We have been here before, in a less connected market, and the legends persisted for centuries.

There is a word sitting quietly in the press release Nvidia published on the tenth of August, and the word is offtakers. The term belongs to the paperwork of power plants and copper mines, to the long-term purchase agreements that let a lender believe in a hole in the ground before the ore comes out of it, and the value that ore will have once the hole is sufficiently expanded and exploited. Nvidia used it to describe the laboratories that buy its chips. The release announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR – six houses that among them manage sums measured in trillions – to establish financing platforms intended to mobilize more than 500 billion of third-party capital for AI infrastructure, and it described the product at the center of the arrangement as “an investable asset” (NVIDIA, 2026). Jensen Huang added, on X, that Nvidia retains the option to backstop up to 125 billion of whatever gets written, roughly a quarter of the potential book (Reuters, 2026). This chipmaker super company, who spent July absorbing criticism for financing its own customers, doubled down weeks later, to declare their platforms a hole in which to put money, and to invite the largest pools of capital on the planet to come put it there, on the prediction that what comes out will be worth more, once everything is expanded and exploited.

The last few weeks produced a second kind of blip for those who know history, quieter and stranger. Since the second of August, under Article 50 of the EU’s AI Act, providers of generative systems operating in Europe must mark synthetic output in machine-readable form and label their deepfakes, so that a person encountering the artificial can know it for what it is (European Commission, 2026). Eighteen days later, Apple Music told its industry partners that songs built from “a material portion of the content” generated by AI will carry a visible Made With AI label by year’s end, converting the voluntary disclosure tags of March into a listener-facing mark (Variety, 2026). And peer-reviewed work has already measured what such a mark does. When the phrase “artificial intelligence” appears in a product description, emotional trust drops, and purchase intention drops with it, most sharply for the goods people are most afraid to get wrong (Cicek, Gursoy, & Lu, 2025).

Hold these two developments in the market, around the same moment in time, side by side. One august institution stamps the chip as extractable investment. Another set of institutions stamps the chip’s output as indelible, with clear market research on what that means to salable product. The same civilization, in the same month, is underwriting a thing and warning itself about it.

Somebody will say tulips. Somebody always says tulips, and this essay is going to jump headfirst into the tulip parallel, because the flower earns its place properly, once you stop misremembering it.

The financing of machine intelligence has revived vendor finance, the arrangement by which a supplier funds its customers’ purchases of its own product – a structure last run at national scale by the telecom equipment makers between 1998 and 2001 and adjudicated afterward, in writing, by the Securities and Exchange Commission. Stacked on top of that structure sits a valuation that depends on a good, future intelligence, whose worth no party inside the loop can independently price, so the parties price it for each other. Running underneath both these stories, older than either, is the thing the Dutch actually lost in the winter of 1637, the loss Anne Goldgar recovered from the notarial archives after the drowned speculators of legend turned out never to have drowned: “a whole network of values was thrown into doubt” (Goldgar, 2007).

We are in a bubble. The productive question is not whether it pops. The productive question is what breaks when it does, and who is standing inside the circle of ruin and distrust when the vouching stops.

The Flower, Corrected

The tulip story you might know arrived two centuries late. Charles Mackay published Extraordinary Popular Delusions and the Madness of Crowds in 1841, and from him descends nearly every vivid particular in the popular account: the chimney sweeps and maidservants bidding on bulbs, the merchant houses ruined beyond redemption, the grief-mad burghers going into the canals. Mackay built his popular account largely out of songs and pamphlets printed in 1637 to mock the newly rich – satire he mined as though it were a court record. As Goldgar puts it, later writers who relied on him amount to virtually everybody (Roos, 2020).

She went where Mackay never did, into the notarial registers and small-claims files of Haarlem, Amsterdam, Alkmaar, and Enkhuizen, and the legend dissolved on evidentiary exploration. Goldgar identified roughly 350 people meaningfully involved in the trade and its collapse – a closed set bound to one another through profession, family, and congregation. She found 37 of them who ever paid more than 300 guilders (somewhere between 18,000–20,000 of today’s money) for a single tulip bulb, about a skilled craftsman’s year. She found receipts topping out near 5,000 guilders, the price of a decent house, and she found them to be outliers. Prices for the coveted varieties did spike twelvefold between December 1636 and February 1637, and then they fell out of the sky.

What she could not find, anywhere, was a single bankruptcy caused by the collapse. The painter Jan van Goyen, the crash’s most famous supposed casualty, appears to have been undone by land speculation. The drowned Dutchman is a character invented so that later centuries could feel clever, or, at least, more clever than yesteryear’s floral investors.

Goldgar understands the much-mistold legend’s durability better than anyone: it survives, she says, because “it makes people look stupid” (Roos, 2020).

So the skeptic of the present panic has a fair opening. If the canonical bubble is mostly a sermon, perhaps the bubble-callers of 2026 are preaching from the same pulpit. Goldgar herself, retelling the research for a general audience during the crypto winter, closed with a warning against fitting each new mania for seventeenth-century Dutch clothes (Goldgar, 2018).

And the skeptic has a second, sharper witness – the man at the center of the loop. Asked directly about the circular-financing charge, Huang’s answer ran five words: “OpenAI will pay the lease” (Yahoo Finance, 2026). His fuller case deserves a fair hearing. Frontier laboratories are growing faster than their balance sheets and credit profiles can carry, and when an industry’s ambition outruns its collateral, somebody has to bridge the gap. That is what capital formation is for. The hardware is real, the data centers are being built, and the chipmaker is merely providing the bridge financing that would normally come from banks – except that banks don’t yet understand how to value GPU-backed collateral, so the chipmaker has to become the bank.

This is not a bad argument. It is, in fact, the argument a rational actor makes when standing at the center of a self-reinforcing loop. The scaffolding is ugly and temporary, yes – but buildings go up inside scaffolding every day, and the scaffolding comes down when the structure is self-supporting. The question is whether this building has a foundation, or whether the foundation is also scaffolding.

Notice what Goldgar’s archives actually yielded once the fictional corpses were cleared away. Court records full of promises broken. Buyers who had pledged a hundred or a thousand guilders in a tavern, before witnesses, and then refused to pay when the price of a summer flower gyrated through a winter. Reputations cracked in an economy that ran on credit and honor, in a merchant culture where a man’s signature and his standing were essentially the same instrument. Modern credit ratings are mutable – you can get a credit card a handful of years after bankruptcy and insolvency. Those times were different.

Goldgar calls the aftermath a shock to the culture rather than to production, a season in which the meaning of value itself came unfixed, because if a bulb could be worth a house in January and an onion in February, then the whole apparatus by which the Dutch knew what things were worth stood exposed as a flimsy agreement, which, apparently, people could walk away from (Goldgar, 2007).

This insight is the bridge to the present. The tulip trade’s collapse was not about the bulbs being worthless – it was about the network of mutual assurance that priced them shattering. That is the exact structure now at work in the compute economy, and the Securities and Exchange Commission wrote it down for us in New Jersey, not Haarlem.

The tulip tale is a fable which would never afterward stay buried. Satirists reached for it when the South Sea and Mississippi schemes broke in 1720. Beckmann reached for it in the 1780s. Galbraith, a serious economist writing in the shadow of the 1987 crash, repeated Mackay’s inventions as history in a book about how euphoria forgets (Galbraith, 1990). Four centuries of borrowers keep taking the story out on loan. A legend that useful is telling you something true about the fear it services, whatever it gets wrong about Haarlem.

What Lucent Knew

For the now, let us leave the taverns of the Netherlands, and go to New Jersey.

The Telecommunications Act of 1996 opened the American telecommunications network to competitive entry, and capital arriving to that market showed up the way water arrives through a failed levee. Paul Starr, writing the autopsy of the whole affair in September 2002, recorded an industry insider’s memory that the business plans of the new carriers all looked alike – massively parallel systems rising in parallel, each priced on forecasts of internet traffic that treated the future as already invoiced (Starr, 2002). When the traffic failed to arrive on schedule, some executives improved the arithmetic.

Starr describes the era’s signature maneuver with a bookkeeper’s calm: two companies holding excess capacity would sell each other rights to use portions of each other’s networks, each booking, say, 150 million in revenue on the exchange, though no real revenue existed, leaving an industry that appeared 300 million larger than it was. Two firms, one handshake, zero customers, three hundred million.

The equipment makers ran their own version. Lucent committed on the order of 8.1 billion in vendor financing, Nortel roughly 3.1 billion, Cisco about $2.4 billion – loans extended to carriers, disproportionately the thin new competitive local exchange carriers, so those carriers could buy the lenders’ switches and optical gear (Tunguz, 2025). A vendor loan performs a specific piece of alchemy. It converts a sale into an act of faith while recording it as fact, revenue booked today against a receivable that exists only if the borrower’s projected future arrives on time. The receivable is the belief. When the futures stopped arriving, twenty-three telecom companies went into bankruptcy in a wave capped by WorldCom, at that point the largest failure in American history; half a million jobs went with them; roughly two trillion dollars of market capitalization evaporated from the sector; and the industry sat under a trillion dollars of debt that the chairman of the FCC told the Senate would largely never be repaid (Starr, 2002). Not so dissimilar to tavern-smoke tulip pledges.

Then came the part the skeptic should not wave off, because a regulator wrote it down for us. On May 17, 2004, the Securities and Exchange Commission charged Lucent with securities fraud, alleging the company “fraudulently and improperly recognized approximately $1.148 billion of revenue” in its fiscal 2000, and charged nine of its current and former officers and employees along with one officer of Winstar Communications, a customer (SEC, 2004). The SEC’s complaint was precise: the violations were not that Lucent had built nothing – the fiber was real, the switches were real, the capacity was real. What the SEC found was that the revenue recognition itself was structurally circular. The equipment vendors had financed their customers’ purchases, those customers had booked the equipment as assets, and the vendors had booked the sales as revenue – but the customers could only afford to “buy” because the vendors had lent them the money. The accounting was the bubble, not the underlying asset.

The Winstar file shows the machinery at its most naked: a $125 million software purchase engineered at the close of Lucent’s fourth quarter, revenue conjured at the deadline through a scheme run between a Lucent sales vice president and an officer of the customer buying with, in effect, the vendor’s own conviction (SEC, 2005). Lucent had entered that relationship holding Winstar out as a strategic partner. Every vendor’s book is full of customers who are certainly going to pay, right up until the certainty is the only asset left.

Now the two objections from the previous section can be addressed in a single pass. “OpenAI will pay the lease” is the password the structure requires, the exact sentence a receivable needs whispered over it to stay on the books as an asset, and it was true of every CLEC in 1999 until the morning it wasn’t. Goldgar’s demolition of the tulip market drama, far from sheltering the present, strips its best camouflage. The lazy critic said tulips and could be laughed out of the room with the archives that disproved the drama. Bury the myth, and the precise analogue stands up in its place wearing a docket number. A skeptic may keep the flower. The filings remain.

The filings are, in fact, multiplying in the present tense. In July 2026, reports that Nvidia might guarantee as much as 250 billion of OpenAI’s Ohio data-center commitments knocked roughly 4.5 percent off the company’s shares in a day, on exactly this anxiety (Fortune, 2026). On August 17, a securities filing fixed the figure at up to 105 billion of credit support for an initial 4.25 gigawatts of capacity, with an option on 3.75 more, compute arriving in phases in 2028 (CNBC, 2026). Fortune’s gloss on the arrangement was clinical: the dominant chipmaker is increasingly helping finance the infrastructure that generates the demand for its own chips, and the guarantee’s shrinkage of $145 billion against the July reports reads as a market forcing a vendor to admit how much of its demand it had been prepared to manufacture (Fortune, 2026). Bloomberg’s mapping of the wider web, the equity stakes and cloud commitments and chip purchases flowing among Microsoft, OpenAI, Nvidia, Oracle, and the rest, shows cloud providers and chipmakers funding the model builders who then rank among their largest customers (Bloomberg, 2026a). Michael Burry, watching the guarantee structure take shape, filed his analysis in six words: “around and around we go” (Benzinga, 2026). A short-seller reduced to a nursery rhyme is its own kind of market signal.

What’s Different This Time

The skeptic has a third line of concern, and it is the sharpest: AI is not tulip bulbs, and it is not telecom fiber that sits idle. GPUs produce inference. Models are being used. Google reports that Gemini has crossed one billion monthly active users – its fastest-growing product ever, announced by CEO Sundar Pichai on August 11, with 63 percent of users interacting by voice and the app generating over 150 million images daily (Ars Technica, 2026). OpenAI and Anthropic are engaged in a price war, with GPT-5.6 Luna dropping 80 percent in cost to $0.20 per million input tokens. Real adoption, real competition, real infrastructure – Amazon has announced its first off-the-grid data center, powered by dedicated energy generation rather than grid contracts.

This is not wrong. The GPUs are real. The users are real. The price competition is real.

But the reality of the asset does not negate the circularity of its financing. The tulip bulbs were real too – living, flowering, tradeable biological objects. The telecom fiber was real too – it still carries your data. The problem was never whether the thing existed; the problem was whether the promise about the thing was priced on independent evidence, or on a loop of mutual vouching.

When Nvidia guarantees $105 billion of OpenAI’s data center lease, and OpenAI signs a ten-year commitment to buy Nvidia chips, and Apollo books that arrangement as an “investable asset,” the pricing of the GPU’s future revenue depends on the GPU’s future revenue depending on the GPU’s future revenue. The loop is not broken by the fact that the loop produces something useful. It is broken by the fact that the loop is priced as though it were not a loop.

A Network of Values

Let us return to Haarlem, 400 years ago, with better questions. What actually failed there, if the economy held?

The tulip trade of the early 1600s ran on the same instruments as the rest of Dutch commerce, which is to say on notarized promises and personal standing among people who would meet again at market. A florist’s contract was a wager witnessed by neighbors. When the February collapse came and bulb buyers walked away from their winter promises, what tore was the fabric of belief in obligation itself, in a mercantile culture where credit and character shared a root. Goldgar’s court records show the scars of those oozing wounds, in suits over honor as much as guilders, in the vertigo of a community discovering that the price of a flower that blooms in June could swing wildly through January, and that value, which had felt like a property of things, was revealed as a property of agreement (Goldgar, 2007). Haarlem’s economy held. Haarlem’s handshake broke. The satirical pamphlets that so misled Mackay were the culture dressing that wound, moralizing the earthly flower because the alternative was admitting that the whole structure of worth at the foundation of a prospering nation had blinked.

The present press releases dress the same wound. A technology is being pressed into products at a pace the products’ buyers did not request. One American consumer in three tells Circana they do not want AI in their devices at all, need and privacy leading the reasons (Circana, 2026). TechCrunch, surveying the polling in mid-August, found resistance running deeper than anything the personal computer, the iPhone, or the internet itself met at comparable stages of adoption, with over seventy percent of Americans saying the technology is moving too fast and a majority of younger adults declining to trust the industry’s own leaders to behave responsibly (TechCrunch, 2026). And so the institutions have begun to legislate provenance. Brussels requires the artificial to be machine-marked. Cupertino requires the AI-made song to confess itself, whatever “a material portion” of a song turns out to mean in practice. These are our satirical pamphlets, anxiety codified instead of sung, a civilization writing itself assurances that it can still tell the grown thing from the generated one. The bitter joke inside the regime is the one Cicek and colleagues measured before the mandates landed: the mark meant to inform behaves in the market as a warning label, draining the very trust it was designed to secure (Cicek et al., 2025).

Goldgar’s 350 traders knew one another through guild, family, and church. The compute economy’s principals fit in a smaller room, and their obligations interlock more tightly than any florists’ syndicate ever managed. Consider one August afternoon’s worth. On August 10th, 2026, the same day Nvidia summoned its six underwriters, a Texas company called Riot Platforms disclosed a twenty-year lease of 191 megawatts at its Rockdale campus to a frontier AI laboratory, later identified as Anthropic, a $9.1 billion commitment running to June 2048, and Riot’s shares jumped 25 percent on the news (Bloomberg, 2026b). Riot mines Bitcoin. Or it did. The sheds it raised on cheap Texas power to hash one speculative asset are being re-let, hall by hall, to house the training runs of the next, the latest in a run of enormous compute agreements the laboratory has signed in recent months to feed models whose revenues remain a fraction of the commitments. One boom’s carcass, refitted as the next boom’s real estate, leased for two decades to a tenant priced on a future. Somewhere in Haarlem, the ghost of a bulb dealer is nodding.

The future being priced is, so far, largely failing to appear where it was promised to appear. The most careful field audit available, run out of MIT’s Project NANDA across three hundred public deployments, dozens of interviews, and 30 to 40 billion in enterprise spending, found 95 percent of organizations seeing zero measurable return on generative AI, with only five percent of task-specific tools surviving to production (Challapally, Pease, Raskar, & Chari, 2025). Read past the headline number and the report gets stranger and more relevant. Procurement, it finds, runs on referral and reputation; one purchasing chief, drowning in demos, told the researchers that “establishing trust is the real challenge” (Challapally et al., 2025). MIT went looking for returns and found, instead, a market coordinating on trust precisely where verifiable value is absent, which is a sentence Goldgar could have filed from the Haarlem archives without changing a word of method. The valuation resting on this techno-progress speculative field has left the ground of earnings entirely. Torsten Sløk, chief economist at Apollo, published the comparison in a single chart: the ten largest companies in the S&P 500 are “more overvalued than they were in the 1990s,” at the peak of the bubble everyone swears they remember (Sløk, 2025). Value unmoored from demonstrated use, adjudicated inside a closed ring of the mutually obligated, marked to one another’s confidence. That was never the legend of the tulip. That was the finding.

The Case for the Frenzy

Under the prairies of the American Midwest there is glass that spent a decade in the dark. The fiber went into the ground between 1996 and 2001 on demand forecasts that were off by years, and Starr’s autopsy lists it among the ruins, billions of dollars of networks unused for want of any prospective traffic, their builders broke (Starr, 2002). Then the years passed, video arrived, the cloud arrived, and the dark strands lit. Whoever streams this essay’s sources tonight will pull them through capital that bankrupted its investors and served its civilization like fruiting bodies on a waterlogged corpse.

Carlota Perez’s study of technological revolutions finds the same shape five times in 250 years: a new technology irrupts, finance floods in, the flood becomes a frenzy that decouples paper values from present earnings, the frenzy crashes, and then, on the far side of the wreckage, the golden age arrives, deployed across the very infrastructure the frenzy overbuilt (Perez, 2002). Canals, railways, electrification, the motorized economy, the internet. On this account the bubble is how a society talks itself into installing more capacity than any sober committee would approve, and the crash is the tuition. If Perez is right, then every loop documented in this essay can be real – the swaps, the guarantees, the vendor’s conviction booked as revenue – and the verdict of history will still be gratitude, because the data centers, the substations, the silicon, and the accumulated craft of running them at scale will outlive the paper that paid for them, waiting for their own dark decade to end.

The measured version of the case comes from Noah Smith, who walked through the circular deals as they multiplied and concluded that the anxiety, while reasonable, may be overdrawn: much of the spending flows from the operating cash of the most profitable firms in history rather than from the fragile leverage of thinly capitalized carriers, the assets being built are physical and durable, and a correction, if it comes, could plausibly stay inside the walls of the people who can afford it (Smith, 2025). Grant Huang his strongest ground here too. Somebody genuinely does have to bridge the gap between a laboratory’s ambitions and its credit profile, and a supplier with a fortress balance sheet stepping into that gap is not automatically a con; it can be, and sometimes has been, how frontiers get financed. The honest reader should feel the pull of all this. The frenzy builds real things, the builders are rich, and the last dark fiber eventually carried the world.

Three Questions, One Circuit

Question one: does the frenzy leave durable capital behind?

Often, yes. Perez has the receipts, and Ohio’s gigawatts will exist whatever happens to the paper.

Question two: do the people financing the frenzy get their money back?

Here the telecom record answers without hedging. The fiber survived and eventually lit, and its survival consoled nobody holding Nortel at the top, nobody holding the trillion in debt the FCC chairman wrote off in congressional testimony, and none of the nine Lucent officers reading their names in a federal complaint (SEC, 2004). Infrastructure redeemed the civilization, but redeemed the ledgers of no one.

Question three, the tulip’s question: does the web of trust survive?

Rails do not repair that, and Perez’s framework, honest about its own scope, has nothing to say about it. Haarlem kept its wealth and lost, for a season, its ability to believe a neighbor’s promise about worth. Modern interconnected markets may not be able to sustain that window of disbelief.

What makes the present cycle genuinely novel, the thing neither Haarlem nor Murray Hill quite managed, is that the second and third questions have been fused. In 1637 the network of values and the network of finance were adjacent; in 2000 the vendor and its customer at least kept separate books that could fail separately. In 2026 the parties vouching for one another and the parties financing one another are the same parties, holding one another’s equity, guaranteeing one another’s leases, buying one another’s output, and pricing the collateral, which is a forecast of intelligence, by mutual agreement, since no external meter for it exists. Dark fiber’s worth was always eventually countable; a bit either moves or it doesn’t. Trained capability at the promised scale is a claim, and the claim’s principal underwriters are its principal beneficiaries. When trust and solvency run on one circuit, they fail as one event.

And the circuit, having spent July under inspection, spent August getting bigger and moving its risk outward. The same Fortune reporting that tracked the shrunken guarantee noted a quieter July development: SEC staff guidance concluding that certain data-center debt sits outside the Dodd-Frank securitization rules that require sponsors to keep a slice of the risk on their own books, an interpretation that makes it markedly easier for a vendor to mobilize outside capital rather than carry the exposure itself (Fortune, 2026). Weeks later came the six-house platform, dedicated pools of third-party money for Nvidia’s offtakers, the vendor’s residual backstop capped at a quarter. Anyone who lived through 2008 recognizes the choreography of originate and distribute, risk written by the party with the information and held, in the fullness of time, by parties without it, and Nvidia’s own release supplies the destination in its partners’ boilerplate, where Apollo introduces Athene, its retirement-services business, keeper of other people’s old age (NVIDIA, 2026).

Apollo’s academy published the sharpest public warning that this bubble exceeds the last great one. Apollo’s platform now helps underwrite it. One house, both sides of the prophecy. The credit market has started to price the arrangement on its merits; when the guarantee reports first surfaced, default swaps on Nvidia’s bonds recorded their steepest intraday jump since they began actively trading (Axios, 2026). The tulip trade never securitized the promises that broke it. We fixed that.

Before the Vouching Stops

The canonical tulip crash is a sermon Mackay mistook for a deposition, and anyone waving it as economics deserves Goldgar’s archives dropped on their desk. The frenzy does build; Perez’s five surges are real history, and the odds are decent that Ohio’s capacity, like the prairie glass, serves people not yet born under owners not yet imagined. And the demand may yet arrive on schedule, in which case the leases get paid, the receivables mature into cash, and Huang’s five-word answer ages into a modest statement of fact.

Here is why the thesis stands anyway, and it comes down to a test any reader can apply. Every plank of the counterposition is a forecast. It asks for patience while the future validates a structure of scrap wood and penny nails. Every plank of the thesis is a filing. The vendor financing its buyers is in an August securities filing. The circular booking of a supplier’s conviction as revenue is in a 2004 federal complaint, with names. The absence of the returns that would anchor the valuation is in MIT’s field data. The detachment of the valuation from earnings is in Apollo’s own chart. The migration of the risk toward pooled and retirement capital is in a press release and a staff interpretation of Dodd-Frank. The erosion of the surrounding trust is in the polling, in the labeling mandates, and in a peer-reviewed measurement showing the label itself now functions as a warning. An argument that strengthens as you check it beats an argument that asks you to wait, and that asymmetry represents the confidence in the matter.

Where does this all lead? Probably, hopefully, nothing so cinematic as 1929, and nothing so tidy as containment either. The likeliest path looks a lot like the tulip’s break wearing Lucent’s clothes: a failure of vouching inside the interlocked circle, leases renegotiated, guarantees tested or failed, one set of marks no counterparty will initial, propagating through balance sheets that were engineered to hold one another up, with two modern amplifiers the Dutch were spared. The circle’s members now weigh so much that the bulk and momentum of the modern stock index is mostly them, which is the true edge of Sløk’s warning, and the circle spent this August piping its obligations outward into the savings of people who have never heard the word offtaker. The infrastructure will survive the paper; it generally does. The fable will survive everything, because that is what it is for, and its next edition is already typeset in Brussels and Cupertino, our anxieties about the generated thing filed as metadata the way Haarlem filed its own as verse.

Four centuries from the winter of 1637, one detail still divides the cases. However badly the florists misjudged the market, however many promises died in the taverns of Haarlem, every buyer in that trade could, in the end, walk out into the cold air holding the thing itself, damp, brown, alive, worth something to somebody with a garden. The compute loop trades a claim on an intelligence that does not yet exist at the scale being financed, collateralized by the confidence of the parties selling it, marked to a market they constitute. When the vouching stops, and every prior version of this arrangement teaches that the vouching stops, honor and solvency will go in the same hour, because we wired them into one circuit and called the wiring innovation. The Dutch at least got a tulip. Check what is in your hand for that LLM subscription fee.

References

Axios. (2026, July 27). Nvidia reignites “circular” AI concerns as it weighs OpenAI financing guarantee. https://www.axios.com/2026/07/27/nvidia-openai-financing-ai-jensen-huang-ssi

Benzinga. (2026, July). Nvidia funds AI frenzy: Timeline of its circular financing deals … so far. https://www.benzinga.com/markets/tech/26/07/60713664/nvidia-funds-ai-frenzy-timeline-of-its-circular-financing-deals-so-far

Bloomberg. (2026a, January 22). AI circular deals: How Microsoft, OpenAI and Nvidia keep paying each other. https://www.bloomberg.com/graphics/2026-ai-circular-deals/

Bloomberg. (2026b, August 11). Anthropic strikes $9 billion deal with cloud computing firm Riot. https://www.bloomberg.com/news/articles/2026-08-11/anthropic-strikes-9-billion-deal-with-cloud-computing-firm-riot

Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025, July). The GenAI divide: State of AI in business 2025. MIT Project NANDA. https://www.artificialintelligence-news.com/wp-content/uploads/2025/08/ai_report_2025.pdf

Cicek, M., Gursoy, D., & Lu, L. (2025). Adverse impacts of revealing the presence of “Artificial Intelligence (AI)” technology in product and service descriptions on purchase intentions: The mediating role of emotional trust and the moderating role of perceived risk. Journal of Hospitality Marketing & Management, 34(1), 1-23. https://doi.org/10.1080/19368623.2024.2368040

Circana. (2026, January 27). Most consumers are aware of AI, but one-third don’t want it in their devices. https://www.circana.com/post/most-consumers-are-aware-of-ai-but-one-third-don-t-want-it-in-their-devices-circana-reports

CNBC. (2026, August 17). Nvidia backing $105 billion in financing for OpenAI data center in Ohio. https://www.cnbc.com/2026/08/17/nvidia-financing-open-ai-data-center-ohio.html

European Commission. (2026). Guidelines on transparency obligations for providers and deployers of certain AI systems (Article 50). Shaping Europe’s Digital Future. https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content

Fortune. (2026, August 18). OpenAI data center deal with Nvidia comes in $145 billion lower than reported. https://fortune.com/2026/08/18/openai-data-center-deal-with-nvidia-comes-in-145-billion-lower-than-reportedsignaling-concerns-of-artificial-demand-for-chips/

Galbraith, J. K. (1990). A short history of financial euphoria. Whittle Direct Books.

Goldgar, A. (2007). Tulipmania: Money, honor, and knowledge in the Dutch Golden Age. University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/T/bo5414939.html

Goldgar, A. (2018). Tulip mania: The classic story of a Dutch financial bubble is mostly wrong. The Conversation. https://theconversation.com/tulip-mania-the-classic-story-of-a-dutch-financial-bubble-is-mostly-wrong-91413

NVIDIA. (2026, August 10). NVIDIA partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish AI compute infrastructure financing platforms to mobilize over $500 billion of third-party capital. NVIDIA Newsroom. https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital

Perez, C. (2002). Technological revolutions and financial capital: The dynamics of bubbles and golden ages. Edward Elgar. https://carlotaperez.org/publications/

Reuters. (2026, August 10). Nvidia partners with Wall Street giants to raise $500 billion for AI buildout. Yahoo Finance. https://finance.yahoo.com/technology/ai/articles/wall-street-giants-partner-nvidia-163441421.html

Roos, D. (2020, March 16; updated 2025, May 27). The real story behind the 17th-century ‘Tulip Mania’ financial crash. History, A&E Television Networks. https://www.history.com/articles/tulip-mania-financial-crash-holland

Sløk, T. (2025, July 16). AI bubble today is bigger than the IT bubble in the 1990s. Apollo Academy. https://www.apolloacademy.com/ai-bubble-today-is-bigger-than-the-it-bubble-in-the-1990s/

Smith, N. (2025, October). Should we worry about AI’s circular deals? Noahpinion. https://www.noahpinion.blog/p/should-we-worry-about-ais-circular

Starr, P. (2002, September 9). The great telecom implosion. The American Prospect. https://www.princeton.edu/~starr/articles/articles02/Starr-TelecomImplosion-9-02.htm

TechCrunch. (2026, August 19). AI was supposed to win people over by now – it hasn’t. https://techcrunch.com/2026/08/19/ai-was-supposed-to-win-people-over-by-now-it-hasnt/

Tunguz, T. (2025, October 3). Circular financing: Does Nvidia’s $110B bet echo the telecom bubble? https://tomtunguz.com/nvidia_nortel_vendor_financing_comparison/

U.S. Securities and Exchange Commission. (2004, May 17). Lucent settles SEC enforcement action charging the company with $1.1 billion accounting fraud (Press Release No. 2004-67). https://www.sec.gov/news/press/2004-67.htm

U.S. Securities and Exchange Commission. (2005, October 20). SEC v. Lucent Technologies Inc., et al.: Court enters final judgment against David Ackerman (Litigation Release No. 19437). https://www.sec.gov/enforcement-litigation/litigation-releases/lr-19437

Variety. (2026, August 20). Apple Music to label AI-made tracks later this year. https://variety.com/2026/music/news/apple-music-to-label-ai-made-tracks-1236839371/

Yahoo Finance. (2026, August). Nvidia-OpenAI deal isn’t ‘circular financing,’ says Jensen Huang. https://finance.yahoo.com/technology/ai/articles/nvidia-openai-deal-isnt-circular-172204839.html