Frédéric Lordon

18 September 2026Economics

There they all are, gathered around the table of the CNBC studio. The giants, the big names who really matter – the bosses. Breaking news, live coverage, news ticker along the bottom of the screen. An unprecedented spectacle. In short, history is being made, right here on set. The executives of Nvidia, Goldman Sachs, Blackstone, BlackRock, Apollo and KKR are all in attendance. OpenAI and Anthropic? They didn’t make the cut. Google, Microsoft, Oracle, Amazon? Ditto. What did this gathering in early August represent? The titans of chips on the one side and the cream of finance on the other. Such is the reality of the AI economy. The official reason for the change in programming? The announcement of a $500 billion mega-deal to finance AI infrastructure. It’s easy to lose your head with the sums being thrown around these days. The last bombshell was OpenAI raising $122 billion in its most recent funding round. That now pales into insignificance.

Fortunately, there are the contrarians, the Schadenfreude enthusiasts, reading the small print. By the evening, a prominent dissenter is already imploring on X: ‘It’s a god damn MOU’. And indeed, the press release on Nvidia’s website reads: ‘Memorandums of understanding signed with six of the world’s premier financial institutions  . . . ’ What is an MOU? A declaration of intent. In other words: hot air. Look beyond the spectacle and this is not an investment plan. It is a PR plan. We might even call it a psy-op. Why throw $500 billion across the table live on television? Because morale is flagging. Doubt is beginning to creep in. Are the trillions flying in every direction really sensible? That is the question that must not be asked.

The financial trajectory of AI is so extraordinary that it requires faith as hard as titanium to sustain it. And yet it is beginning to wobble – even before doomsaying calls for a ‘coordinated slowdown’ began emanating from AI labs. Hence the need for a psy-op, and they’re not holding back. Nothing but heavy artillery. Heavy in financial terms, of course, but also in symbolic terms. In a psy-op, you have to marshall vast amounts of symbolic capital to grip the imagination. In finance, you do this with numbers. And $500 billion, announced by Nvidia itself – who would not be sent into a frenzy? The whole operation is designed to confirm that everything is fine. The proof: we’re upping the stakes. Yet operations designed to insist that everything is fine inevitably betray the fact that this is not the case. And indeed, everything is not fine. We should pay close attention, since this is a moment of great historical interest – and incidently, great danger. For we are witnessing the great reversal unfolding in real time. The reversal the $500 billion announcement was designed to ward off. We can recognize it by the telltale signs: an exasperated headlong rush, frantic ingenuity to overcome dead ends, and a return to underhand methods seen during the subprime era.

There is certainly cause for some concern. Anthropic has committed to paying $330 billion to Google, AWS and Microsoft over the next decade; OpenAI, $770 billion to AWS, CoreWeave, Cerebras, Oracle and Microsoft by 2030. In total, these hyperscalers have racked up $2.1 trillion in . . . what exactly? Revenue? No: promises. The question is whether we can reasonably expect these promises to be fulfillled. If we were dealing with well-established giants, we might be moderately worried. But alas, this is not the case. OpenAI closed the fiscal year of 2025 with $13 billion in revenue and $21 billion in losses – or $39 billion, depending on whom you believe. Anthropic reported $9 billion in revenue for the same fiscal year, with losses estimated by some at $4 billion, by others at . . . $42 billion. All of this is completely opaque – not least because, as these companies are not yet publicly listed, they are under no obligation to disclose financial information; outside observers are left to piece the reality together as best they can.

And now the commitments – IOUs essentially – signed by these two companies are reaching their expiration date. These were formalized through what are known as ‘take-or-pay’ contracts, meaning that the buyer is obligated to pay even if they would rather forego the transaction by the time the contract is due – perhaps because market conditions now make the service unnecessary, in this case, computing time in a data centre operated by a hyperscaler. We must delve into the details of these arrangements to understand just how quickly the wall is approaching. There is a striking homology here with the subprime crisis. The business model of subprime mortages hinged on the ‘reset’ clause. Low-income people were lured to borrow money – when they in fact couldn’t afford to take on a mortgage – by a very low ‘teaser’ rate that abruptly defaulted to the regular, much higher rate after only two years, and remained there for the subsequent twenty-eight.

Those who didn’t read the small print collided with the reset wall after two years, unable to meet the costs of their interest payments which had suddenly quadrupled or quintupled without warning. Personal bankruptcy, foreclosure and eviction followed. Those who had read the small print, on the other hand, fully intended to get out in time, borrowing for speculative purposes and banking on the continued rise in property prices to sell and pocket the gain before the reset rate struck, and then perhaps even do it again. But clever though they were, the reversal in the property market sent them into bankruptcy just the same. In its own way, AI is repeating the pattern. Just as with subprime, you sign now and pay later. For the time being, everyone is celebrating: the labs are announcing legendary cash-burn rates, the hyperscalers are announcing sensational revenue projections – the financial community neglecting the word ‘projection’, preferring to look at these figures as actual revenue. A data centre takes two to three years to build – plenty of time to prolong the euphoria.

Then the wall is hit. Because now the commitments have come due, and it will be take-or-pay. Usage, non-usage – no one wants to hear about it. It was signed; now it must be paid. We were drunk on spectacular announcements; now the morning-after. And these payment obligations do not rise gradually but in steps, each of considerable height. In 2025 and 2026, OpenAI and Anthropic had to shell out $46 billion and $131 billion respectively. The reset hits in 2027: $412 billion to be paid out. 2028, the next step: $440 billion. The impending collision promises to be violent. For the gap between current economic means and order commitments signed is dizzying. And the entire financial edifice of AI – with its trillion-dollar valuations – depends on these two financial minnows turned mega-customers.

From the beginning, one could legitimately ask how chronically loss-making companies such as OpenAI and Anthropic could support future spending commitments amounting to hundreds of billions of dollars. The answer is that, in the absence of customers – and therefore sufficient revenue – investor capital makes up the difference. This is not entirely unusual during a company’s launch phase: they must invest before finding their market, and can only do so with funds from creditors or shareholders. But this gives rise to two questions. Will the industry find its market? And will it have been forced to engage in excessive financial manoeuvring while it seeks it? Until now, the answer to the first question appeared self-evident; the second has rarely even been asked. Now it is unavoidable.  

We have seen the situation facing the labs, trapped by the take-or-pay obligations coming due. But the hyperscalers are struggling too. For years they have poured their abundant cash flow into building data centres. But the race for gigantism – as is the nature of such epic ventures – eventually caught up with them: their net cash flow has slipped into the red. This is a major event for such profit machines, which are now forced to borrow. To be sure, the hyperscalers are viewed favourably by the world of finance. But it takes more than a friendly face to secure trillions in debt. What can they offer as proof of solvency? The trillions in future orders, of course. If you won’t lend to people on track to pocket $2.1 trillion, then who exactly will you lend to? So the financial sector – ignoring the distinction between promised orders and realized revenue – goes ahead and lends. Meanwhile, everyone either fails to see the problem or crosses their fingers. Baron Munchausen thought he could pull himself out of the mud by tugging on his own bootstraps, and AI is doing the same.

What are the chances of bootstrapping? The sustainability of the boom rests on two complementary expectations: a surge in revenue for the AI labs, and the continued refinancing of those labs by investors. The first expectation has fallen considerably behind schedule, if it is not collapsing altogether. The price war between OpenAI and Anthropic is already hurting both, though it is nothing compared to the threat posed by Chinese competition, which offers near-equivalent products at a hundredth of the price. As the prospect of revenue catching up fades, the only remaining option is investor backing. Driven by the kind of belief that fuels bubbles, investors are eager to join the epic adventure. Not without conditions, however. One above all: the projected valuation must continue to rise without end, justifying ever-larger capital injections.

Notice the correlation between OpenAI’s expected valuation and its fundraising. In 2023, the company’s projected valuation was $29 billion. It jumped to $157 billion at the end of 2024; that same year it raised $6 billion. By March 2025, the valuation had risen to $300 billion; fundraising was $40 billion. In March 2026, it was valued at . . . $852 billion; this year’s funding round ends at $122 billion. In the absence of revenue and profits, everything else is experiencing take-off. Acceleration is the decisive point here. It embodies the promise that the future will not merely be brighter, but increasingly brighter. To remain balanced, a bicycle requires only constant speed; the financial bicycle of AI, by contrast, requires increasing speed (spare a thought for the cyclist). In mathematics, this is known as the second derivative. The first derivative is speed; the second is the rate of change. And here lies the exponential condition for the system’s sustainability: for projected valuation, the second derivative must not turn negative. Yet, in the real economic world, sustained exponential growth does not exist.

When will this reality hit? At the very latest with the IPO – the moment when the true price will, in a sense, be revealed. Much is being made of the IPOs of both OpenAI and Anthropic; and rightly so. Yet we see not the slightest sign of a return to sanity. In Anthropic’s case, the anticipated valuation has doubled: $2 trillion is now expected, hoped for – a final spasm of belief. As for OpenAI, the aim was to go public in 2026, only for plans to change. It would have been difficult, we were told, to follow the IPOs of SpaceX and Anthropic – and ‘ill-advised’ given the furore about the technology’s risks. We also understand that CEO Sam Altman would like to take the time to secure a valuation of at least $1 trillion – the recent $852 billion figure was apparently a little lacklustre. But let us assume a final rush of euphoria takes the company to the IPO, and the popping of champagne corks. What matters is what comes next. The companies are publicly listed – what happens to the share price? SpaceX’s stock has seen some unfavourable fluctuations. But for the two labs, this will not be an option. The ascent will have to continue, as it is far from certain that the $60–70 billion raised at the IPO will be enough to settle their accounts. If revenue growth remains stubbornly linear, if the second derivative turns negative, then colossal problems will arise for everyone. Labs deprived of the means to meet their commitments; hyperscalers deprived of their future revenue; hyperscaler creditors finding their debtors in dire straits. That is how a house of cards collapses.

We are entering what one might call the latency period – that characteristic phase in which the belief which sustains a bubble is still in place but faces relentless assault, where verywhere there are indicators that the previous dynamic is unsustainable. And yet the logical consequence – a crash – has not yet ensued. During this period, a realization begins to dawn, though struggles to assert itself because it runs counter to the dominant belief, even as it gradually destabilizes it. What is this realization today? That the AI edifice is a gigantic, inverted pyramid resting on the tip of a single hypothesis: the constantly accelerating upward revision of the labs’ future valuations.

*

Clearly, the atmosphere is starting to change. The AI ​​sector is now rife with nervousness and agitation. Dirty tricks are spreading. The most glaring symptom is the massive growth in hyperscalers’ off-balance-sheet commitments. The corrosive effect of doubt naturally produces a reluctance among banks to lend – as seen in the case of a banking consortium led by Morgan Stanley seeking to offload the $15 billion it provided to Anthropic for a data centre operated by Google. One mechanism being used for this shedding of exposure is something called a Synthetic Risk Transfer. These are essentially insurance contracts, where an investor agrees to assume risk on behalf of a lender, generally a bank, which it compensates in the event that the borrower defaults. The subprime era had similar hedging instruments in the form of the infamous Credit Default Swap, though with one key difference – back then, the underlying loans remained on the bank’s balance sheet, entailing various drawbacks in terms of the regulatory capital required to back such risky exposures. Today, once the risk is transferred through an SRT, the associated capital requirements are reduced accordingly, allowing the bank to resume lending – to entities other than AI players, whom banks are no longer so keen on. The result is that risk is spread throughout the financial system, becoming increasingly diffuse and hard to track.

Finding traditional banks no longer hospitable, AI players have ventured into every corner of the unregulated world. And they have found what they were looking for. Hedge funds, private credit, private equity. AI finance began tapping into every available resource. The instrument of the moment is the Special Purpose Vehicle: an empty legal shell, usually domiciled in the Cayman Islands or the Bahamas, designed to hold massive debts that no one else wants cluttering their balance sheet. Take ‘Beignet’, for example. Beignet is the deceptively sweet-sounding name given to a $30 billion data centre project for Meta. Meta, in a joint venture with Blue Owl, one of the leading private equity firms, capitalizes the SPV with $3 billion in equity, only to immediately load it with $27 billion in debt. The SPV spends its $30 billion on infrastructure (land, buildings, chips, power), whose capacity Meta then leases, paying for the service to – and through – the SPV, which can thereby service the debt.

The motto of both labs and hyperscalers has become ‘it’s not me’. The chips in the data centre? I don’t own them, I only lease them, protecting myself from their obsolescence. The debt? Not me either – legally it belongs to the SPV. Debt creates unsightly blotches on a balance sheet, and I want mine spotless. Of course, to get other investors to subscribe to the $27 billion loan, it helps to offer some guarantees. First, there is the interest rate, complete with a handsome risk premium. Above all, there is the backstop provided by Meta itself, committing, come what may, to channel sufficient computing purchases through the SPV to service the debt. This offers the added advantage that a backstop is also an off-balance-sheet commitment – meaning it remains invisible.

Everywhere you look, arrangments like this are springing up. Apollo and Blackstone have joined forces with Anthropic on a $35 billion project to lease computing capacity from a Google data centre. In this case, there is a twist: the chip supplier, Broadcom, is providing the backstop and would compensate creditors suffering a default with . . . its own merchandise. We are not far from being asked to believe – literally, not just figuratively – that chips are gold. At the very least, they are being presented as themselves a means of payment. This is the logic generalized by the $500 billion Nvidia & co. mega-deal. Nvidia is no longer merely the boss: it is the godfather. Its gold-standard chips are being offered as the backstop for every deal that deploys the $500 billion. Finance has always been known for its creativity, but with this AI saga, it has truly outdone itself.

And now we discover that the CEO of Nvidia aims to do nothing less than transform ‘compute’ – that is, an hour of computing power – into a financial asset tradable on markets. Here, arrogance reaches new heights. First, because creating financial assets requires the underlying asset to be fungible: when an investor buys compute futures, they need assurance that an hour of computing power here is identical to an hour there. Jensen Huang asserts that the hour of computing power underpinning these financial products will be the GPU hour – specifically, the Nvidia chip hour. Nvidia declares: the hour of compute on my chip is universal. Consequently, the prerequisite for the financialization of compute is satisfied (by me). When asked whether chip obsolescence might pose a problem for their quality as collateral assets, he says, no, there is no problem, because they are his chips – they defy obsolescence. It is enough to take one’s breath away. Yet not as much as imagining the speculative circus destined to be built upon this foundation, no doubt featuring the full, garish array of finance’s worst excesses.

We need to step back a little to gauge the consequences of all this finanical creativity:

1. Off-balance-sheet practices have gone from being incidental to central and widespread. The amount of debt involved is so mind-boggling that there is a preference for keeping it out of sight. Expenditure commitments – whether those made by AI labs to hyperscalers or by hyperscalers to chip manufacturers and suppliers – stem from the same logic of concealment, albeit in a slightly different form. Formally speaking, they are debts (as they represent commitments to order or pay), yet their specific nature falls outside standard accounting records. Consequently, they do not appear anywhere; Goldman Sachs reports that, of the $1.5 trillion in identified IOUs, $1 trillion is not reflected in financial statements.

2. The chip suppliers providing backstops claim they can pay, or provide compensation, by delivering goods with a notoriously short shelf life – a downside of technological progress moving so wonderfully fast; in this case, perhaps too fast. No matter. Nvidia declares that its colossal balance sheet can underpin almost the entire edifice of AI debt. Never before has an industrial company positioned itself as the guarantor of such a gigantic financial pyramid.

3. With banks withdrawing, private finance has stepped in. But with whose capital? Increasingly, that of the great institutional savings collectors: pension funds and mutual funds. Future retirees and small savers are being unwittingly drawn into the most precarious schemes supporting the inverted pyramid – one that would inevitably bury them should it collapse.

4. The involvement of public savings extends throughout the entire private credit universe. Private credit funds have been under severe pressure since the beginning of 2026, with borrower default rates rising on the one hand and worried investors seeking to withdraw their money on the other. There is a growing desire to unload commitments that might charitably be described as malodorous. How will they go about it? Just as in the heroic days of subprime, of course: through securitization and ‘structuring’. While banks may have been urged to tone this down since 2008, private credit funds do not consider themselves so bound. Following a tried-and-tested formula, they pool together loans of varying quality, slice the resulting portfolio into different tranches of securities each carrying a particular level of risk – and corresponding remuneration – and put them up for sale. Investors choose their poison.

But this is where it gets truly insidious: insurers have developed a passion for this game and are playing it from every angle. First, on the investor side, they eagerly buy these structured products, presumably favouring the lower-risk categories, though we know what became of those supposedly AAA-rate tranches in the subprime era. Then – and this is a new development – insurers have discovered that it is lucrative to serve as ‘wrappers’, essentially acting as guarantors for other investors in these products, which are not exactly prudent, family-friendly investments. If defaults occur, the insurers must compensate the investors holding the bag; the overall strategy is to collect insurance premiums while hoping the dreaded event never comes to pass. Playing both sides, insurers – who managed to keep a low profile in 2008 – stand on the front lines when it comes to making headlines should a collapse occur.

The warning signs of the coming crash keep multiplying. Recall the bicycle: the trouble begins not when valuations collapse, but simply when they slow down. We will soon know where we stand on that front, given that there is now very little left in the AI business model capable of producing a second derivative. The proliferation of private credit unloading structures and other murky practices we thought we had left behind in the ruins of the subprime crisis bear witness to this. Yet we also learn that the Securities and Exchange Commission, the US market regulator, has just authorized the securitization of hyperscalers’ debt, on the basis of highly specious arguments. Here we go again. . .

Read on: Julian Stallabrass, ‘Ultra-Processed Images’, NLR 160.