The Loop: How the AI Industry Learned to Fund Its Own Demand

On July 27th of this year, the Wall Street Journal reported that Nvidia was negotiating a guarantee of up to $250 billion to backstop OpenAI's data center campus in Ohio. Bloomberg and Reuters confirmed the figure over the same weekend. The number moved markets, generated a week of commentary about whether the AI industry had finally lost its mind, and became the reference point for every conversation about circular financing for the next three weeks. Then, on August 17th, Nvidia filed a Form 8-K with the Securities and Exchange Commission describing what it had actually signed, and the instrument turned out to be something rather different. Not a $250 billion guarantee. A set of residual value guaranties with SB Energy Corp, covering leases for approximately 4.25 gigawatts of IT load at the Portsmouth site in Pike County, with Nvidia's aggregate payment obligation cumulatively capped at $105 billion for the initial commitment. Payments conditional on OpenAI defaulting or becoming insolvent. Effective only as individual leases commence, expected to begin phasing in from 2028. And crucially, if triggered, Nvidia holds a menu of remedies including assuming the lease, requiring SB Energy to find a replacement tenant, initiating a sale, or deferring for up to a year, with OpenAI contractually obligated to reimburse Nvidia for any amounts actually paid.

I want to start here, with the gap between the reported number and the filed instrument, because it is the single most useful thing I can tell you about this entire subject. The AI financing story is being narrated almost entirely in headline figures, and headline figures in this space are not a reliable guide to what anyone has actually agreed to do. A $250 billion unconditional guarantee and a $105 billion conditional residual value guaranty are different animals with different risk profiles, different accounting treatment, and different implications for every counterparty in the chain. The first would be a genuinely staggering assumption of credit risk. The second is closer to what a sophisticated lessor would demand from a strategically motivated third party in any large sale-leaseback, and while the number is enormous, the probability-weighted exposure is a fraction of the ceiling. Most people reading about this deal will remember $250 billion. Almost none of them will read the 8-K. That asymmetry, repeated across dozens of deals over eighteen months, is how a genuinely important structural question has become nearly impossible to reason about from the outside.

So let me try to do it properly. What follows is an attempt to map what people are calling the AI circular financing loop using primary sources wherever they exist, to separate the parts that are genuinely circular from the parts that merely look circular, to take the historical analogy seriously rather than reflexively, and to identify the one measurement that actually settles the question. I have spent most of my career thinking about counterparty risk and position sizing, and the instinct I have developed is that when a structure becomes too complicated to explain simply, the complexity is usually doing work for someone. That is not always sinister. Sometimes complexity is just what happens when you finance something genuinely novel. But it is always worth asking who benefits from the confusion.

Start with the mechanics, because the basic shape is simple even if the instances are not. Circular financing, in the sense people mean it here, describes an arrangement where a supplier invests capital in a customer, and the customer uses that capital to buy the supplier's products. Nvidia invests in an AI lab, the lab buys Nvidia GPUs. Microsoft invests in an AI lab, the lab commits to buy Azure capacity. A cloud provider takes equity in a model developer, the developer signs a multi-year compute contract with that same provider. The money describes a circle, and the concern is that this circle can manufacture the appearance of demand. If I give you a dollar and you hand it back to me in exchange for a widget, I have booked a dollar of revenue and you have booked a widget, and nothing has happened economically except that I now own a claim on you and you own a depreciating asset. Repeat this at sufficient scale and you can produce revenue growth, backlog, and market capitalization that reflect the velocity of money inside a closed system rather than any demand from outside it.

It is worth being precise about a distinction that gets collapsed constantly in commentary. Circular financing is not the same thing as round-tripping. Round-tripping, in the regulatory sense, refers to sham transactions with no economic substance constructed specifically to inflate reported revenue. Two companies agree to buy equivalent amounts of nothing much from each other, both book revenue, no net cash moves, and the financial statements lie. That is fraud, and it is what several telecom operators did with fiber capacity swaps in 2000 and 2001. Circular financing is generally not that. When Nvidia sells a GPU to a company it has invested in, the GPU is real, it is delivered, it is racked, it draws power, and it computes. The transaction has economic substance. The question is not whether the sale happened but whether it would have happened absent the investment, and whether the end demand exists to make the whole chain solvent over time. That is a much harder question than fraud, and it does not have a clean answer, which is precisely why it has generated so much argument.

Now the map. In September of 2025, Nvidia announced it would invest up to $100 billion in OpenAI as part of a strategic partnership under which OpenAI would deploy at least ten gigawatts of Nvidia systems. The announcement was universally reported as a deal. It was, in fact, a letter of intent, structured to fund progressively as each gigawatt was deployed, and it never became a binding contract. On January 30th of this year, the Wall Street Journal reported that the plan had stalled after people inside Nvidia expressed doubts, and Reuters picked up the report the same day. Jensen Huang subsequently stated publicly that the $100 billion figure was never a commitment, and later that the full amount was probably not in the cards. What Nvidia actually did was contribute $30 billion to the funding round OpenAI closed in March, a round that raised $122 billion at a valuation of roughly $852 billion. Huang further indicated it might be the last time Nvidia invests in OpenAI before the company goes public.

Trace that arc carefully, because it is instructive. A $100 billion announcement became a $30 billion equity check, and the intervening period was filled with commentary treating the larger number as a fact about the world. This is not deception on anyone's part. A letter of intent is a real thing, disclosed as such, and Nvidia's securities filings had flagged the conditionality. But the market received it as a commitment, priced it as a commitment, and built a narrative about circular financing on top of a number that never existed as a binding obligation. Then in July the $250 billion backstop was reported, and in August the actual filed instrument came in at $105 billion of conditional exposure. Twice now, the gap between the reported figure and the executed document has been substantial, and in both cases the direction of the error was the same.

The AMD arrangement is more interesting because it is genuinely novel and because it is fully documented in a public filing. In October 2025, AMD entered a product purchase agreement with OpenAI OpCo, LLC to deploy six gigawatts of AMD GPUs, beginning with the Instinct MI450 series. Concurrently, and this is the part with no clean precedent, AMD issued OpenAI a warrant to purchase up to 160 million shares of AMD common stock at an exercise price of one cent per share. That is roughly ten percent of the company. The warrant vests in tranches tied to GPU purchase milestones by OpenAI, achievement of specified AMD stock price targets, and satisfaction of further technical and commercial conditions before exercise. The final tranche is tied to AMD reaching $600 per share. The warrant is exercisable through October 5, 2030. All of this is in AMD's 10-K, not in a press release, which means it is audited language rather than marketing.

Think about what that structure actually does. AMD is paying its customer, in equity, to become its customer. The stated logic is alignment: if OpenAI deploys six gigawatts of AMD silicon and AMD's stock triples as a result, both sides win, and OpenAI has a direct financial interest in AMD's success as a second source against Nvidia. That logic is coherent. But look at the vesting condition tied to AMD's share price, because it creates a reflexive loop that is unusual even by the standards of this industry. The warrant becomes valuable if AMD's stock rises. AMD's stock rises partly because of the announced OpenAI relationship. On the day the partnership was disclosed, AMD's shares jumped and the company added roughly $80 billion in market value. The customer's equity compensation is therefore partly a function of the market's reaction to the announcement of the customer relationship. I do not know of a clean historical analogue for that, and I have looked. It is not fraudulent, it is disclosed, and the vesting conditions mean OpenAI could end up buying GPUs and receiving nothing if the milestones are missed. But it is a genuinely new kind of instrument, and new instruments in capital-intensive booms have a history worth respecting.

The Anthropic side of the map has become, if anything, more entangled than OpenAI's. In November 2025, Microsoft and Nvidia announced investments of up to $5 billion and up to $10 billion respectively, alongside Anthropic's commitment to purchase $30 billion of Azure compute capacity and to contract additional capacity up to one gigawatt. Microsoft's own blog post is the primary source and states the terms plainly. Then in April of this year the pace accelerated sharply. On April 20th, Amazon announced a further $5 billion investment with up to $20 billion more tied to commercial milestones, alongside Anthropic committing to spend over $100 billion on AWS technologies over ten years. Four days later, on April 24th, Google confirmed an investment of up to $40 billion, with $10 billion immediate and the remaining $30 billion contingent on performance milestones, tied to gigawatts of TPU capacity. The result is that Microsoft, Nvidia, Amazon, and Google are all simultaneously on the cap table of a company that competes directly with products each of them sells, and each of those investments is paired with a compute purchase commitment flowing back to the investor.

Microsoft's own relationship with OpenAI was restructured in October 2025 in a way that is worth understanding because it changed the shape of the whole network. Following the recapitalization, Microsoft holds an investment in OpenAI Group PBC valued at approximately $135 billion, representing roughly 27 percent on an as-converted diluted basis, down from a 32.5 percent stake excluding the impact of recent funding rounds. OpenAI contracted to purchase an incremental $250 billion of Azure services. In exchange, Microsoft gave up its right of first refusal to be OpenAI's compute provider, which is precisely what enabled the Oracle, Amazon, and Broadcom arrangements that followed. Microsoft's IP rights extend through 2032. OpenAI's own commitments to Oracle exceed $300 billion over five years for up to 4.5 gigawatts of Stargate capacity, per OpenAI's own announcement, and Oracle's remaining performance obligations grew to $455 billion at the end of its fiscal first quarter 2026 and to roughly $523 billion by the third quarter.

So that is the picture at the model layer: a small number of suppliers holding equity in a small number of labs, with enormous purchase commitments running in the opposite direction. Now consider the layer beneath it, because this is where the structure stops being merely circular and starts being leveraged, and leverage is what turns an uncomfortable structure into a systemic one.

The neoclouds are specialized providers that do essentially one thing: buy GPUs and rent them out. CoreWeave is the archetype. In August 2023, it borrowed $2.3 billion in a facility led by Magnetar Capital and Blackstone, using Nvidia H100 processors as collateral, which Reuters reported at the time as the first use of H100-based hardware as security for a loan. That structure became the template. By 2024 CoreWeave's total debt was under $8 billion. By May of this year it exceeded $21 billion, and by August, sector reporting put it around $35 billion. The financing is typically done through special purpose vehicles that pledge the GPUs as collateral alongside the customer contracts, and the critical feature is that the credit quality of the paper derives not from the neocloud's own balance sheet, which carries a speculative-grade rating, but from the creditworthiness of the customer on the other side of the contract. CoreWeave's $8.5 billion delayed-draw term loan facility was rated A3 by Moody's and A (low) by DBRS on the strength of a customer contract, not on the strength of CoreWeave.

Follow the chain all the way through and you get something that should give any credit analyst pause. Nvidia invests equity in a neocloud. The neocloud uses that equity, plus debt raised against Nvidia GPUs as collateral, to buy more Nvidia GPUs. The debt is rated investment grade because a hyperscaler has signed an offtake contract. The hyperscaler is itself spending a historically unprecedented share of operating cash flow on AI infrastructure. And the collateral underlying the whole structure is a depreciating asset whose residual value depends on the pace of the product cycle set by the same company that provided the initial equity. Every link in that chain is individually defensible. The aggregate is a structure where the same handful of balance sheets appear at multiple points, and where a correlated shock does not have many places to be absorbed.

Nvidia's response to this, beginning around July 1st of this year, has been to formalize its role rather than retreat from it. The company introduced a backstop program under which it provides take-or-pay commitments to neoclouds, guaranteeing a minimum revenue floor on deployed GPU capacity in exchange for a share of revenue earned above that floor. The stated purpose is to break a circular dependency of a different kind: building a cluster requires capital, an offtake contract, and datacenter space, and each of those typically requires proof of the other two, which locks out everyone except the largest players. An investment-grade guarantee from Nvidia lets a smaller operator secure debt, which lets it place equipment deposits, which lets it sign customers. SemiAnalysis, which has covered this most closely, has described Nvidia in this role as functioning as something like a central bank for the AI sector, supplying liquidity where private lenders will not yet go. Then on August 10th, Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build compute financing platforms targeting more than $500 billion of third-party capital, and Huang subsequently addressed the circularity criticism directly, indicating Nvidia may provide residual value support for up to 25 percent of a given opportunity, assessed project by project.

Whether you find that reassuring depends on what you think the constraint actually is. If the binding constraint on AI buildout is that lenders lack the tooling to price GPU residual value, then an investment-grade party with better information stepping in to underwrite that specific risk is genuinely useful market-making, and the spread between roughly 5.9 percent on backstopped hyperscaler-contracted paper and roughly 10 percent on CoreWeave's unsecured bonds represents real value being created by transferring risk to whoever can bear it most cheaply. If the binding constraint is that end demand does not yet justify the buildout, then the same mechanism is a way of extending the boom past the point where private capital would have stopped it, and every dollar of additional deployment increases the eventual adjustment. Both readings are consistent with the observable facts today. That is the uncomfortable part.

Which brings me to the historical parallel, invoked so often it has become a reflex, and which deserves better treatment than it usually gets. Between roughly 1998 and 2001, telecom equipment manufacturers extended enormous vendor financing to the carriers that bought their gear. Lucent committed approximately $8.1 billion. Nortel extended around $3.1 billion with $1.4 billion outstanding. Cisco promised roughly $2.4 billion in customer loans. The logic was identical to today's: lend money to capital-constrained customers so they can buy your equipment, book the revenue, and everybody grows. Lucent's revenue peaked at $37.92 billion in 1999. By 2002 it had fallen to $11.80 billion, a decline of roughly 69 percent, and the company never recovered as an independent entity, merging with Alcatel in 2006. When carriers began failing, 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. Its exposure to Winstar Communications alone, a single customer, ran into the hundreds of millions when Winstar went bankrupt in April 2001. Between 2000 and 2002, global telecom stocks lost more than $2 trillion in market value.

And then there was the accounting. On May 17, 2004, the SEC charged Lucent with securities fraud, alleging the company had fraudulently and improperly recognized approximately $1.148 billion of revenue and $470 million in pre-tax income during fiscal year 2000. The complaint described falsified documents, undisclosed side agreements with customers, and circumvention of internal controls. Lucent settled without admitting or denying, paying a $25 million penalty that the SEC explicitly stated was for lack of cooperation with the investigation rather than for the underlying conduct. Nine current and former Lucent employees were charged, along with a former Winstar officer. That is the primary source, SEC Press Release 2004-67, and it remains the cleanest available account of what happens when vendor financing and revenue recognition pressure operate on the same set of books.

Now, what actually rhymes and what does not. The rhyme is real. In both cases, a supplier facing a customer base that could not independently fund purchases at the desired scale chose to supply the funding, in both cases the resulting revenue was reported as ordinary revenue, and in both cases the concentration of counterparties meant that a demand disappointment at one node would propagate rather than be absorbed. The most important structural similarity is the one nobody likes to state plainly: the suppliers had a strong incentive to keep the customers buying regardless of whether the customers should have been buying, because the suppliers' own valuations depended on the customers' purchases continuing.

But the differences are substantial and they cut in the direction of today being more robust, not less. The 1999 vendor financing was extended overwhelmingly through debt, often unsecured, frequently to companies that had already been declined by banks, and by vendors whose own balance sheets could not absorb the losses when they came. Nvidia today is funding its equity investments from operating cash flow, not from borrowings, and the hyperscalers at the center of the network are among the most profitable enterprises in history with real revenue from real external customers. The 2001 carriers were building capacity for demand that was speculative, and the famous statistic about internet traffic doubling every hundred days turned out to be largely fabricated. Today's AI demand is measurable: ChatGPT reports more than 900 million weekly active users, enterprise API consumption is billed monthly, and the revenue trajectories at both major labs have been extraordinary by any historical standard. Anthropic's annualized revenue reportedly grew roughly thirty-fold over about fifteen months. Whether that demand is durable at the price points required to service the infrastructure is exactly the open question, but it is not fabricated, and that is a meaningful distinction from 2000.

There is a second-order question that has generated more heat than any other, and it belongs here because it determines whether the reported profitability of the whole structure is real. In November 2025, Michael Burry began publicly arguing that hyperscalers are inflating earnings by depreciating Nvidia-based hardware over five or six years when the true economic life, against a chip cycle that renews every year or two, is closer to two or three. He put the aggregate understatement at roughly $176 billion between 2026 and 2028, suggested reported operating income at some companies could be materially overstated as a result, and characterized the practice on X as one of the more common frauds of the modern era. He backed the view with put positions on Nvidia and Palantir.

I think the strong form of that claim is wrong and the weak form is important, and the distinction matters. The strong form treats a longer useful life as prima facie fraudulent, which does not survive contact with how these assets actually get used. GPUs cascade. A chip that is no longer economical for frontier training gets demoted to inference, then to smaller model serving, then to general compute, and it can generate revenue in each of those roles. Five-year-old A100s are still being rented today, which is the empirical rebuttal, and Nvidia has argued that observed customer utilization supports four-to-six-year lives. Useful life estimates are audited, defended with utilization data, and the companies keep passing. The weak form, though, is a genuine concern and worth holding onto: depreciation policy is an estimate, estimates in this industry have been drifting longer for years, and the direction of drift has been convenient. It is telling that in 2025 Amazon shortened the useful life of a subset of servers while Meta extended its estimate further. When two sophisticated companies with access to the same technological facts move their assumptions in opposite directions, at least one of them is making a judgment call that could be wrong, and the earnings impact of being wrong is measured in tens of billions.

The concentration numbers deserve a paragraph of their own because they are the clearest quantitative expression of how narrow this structure has become. In fiscal years 2022 and 2023, no single customer accounted for ten percent or more of Nvidia's revenue. By fiscal 2024, one unnamed customer was at 13 percent. In the second quarter of fiscal 2026, Customer A was 23 percent and Customer B was 16 percent, meaning two unnamed entities represented 39 percent of revenue, with four more customers at 14, 11, 11 and 10 percent. By the third quarter of fiscal 2026, four direct customers each exceeding ten percent collectively represented 61 percent of revenue, against 36 percent for the comparable prior-year quarter. Nvidia's own filing language notes that it has experienced periods of receiving significant revenue from a limited number of customers and that the trend may continue. That is a company whose revenue base has concentrated from effectively diversified to majority-dependent on four counterparties in roughly two years, during a period in which it was also taking equity positions in, and extending credit support to, entities in that same customer set.

What moved this from an analyst debate to something more serious was the Bank for International Settlements. On June 28th of this year, the BIS published its Annual Economic Report, and it named the AI capex boom and its financing structures among the pressure points demanding immediate policy attention, alongside inflation risk and fiscal fragility. The report notes that the five largest hyperscalers are set to spend more than a trillion dollars on AI-related capital expenditure across 2025 and 2026, that these commitments are outpacing earnings and free cash flow, and that some firms are issuing debt as a result. It flags circular financing specifically, observing that the terms of such deals are typically poorly disclosed and that the same asset may be pledged more than once. A companion working paper, BIS Working Paper No. 1367, models the buildout as a race and estimates over-investment at roughly 1.5 times the efficient level, rising toward three times where demand is less elastic. When the institution that serves as the central bank for central banks puts a structure on its financial stability register, the question has stopped being whether commentators find it aesthetically troubling.

It is worth noting, for balance, that the BIS itself and the IMF have both assessed the direct systemic risk as moderate for now, on the reasoning that the primary spenders are profitable firms with real customers and strong cash generation, and that aggregate business and household leverage relative to GDP is nothing like 2007. That is a fair read of the current state. But the thing about a structure being robust because its participants are currently profitable is that the profitability is the variable under examination.

So here is the test, and I think it is the only one that matters. Every argument about whether this is a healthy financing innovation or an elaborate way of moving the same money in a circle reduces to a single measurable quantity: what fraction of the revenue in this system originates from outside the system. Not from a supplier who invested in you. Not from a lab that took your equity. From a hospital paying for clinical documentation, a bank paying for research automation, a law firm paying for discovery, a manufacturer paying for defect detection, an individual paying twenty dollars a month. That is the number. If external revenue grows fast enough to service the infrastructure being built for it, then every circular arrangement described above was a sensible piece of financial engineering that pulled forward capacity the world genuinely needed, and the vendor financing will look, in retrospect, like the railroads and the fiber build, which is to say wasteful at the margin and transformative in aggregate. If it does not, then the circularity has been functioning as a mechanism for deferring the recognition of a demand shortfall, and the deferral will have made the eventual adjustment larger.

What makes this genuinely hard to assess from outside is that the disclosure does not decompose that way. Nvidia does not report revenue split by whether the customer received Nvidia capital. Oracle's remaining performance obligations do not distinguish contracts backed by end demand from contracts backed by a lab's expectation of future end demand. The hyperscalers report cloud revenue without disaggregating AI-lab compute purchases from enterprise workloads. There is no line item anywhere that answers the question. The BIS point about poor disclosure is not a complaint about opacity for its own sake, it is an observation that the single most decision-relevant number in a trillion-dollar buildout is not published by anyone.

The IPO filings may change that, which is the most interesting near-term development in this whole story. Anthropic submitted a confidential draft S-1 to the SEC on June 1st at a reported valuation of $965 billion. OpenAI followed a week later, announcing on June 8th via its own blog under Rule 135 that it had submitted a confidential S-1, noting it expected the news to leak and had not decided on timing. Both companies are therefore in nonpublic SEC review. Whichever files publicly first will produce the first genuine prospectus for a frontier AI lab, and a prospectus has to disclose things a press release does not: cost of revenue, gross margin on inference, the actual terms of compute commitments, customer concentration, related party transactions, and the risk factors a securities lawyer insists on. For the first time, someone will have to write down, under liability, what the economics actually are. I would rather read one of those documents than a hundred more headlines about billion-dollar partnerships.

If I were sizing exposure to any part of this structure, and I want to be careful here because I am not offering anyone investment advice and I have no idea what your situation is, the framework I would use has three questions and they apply to every headline figure in the sector. Is this committed capital, a guarantee, or an option? What specifically triggers payment? And who pays, out of what source of funds? Applied to the Nvidia Ohio arrangement, the answers are: a guarantee, an OpenAI default, and OpenAI itself out of its own revenue, with Nvidia standing behind it and holding reimbursement rights. Applied to the AMD warrant: an option, a combination of deployment milestones and share price targets, and AMD's existing shareholders through dilution. Applied to the original $100 billion Nvidia announcement: none of the above, because it was a letter of intent that never became binding and ultimately resolved into a $30 billion equity check. Those three questions dissolve most of the confusion in this sector, and almost nobody asks them because the headline number is more interesting than the instrument.

The thing I keep returning to is that the participants themselves are behaving as though they take the risk seriously, which is more informative than any of the public rhetoric. Nvidia walked away from a $100 billion letter of intent when internal doubts surfaced, which is not the behavior of a company indifferent to its customer's discipline. Huang has said publicly that the March round may have been the last Nvidia investment in OpenAI before a listing. The Ohio guarantee came in at 42 percent of the reported figure and with conditionality that reporting had not captured. Amazon's and Google's Anthropic investments are structured with the large majority contingent on milestones rather than paid upfront. AMD's warrant requires its own stock to triple before the final tranche vests. Across the board, the executed instruments are more conditional, more milestone-dependent, and smaller than the announcements that preceded them. You can read that as prudent structuring by sophisticated parties who understand exactly what they are underwriting. You can also read it as a set of counterparties who very much want the option to walk away, retaining flexibility precisely because they are not certain the demand will arrive on schedule. Those readings are not mutually exclusive, and I suspect both are true.

The 2001 lesson was not that vendor financing is inherently corrupt. Railroads were vendor-financed. So was fiber, and the fiber that bankrupted its builders is still carrying traffic today at a cost that made the modern internet possible. Capital-intensive industries with long build cycles and uncertain early demand have always required someone to bridge the gap between what customers can fund and what the technology requires, and the people who provide that bridge have often been the suppliers, because they have the best information and the strongest incentive. The lesson is narrower and more practical: vendor financing becomes dangerous at the exact moment when it stops being a bridge to external demand and starts being a substitute for it, and the participants are structurally the last people to notice the transition, because their own reported numbers look fine right up until they do not. Lucent did not know it was extending credit to companies that would not survive. It thought it was winning market share. The revenue was real, the shipments were real, and the receivables were real until they were written off.

We will find out which of these we are in, and probably sooner than most people expect. The first public S-1 from a frontier lab will disclose more than the last two years of press releases combined. Nvidia's 10-Q for the quarter ended July 26th will include the full form of the Ohio guaranty agreements as an exhibit, which will tell us considerably more about the trigger mechanics than the 8-K summary does. The neocloud refinancing wall between 2026 and 2028 will test whether GPU residual values hold across hardware generations, and it will do so with real money at real spreads. And the depreciation question resolves itself mechanically over the next several reporting cycles, because either the assets keep generating revenue over five and six year lives or they do not, and the impairments show up either way. None of this requires anyone to be right about the future. It requires only patience and a willingness to read the filings rather than the headlines, which is, as far as I can tell, the entire edge available in this sector at the moment.

References

NVIDIA Corporation. Form 8-K, filed August 17, 2026. Residual value guaranties with SB Energy Corp. relating to the PORTS Technology Campus, Pike County, Ohio. SEC EDGAR, accession 000104581026000069.

Advanced Micro Devices, Inc. Form 10-K for fiscal year 2025. Disclosure of the October 2025 OpenAI OpCo, LLC product purchase agreement and the 160 million share warrant at $0.01 exercise price, exercisable through October 5, 2030. SEC EDGAR.

AMD. "AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs." Investor Relations press release, October 6, 2025.

Microsoft Corporation. Form 8-K, October 2025, exhibit describing the revised OpenAI partnership terms. SEC EDGAR.

Microsoft. "The next chapter of the Microsoft-OpenAI partnership." Official Microsoft Blog, October 28, 2025.

Microsoft. "Microsoft, NVIDIA and Anthropic announce strategic partnerships." Official Microsoft Blog, November 18, 2025.

OpenAI. "OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites." Company announcement, 2025.

OpenAI. "Confidential submission of draft S-1 to the SEC." Company announcement under Rule 135, June 8, 2026.

U.S. Securities and Exchange Commission. "Lucent Settles SEC Enforcement Action Charging the Company with $1.1 Billion Accounting Fraud." Press Release 2004-67, May 17, 2004.

Bank for International Settlements. Annual Economic Report 2026, Chapter I, "Progress and peril." Published June 28, 2026.

Bank for International Settlements. Working Paper No. 1367, "The AI investment race." July 2026.

Reuters. "Nvidia's plan to invest up to $100 billion in OpenAI has stalled, WSJ reports." January 30, 2026.

CNBC. "Nvidia and OpenAI in talks for up to $250 billion backstop to fund AI infrastructure plans." July 27, 2026.

CNBC. "OpenAI completes restructure, solidifying Microsoft as a major shareholder." October 28, 2025.

CNBC. "Anthropic valued in range of $350 billion following investment deal with Microsoft, Nvidia." November 18, 2025.

CNBC. "Google to invest up to $40 billion in Anthropic as search giant spreads its AI bets." April 24, 2026.

CNBC. "OpenAI confidentially files for IPO, prepping Wall Street for mega AI debut." June 8, 2026.

CNBC. "Nvidia's top two mystery customers made up 39% of the chipmaker's Q2 revenue." August 28, 2025.

Bloomberg. "Microsoft, Nvidia to Invest Up to $15 Billion in Anthropic." November 18, 2025.

Bloomberg. "AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other." Graphics feature, January 2026.

Axios. "Google's $40B Anthropic move is Big Tech's latest huge AI bet." April 24, 2026.

Axios. "Anthropic lands $15 billion investment from Microsoft, Nvidia." November 18, 2025.

Quartz. "GPU-collateralized debt explained: AI financing risks." May 2026. On CoreWeave's 2023 Magnetar and Blackstone facility and the DDTL 4.0 ratings.

SemiAnalysis. "Nvidia GPU Debt Backstop Unleashes the AI Project Trinity: Capital, Offtake and Datacenters." July 2026.

Irwin-adjacent historical sourcing on vendor financing: Princeton, "The Great Telecom Implosion," Paul Starr, September 2002; and contemporaneous reporting on Lucent, Nortel and Cisco customer financing commitments during 1999 to 2001.

Tomasz Tunguz. "Circular Financing: Does Nvidia's Bet Echo the Telecom Bubble?" October 2025. Source for the Lucent, Nortel and Cisco vendor financing commitment figures and Lucent revenue peak and decline.

Michael Burry, posts on X, November 2025, on hyperscaler depreciation and useful life assumptions.

Federal Reserve and IMF assessments of AI-related financial stability risk, 2026, as summarized in contemporaneous reporting.

Stanford Institute for Economic Policy Research, Brookings Institution, and Yale Budget Lab commentary on AI capital expenditure and macroeconomic effects, 2026.