https://www.cnbc.com/2025/10/08/nvidia-huang-amd-open-ai.html.
Sherwood News, "What Is an RPO, the Number That Drove Oracle's Giant Share Move?," September 10, 2025, https://sherwood.news/markets/what-is-an-rpo-the-number-that-drove-oracles-giant-share-move/ (reporting remaining performance obligations of $455 billion, a 359% increase, and quoting Oracle Chief Executive Officer Safra Catz).
The Wall Street Journal, as reported by DatacenterDynamics, "OpenAI Signs $300bn Cloud Deal With Oracle—Report," February 11, 2026, https://www.datacenterdynamics.com/en/news/openai-signs-300bn-cloud-deal-with-oracle-report/ (five-year agreement reported to begin in 2027); Yahoo Finance, "Oracle Made a $300 Billion Bet on OpenAI. It's Paying the Price," December 12, 2025, https://finance.yahoo.com/news/oracle-made-a-300-billion-bet-on-openai-its-paying-the-price-205441863.html.
CNBC, "Oracle's AI-Fueled Debt Load Has Investors on Edge Ahead of Quarterly Earnings," December 9, 2025, https://www.cnbc.com/2025/12/09/oracles-ai-fueled-debt-load-has-investors-on-edge-ahead-of-earnings.html.
Yahoo Finance, "Oracle Made a $300 Billion Bet on OpenAI. It's Paying the Price," December 12, 2025; CNBC, "Oracle's AI-Fueled Debt Load Has Investors on Edge," December 9, 2025 (reporting Oracle's $18 billion bond sale and BBB credit rating).
DatacenterDynamics, "OpenAI Signs $300bn Cloud Deal With Oracle—Report," February 11, 2026; Fierce Network, "Oracle's Massive OpenAI Deal Raises Equally Big Questions," September 11, 2025, https://www.fierce-network.com/cloud/oracles-massive-openai-deal-raises-equally-big-questions.
AOL, "Why OpenAI's $300 Billion Deal With Oracle Has Set the 'AI Bubble' Alarm Bells Ringing," 2025, https://www.aol.com/articles/why-openai-300-billion-deal-150815932.html (noting that remaining performance obligations represent contracted but undelivered services that customers may delay, renegotiate, or cancel); see also note 22 and FASB Accounting Standards Codification 606-10-50-13.
The Motley Fool, "Nvidia Just Piled $2 Billion Into This Key AI Partner. Should Investors Follow Suit?," February 1, 2026, https://www.fool.com/investing/2026/02/01/nvidia-just-piled-2-billion-in-this-key-ai-partner/ (reporting Nvidia's increased stake of approximately 11.5%).
The Motley Fool, "Nvidia Just Piled $2 Billion Into This Key AI Partner," February 1, 2026 (identifying Microsoft, Meta, and Nvidia among CoreWeave's customers); AInvest, "Is CoreWeave-Nvidia Circular Financing a Red Flag or a Misunderstood Strategic Partnership?," January 12, 2026, https://www.ainvest.com/news/coreweave-nvidia-circular-financing-red-flag-misunderstood-strategic-partnership-2601/ (noting that CoreWeave rents capacity to hyperscalers including Microsoft and OpenAI).
The Motley Fool, "CoreWeave's $6.3 Billion Backstop Deal With Nvidia: What It Means for Each Company," October 5, 2025, https://www.fool.com/investing/2025/10/05/coreweave-nvidia-6-3-billion-backstop-explained/ (reporting Nvidia's obligation to pay up to $6.3 billion through 2032 for unsold capacity, under an agreement signed in 2023 and disclosed in a securities filing in October 2025).
CoreWeave, Inc., "CoreWeave Reports Strong Third Quarter 2025 Results," press release, November 10, 2025, https://investors.coreweave.com/news/news-details/2025/CoreWeave-Reports-Strong-Third-Quarter-2025-Results/ (reporting a revenue backlog of $55.6 billion as of September 30, 2025); CNBC, "CoreWeave (CRWV) Q3 Earnings Report 2025," November 10, 2025, https://www.cnbc.com/2025/11/10/coreweave-crwv-q3-earnings-report-2025.html (reporting a quarterly net loss of $110 million, narrowed from approximately $360 million a year earlier); DatacenterDynamics, "CoreWeave Q3 Earnings Show Revenue Backlog Doubled to $55.6bn," February 11, 2026, https://www.datacenterdynamics.com/en/news/coreweave-q3-earnings-show-revenue-backlog-doubled-to-556bn/ (reporting third-quarter interest expense of $311 million, up from $104 million a year earlier).
AInvest, "Is CoreWeave-Nvidia Circular Financing a Red Flag or a Misunderstood Strategic Partnership?," January 12, 2026, https://www.ainvest.com/news/coreweave-nvidia-circular-financing-red-flag-misunderstood-strategic-partnership-2601/ (quoting CoreWeave Chief Executive Officer Michael Intrator); AOL, "Nvidia Just Invested Another $2 Billion Into CoreWeave. Time to Buy Again?," https://www.aol.com/finance/nvidia-just-invested-another-2-145316768.html.
Financial Accounting Standards Board, Accounting Standards Codification 606-10-25-1, Revenue from Contracts with Customers (establishing, among the criteria for the existence of a contract, that it be probable the entity will collect substantially all of the consideration to which it will be entitled).
Financial Accounting Standards Board, Accounting Standards Codification 606-10-50-13, Revenue from Contracts with Customers (requiring disclosure of the aggregate transaction price allocated to performance obligations that are unsatisfied or partially unsatisfied at the end of the reporting period).
Financial Accounting Standards Board, Accounting Standards Codification 850-10-50-1, Related Party Disclosures (requiring disclosure of the nature of the relationship, a description of the transactions, and the dollar amounts of transactions for each period for which an income statement is presented).
Financial Accounting Standards Board, Accounting Standards Codification 323-10-15, Investments—Equity Method and Joint Ventures (describing the presumption of significant influence at 20% or more of the voting interest of an investee and recognizing that significant influence may exist at lower levels depending on the facts and circumstances).
Financial Accounting Standards Board, Accounting Standards Codification 460-10, Guarantees (addressing the recognition and disclosure of guarantees, including the nature and term of the guarantee and the maximum potential amount of future payments the guarantor could be required to make).
Financial Accounting Standards Board, Accounting Standards Codification 280-10-50-42, Segment Reporting (requiring disclosure where revenues from a single external customer amount to 10% or more of an entity's revenues, including the total amount and the segment reporting the revenues).
Financial Accounting Standards Board, Accounting Standards Codification 275-10-50, Risks and Uncertainties (requiring disclosure of certain concentrations, including concentration in the volume of business transacted with a particular customer, that expose an entity to the risk of a near-term severe impact).
NVIDIA Corporation, Annual Report on Form 10-K for the fiscal year ended January 25, 2026, as reported by MarketScreener, February 25, 2026, https://www.marketscreener.com/news/nvidia-annual-report-for-fiscal-year-ending-january-25-2026-form-10-k-ce7e5cd8d18af32d (disclosing that, for fiscal year 2026, one direct customer represented 22% of total revenue and another represented 14%, and that, for fiscal year 2025, one direct customer represented 12% and two direct customers each represented 11%).
The Motley Fool, "'Blackwell Sales Are Off the Charts' for Nvidia—and Worryingly, So Is Its Customer Concentration," November 27, 2025, https://www.fool.com/investing/2025/11/27/blackwell-off-charts-nvidia-customer-concentration/ (reporting that, for the fiscal third quarter of 2026, four direct customers each exceeding 10% of revenue—at 22%, 15%, 13%, and 11%—together accounted for approximately 61% of revenue, up from 36% a year earlier).
CNBC, "Nvidia's Top Two Mystery Customers Made Up 39% of the Chipmaker's Q2 Revenue," August 28, 2025, https://www.cnbc.com/2025/08/28/nvidias-top-two-mystery-customers-made-up-39percent-of-its-q2-revenue-.html (reporting that Nvidia identified its largest customers only as "Customer A" and "Customer B," at 23% and 16% of revenue respectively, and referenced an unnamed "AI research and development company" contributing meaningful revenue through both direct and indirect channels).
Munich Personal RePEc Archive, W. Lazonick and E. March, "The Rise and Demise of Lucent Technologies," 2010, https://mpra.ub.uni-muenchen.de/22012/1/lazonick-march_lucent_final_20100410.pdf (reporting that, at the end of fiscal 2000, Lucent had agreements to provide up to $8.1 billion in customer credit or loan guarantees, of which almost $2.1 billion was outstanding); see also Lucent Technologies Inc., Annual Report on Form 10-K405 for fiscal year 2000, U.S. Securities and Exchange Commission, https://www.sec.gov/Archives/edgar/data/0001006240/000095012300011855/y43690e10-k405.txt.
T. Tunguz, "Circular Financing: Does Nvidia's $110B Bet Echo the Telecom Bubble?," October 3, 2025, https://tomtunguz.com/nvidia_nortel_vendor_financing_comparison/ (recounting Lucent's approximately $2 billion vendor-financing commitment to WinStar Communications, its refusal of a final $90 million advance, WinStar's bankruptcy, and a related write-off of approximately $700 million); Munich Personal RePEc Archive, "The Rise and Demise of Lucent Technologies," 2010 (reporting the $700 million write-off).
Munich Personal RePEc Archive, "The Rise and Demise of Lucent Technologies," 2010 (reporting Lucent provisions for customer bad debts of $2.2 billion in 2001 and $1.3 billion in 2002).
The Register, "SEC to Fine Lucent $25m," May 17, 2004, https://www.theregister.com/2004/05/17/sec_fines_lucent/ (reporting civil fraud charges related to the improper recognition of approximately $1 billion in revenue, a $25 million penalty assessed in connection with a failure to cooperate, and charges against former employees); U.S. Securities and Exchange Commission, Litigation Release No. 21551, https://www.sec.gov/litigation/litreleases/2010/lr21551.htm; Computerworld, "Lucent: SEC Examination of Revenue Reporting Not a Surprise," https://www.computerworld.com/article/2591377/lucent--sec-examination-of-revenue-reporting-not-a-surprise.html.
American Affairs, "Who Lost Lucent?: The Decline of America's Telecom Equipment Industry," August 20, 2020, https://americanaffairsjournal.org/2020/08/who-lost-lucent-the-decline-of-americas-telecom-equipment-industry/ (reporting that Nortel Networks had extended approximately $1.4 billion in vendor financing by 2000 and subsequently became absorbed in downsizing, financial restatements, and regulatory matters); T. Tunguz, "Circular Financing," October 3, 2025 (reporting Nortel vendor-financing commitments of up to $3.1 billion with $1.4 billion outstanding).
T. Tunguz, "Circular Financing: Does Nvidia's $110B Bet Echo the Telecom Bubble?," October 3, 2025 (reporting Cisco Systems' approximately $2.4 billion in customer-loan commitments; Cisco's approximately $2.25 billion inventory write-down was recorded in 2001).
T. Tunguz, "Circular Financing," October 3, 2025 (reporting dozens of competitive local exchange carrier bankruptcies between 2000 and 2003 and the abrupt withdrawal of funding between January and April 2001); American Affairs, "Who Lost Lucent?," August 20, 2020.
Bloomberg, "AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other," March 11, 2026, https://www.bloomberg.com/graphics/2026-ai-circular-deals/ (reporting the "virtuous circle" characterization attributed to asset manager Janus Henderson).
CNBC, "Nvidia's Huang Says He's Surprised AMD Offered OpenAI 10% of Company in 'Clever' Deal," October 8, 2025, https://www.cnbc.com/2025/10/08/nvidia-huang-amd-open-ai.html (reporting Jensen Huang's statement that OpenAI did "not have the money yet" and would need to raise it through revenue, equity, or debt).
AOL, "Why OpenAI's $300 Billion Deal With Oracle Has Set the 'AI Bubble' Alarm Bells Ringing," 2025, https://www.aol.com/articles/why-openai-300-billion-deal-150815932.html (citing a Massachusetts Institute of Technology study reporting that approximately 95% of enterprise generative-AI pilots failed to deliver meaningful returns).
This article is published by Buxton Helmsley USA, Inc. for informational and educational purposes only. It does not constitute investment advice, a solicitation, or a recommendation to buy or sell any security. The views expressed are those of Buxton Helmsley and are based on publicly available information as of the date of publication. Investors should conduct their own due diligence and consult with qualified professionals before making investment decisions.
Briefings on markets, governance, and our investment philosophy—delivered directly.
