https://newleftreview.org/sidecar/posts/awaiting-the-crash
Frédéric Lordon
09 July 2026 Economics

What is a bubble? It is a collective belief. What is a crash? It is the collapse of that belief. AI has given rise to two bubbles. There is a stock market bubble – the two highest-profile players, OpenAI and Anthropic, are expected to launch IPOs heralding astronomical market capitalizations of around $1 trillion each. But this is underwritten by a credit bubble, which is of even greater concern. Stock market crashes are often more spectacular than destructive – causing losses in asset value – whereas credit bubbles, when they burst, trigger a cascade of defaults throughout the financial system.

Two bubbles, generated by a collective belief powerful enough to sustain a mobilization of capital unprecedented in the history of capitalism. Let us grant that US capitalists know a thing or two about spinning a story – that is, generating a belief. This time, they have spared no effort, treating us to the most grandiose visions. But these have taken an unusual and paradoxical form: convincing humanity of the terrible, near-existential risks of the product they are selling. The head of Anthropic, Dario Amodei, has mastered this rhetorical style, which under the guise of contrition delivers an entirely self-serving message:

1) AI poses a massive danger, meaning it is a tool of unprecedented power, which ought to interest you greatly;

2) I have spawned a monster, but I’m admitting it therefore my conscience is clear (so buy from me to get a dose of virtue along with your lethal weapon);

3) The government of the Free World is now on notice that this weapon must not fall into anyone else’s hands – the Chinese, say? The horror! – whereas my own hands, as I mentioned, are clean;

4) Given the world-shaking significance of all this, if things go south financially we should under no circumstances be allowed to fail like a lowly Lehman Brothers.

It is the perfect legend: the future of humanity at stake, villains who must not get their hands on the prize. Of course, a legend is not a business model. Until now, it was sufficient to cast a spell, and what a spell it cast: since 2020, the ‘Big Five’ – Microsoft, Google, Oracle, Meta and Amazon – have poured $1.9 trillion into the cauldron. Now, however, it needs to deliver a return – if not soon, which isn’t on the cards, then eventually, and on a scale commensurate with the investment. Those expressing scepticism about this were initially dismissed as grumblers, killjoys incapable of experiencing the thrill of the miraculous, of envisioning the great civilizational breakthrough of our time. How long can the collective belief in AI withstand evidence to the contrary? The answer: a long time, but not forever – especially when warning signs begin to multiply, as they are now. We may well be witnessing the start of the erosion of that belief; once a critical threshold is crossed, a financial correction will ensue as brutal as the preceding frenzy was manic.

Can the revenue forecasts possibly justify the capital expenditure? Everyone’s fate hinges on this: the hyperscalers – those cloud service providers that build massive factories for collecting, hosting and crunching data: AWS, Google, Microsoft, Oracle, Meta – and the AI ​​labs, which have committed to consuming chips and computing power on a similar scale. Anthropic has committed $330 billion to Google, AWS and Microsoft by 2029, while OpenAI has pledged $852 billion to AWS, CoreWeave, Cerebras, Oracle, Microsoft and others by the end of 2030. Such vast sums are intended to make headlines, feeding collective belief in the epoch-defining nature of AI. Yet the legal status and accounting of these ‘commitments’ (the French is engagements) are highly ambiguous – they range from legally binding contracts to mere suggestions, extravagant fantasies and talk of interstellar horizons.

At some point this will all have to come back down to earth. And whichever way reality bites, there will be casualties. If the AI labs have to dig into their pockets without demand keeping pace, it will spell their ruin; if the hyperscalers find themselves high and dry, that won’t end well either. That is because they’ve already committed $2 trillion in capital expenditure, and are planning to commit much more in the coming years (the target is $5.3 trillion for the period 2025–30, according to Goldman Sachs). No one knows the exact proportion of firm commitments because neither Anthropic nor OpenAI are publicly listed companies yet. What we do know are the figures for their recent fundraising rounds: $95 billion for Anthropic this year, $122 billion for OpenAI, which still leaves a gaping funding gap – and there’s no question of self-financing, as they are currently running massive losses. As if that weren’t enough, Jensen Huang, the boss of Nvidia – and let’s not forget that he is the one supplying the chips that power this whole industry – pointed out that, based on a cost of $80 to 100 billion per gigawatt of power, the planned data centres would end up costing not the $5.3 trillion anticipated by Goldman Sachs but somewhere between $9.5 and $15 trillion. He, too, has his own capital to recoup, after all.

So we are faced with an equation with one variable and one parameter. The variable: demand. The parameter: funding, which will buy us time until the variable deigns to materialize. Demand certainly started at breakneck pace, powered by irrational exuberance; one simply cannot afford to miss out on a revolution – especially a capitalist one. At that stage, the hype was all that was driving the momentum. Then came the initial wave of adoption, accompanied by ecstatic awe at AI’s power. Yet the trouble was that reflection of a more prosaic kind was now possible, particularly among finance departments, which have little patience for marvels if the numbers don’t add up. Having hooked their clients with free access, the AI labs began charging companies based on flat-rate plans. At this point, spending was still predictable. Once the next stage of addiction set in though, the pricing model suddenly shifted to being based on usage, i.e., tokens – the fundamental unit fed into Large Language Models (LLMs). This unannounced change led to AI costs skyrocketing, catching finance departments completely off guard.

Uber, for example, discovered that its projected annual AI budget had been obliterated in a single quarter. This was hardly a surprise: the hype had successfully convinced employees that to be part of the glorious new era and supercharge their personal productivity they had better get on board. To spur them on further, a new concept was invented – ‘tokenmaxxing’ (a mark of true faith), along with rankings and honour rolls for those who tokenmaxxed most effectively. As a result, employees threw themselves into it with gusto, so much so that the very companies that had whipped them into a frenzy had to slam on the emergency brakes. This included giants like Amazon and Meta. Usage was to be capped until the picture became clearer. Yet it remains anything but clear. Spending can only be assessed ex post, and productivity gains are inconsistent and opaque. Companies are willing to invest in AI but demand at least some visibility regarding returns on their investment – which, for the moment, is about as clear as a pool of fuel oil.

Clarity is unlikely to emerge anytime soon, particularly when it comes to pricing, which has been remarkably volatile. In June, the Wall Street Journal reported that OpenAI intends to cut token prices drastically, plainly a move to wrest market share from Anthropic. Yet like its rival, OpenAI has a critical need for revenue. All else being equal, lowering prices does not help in that regard – no matter how elastic, overall demand is ultimately determined by the companies consuming AI. And given the current uncertainty, the prevailing attitude is one of caution or even retrenchment.

Moreover, one might well question how serious the price war between OpenAI and Anthropic is – and, should it escalate, what the consequences might be. There would certainly be damage, not for the hyperscalers (who care little where demand comes from provided it is robust), but for the creditors and shareholders who backed the loser. But the clash looks more like a playground skirmish than the Battle of Guadalcanal. The real hostilities are underway elsewhere: Chinese competition. Despite facing a double embargo (both internal and external) and limited access to Nvidia chips, the Chinese, making the most of ‘creative constraint’, have developed LLMs that may be less sophisticated (though that is debatable) but are better tailored to the actual needs of consumers. After all, not everyone needs an AI to write a thesis on Lacan or to prove the Riemann hypothesis. Whether or not we can see clearly into the black box of Chinese AI, one thing is certain: DeepSeek has been able to offer AI nearly on a par but at prices that defy all competition. The price difference is in keeping with the difference between Chinese and US capital expenditure. The ratio is 1 to 10: $57 billion for China in 2025, compared to $443 billion for the US; 2027 estimates stand at $157 billion versus $1 trillion. It seems that in China, in the absence of trillions of dollars, they are doing some actual thinking.

The unveiling of DeepSeek was hailed as a ‘thunderclap’ – yet the fact that lightning had struck was immediately forgotten. Everyone reverted to the prevailing article of faith: the supremacy of US-made AI. This state of denial could not last long; after an initial spike followed by a lull (the period of denial), the weekly demand for Chinese tokens soon surged from 5 to 20 trillion in a single month, leaving US models, at 5 trillion, in the dust. The rock-bottom token pricing ought to shake the faithful out of their complacency, for what is at stake is nothing less than the potential collapse of a $5 trillion business model. Goldman Sachs – acting much like the Roman Congregation for the Doctrine of the Faith – may well stoke the fires of enthusiasm by predicting a shift from mere ‘chat’ AI to ‘agentic’ AI and forecasting an explosion in monthly token demand to nearly 120 quadrillion by 2030. Yet, curiously, they have omitted the crucial follow-up question: who will capture that market?

Unlike its Vatican counterpart, however, the Congregation of Goldman Sachs is not cut from a single cloth. Diversity of viewpoint is tolerated there – though let us be reassured: this is less a sign of integrity than of a variety of profit centres. One recalls that during the heyday of subprime mortgages, the bank’s sales arm was offloading toxic assets onto ordinary clients while its proprietary trading desk was boldly shorting the market. It is no contradiction, then, to hear it now dangling the prospect of quadrillions in tokens just as one of its in-house strategists predicts that a large share will head to China and Japan. As if to prove him right – and with little regard for its current US partnerships – Microsoft blithely announced the replacement of OpenAI and Anthropic products with DeepSeek. Goldman’s analyst insists that ‘major capital providers are overexposed to risk’ and argues that ‘shareholder value creation would be better served by allocating less capital to AI’. The problem, however, is that less is not an option for the economic model of US AI.

We must consider the issue not only from the demand side, but also that of the capital providers betting on this $5 trillion enterprise. The financial sector ought to be in a position to assess the validity of claims that an innovation is ‘groundbreaking’, to determine its time horizons and the sustainability of the financial backing required. Yet neither discernment nor tempering rationality are among the defining traits of the neoliberal form of finance. We saw this play out with the ‘New Economy’ (a grotesque label we have since forgotten) of the dot-com bubble. Here we go again . . .

We are entering hazardous territory. And we know what contribution to expect from the key players: nothing. OpenAI and Anthropic, haemorrhaging money, have yet to generate a single kopeck of profit. As for the hyperscalers, they aren’t in great shape either. The Financial Times estimates – ‘under the most generous assumptions’ – that, with the exception of Amazon, all hyperscalers will see negative returns on investment for the 2025–30 period. Oracle’s figure is a staggering -35%. Indeed, to sustain the financial pace, hyperscalers are now resorting to unexpected measures – including halting share buybacks, on which they previously lavished colossal sums to appease shareholders – and are beginning to pile up debt.

The bulk of the financial backing for AI comes from external sources. And that support is about to be closed off. Banks are starting to throw in the towel: by the end of 2025, they had already committed $450 billion to the AI complex according to a Chicago Federal Reserve estimate. Goldman Sachs Research notes that financing via ‘private markets’ is expected to become increasingly important – a return to the hushed, coded language of the Holy See. This is where the AI ​​‘financing plan’ stands: having to go foraging in the shadows. People are certainly eager to volunteer – after all, a lack of regulation is a true believer’s best friend.

Goldman Sachs raves about the diverse segments and asset classes ready to commit under the bold banner of ‘alternative investments’. Everyone and their dog is involved: private credit, private equity, infrastructure and real estate funds (real estate is key, given the massive footprints required for data centres). The boundaries between these alternatives are increasingly blurred, as are the lines separating them from players in the regulated financial system. Banks provide leverage to these entities; pension funds and insurers invest a portion of retirement savings in them, while some insurers even lend to them – it’s a free-for-all. Every corner of the financial world, whether shadowy or transparent, is thus implicated in AI financing, setting the stage for disaster.

If the edifice crumbles, the effects will indeed be catastrophic. And how could they not? The business model equation is fundamentally untenable: all the investment is based on the most far-fetched assumptions about demand. The financial world is slowly beginning to realize this, as illustrated by a few cautious abstentions by the banks and conversely, the headlong rush into shady territory where neoliberal finance displays no shortage of enthusiasm or imagination. The deal proposed to Anthropic by two private credit funds, Apollo and Blackstone, to hide its debt will be remembered as a classic of the genre. ‘Big Sky’, as the project is known – these guys have a poetic streak – is offering Anthropic access to Broadcom’s chips via a $35 billion lease that keeps the debt off its balance sheet. We have reached the point where we must avoid giving the impression of overheating, lest we frighten the market. Here are the convoluted mechanics:

1) Apollo and Blackstone create an ad hoc entity – a Special Purpose Vehicle (SPV) – from scratch;

2) The vehicle is funded with $35 billion: $800 million in private equity and $34 billion in private credit;

3) The $34 billion debt is further broken down into: two so-called senior tranches (the safest), one of $6 billion rated A1 (Moody’s) and another of $24 billion rated A2, plus a $4 billion junior tranche;

4) The deal takes a truly exotic turn when one discovers that the two senior tranches (totalling $30 billion) are guaranteed by . . . Broadcom, the very company from which the chips will be leased. This means that if Anthropic (which pays interest to the SPV) were to default, Broadcom would cover the loss;

5) On top of that, Morgan Stanley – advising Broadcom on this venture – also generously offered its services by lending to the investors wishing to buy these securities.

What could possibly go wrong? Entering this territory is the surest sign that a ‘credit cycle’ is veering off the rails. The resort to schemes as baroque as those chips-backed loans – and Big Sky is far from an isolated case – indicates that there is too much to hide. Morgan Stanley has published estimates of hyperscalers’ off-balance-sheet obligations that are truly hair-raising: combining Remaining Performance Obligations – i.e., future commitments to provide computing power based on capacity yet to be built – with other obligations, such as those to acquire land, buildings and chips, yields a total of $1.8 trillion ($800 billion plus $1 trillion). All of it off-balance-sheet, of course.

One might ask: where is all the money actually going? An increasingly rhetorical question, it seems. The AI ​​economic-financial complex has become a tangled, impenetrable thicket; grasping its every convolution is a daunting challenge. Yet, everywhere one looks, one sees dead ends and unpredictable variables which – however they play out – will have dire consequences for some, if not for everyone. And in the midst of it all, the financial sector is in up to its eyebrows, having nonchalantly trespassed every boundary of reason; now deeply anxious in private, it continues to promote the belief in public – a belief that is starting to falter. Strange behaviours, driven by agendas that are at times twisted and at others concealed, are coming to light. Amodei bluntly declares that if Anthropic fails to reach $1 trillion in revenue, he sees nothing that could prevent bankruptcy. Sam Altman, his counterpart at OpenAI, who had been desperate for an immediate IPO to avoid trailing the pack in a market drained by SpaceX and Anthropic, now believes it is time to reconsider, to take a little time to gain clarity on pricing, competition and demand.

A powder keg full to the brim – only a match is needed. And now a box of matches has appeared. First, rising interest rates. Take US Treasury bills, which are being pushed higher by monetary policy intended to counter the anticipated inflation from the conflict in the Gulf. Together with the Federal Reserve funds rate around which they revolve, the yields of these bills act as the benchmark for borrowing costs across the system. When it comes to rising interest rates, it sometimes takes very little to shake an edifice built on the spread between investment returns and the cost of borrowed funds; that spread cannot be allowed to narrow, let alone turn negative. One rate hike too many, and these bets suddenly move into the red, prompting speculators to rush for the exits.

Then there is what is unfolding in the equity markets. There is growing concern over the debt used to finance stock purchases readily extended by the brokers through whom investors place their trades. So-called ‘margin debt’ is anything but marginal: it has now reached an all-time high of $1.4 trillion. What happens if equity markets reverse course? Investors will face the infamous margin calls – a broker’s request that a client post additional collateral to secure a loan when the assets backing that loan are losing value. We should here revise our opening proposition: stock market bubbles are not especially dangerous so long as they remain disconnected from the credit system. They become dangerous when they create the potential for defaults and frantic scrambles for liquidity. To meet funding needs in one segment of the market, investors sell assets in another. That market, in turn, comes under liquidity stress, triggering further sales elsewhere, and so on. If the financial system is already hovering at a critical point of structural instability, this chain reaction can push it toward collapse.

Finally, there is the possibility of a major incident. A failed IPO by an AI lab. A symbolically significant stock market collapse – SpaceX, for instance, whose much-hyped debut failed to prevent its share price, after an initial surge, from beginning its descent to earth. And then there is Oracle, busy carving out a legacy of bankruptcy, one that could trigger a chain reaction: -35% return on investment, debt five times the size of its equity, and plans to lay off staff in batches of 10,000, ostensibly in the name of a massive AI-driven productivity leap but in fact a desperate measure to restore cash flow and stave off default. Oracle’s collapse would be the kind of event seen in every major financial crisis: the moment when the spell is suddenly broken, and a belief – undermined from within and now too fragile – comes crashing down. And then the rout is general.