Try Le Grand Continent. The ideas and analyses you won’t find anywhere else. But you’ll see them everywhere soon. In your inbox from eight euros per month
There are precisely 60 years since Jacques Lacan wrote: “The omnipresence of the human discourse may perhaps one day be embraced openly by an omnicommunication of its text.” We have reached that moment. Through the so‑called “artificial intelligence” technologies, the text of the great psychoanalytic Other – which each of us helps to write and in which our lives swim – becomes objective. And our words, in the infinitely plastic form of an automation of our shared symbolic humanity, meet an inexhaustible interlocutor. Even more troubling, the machinic form of “the omnipresence of human discourse” uncovers combinations that remain inaccessible to us, mere mortal bodies. Recently, this has plunged the mathematics community into the greatest disarray.
It is certain that what is happening represents a qualitative leap in the history of society and the mind; but that does not mean we are witnessing the end of history. Against the stupefying fantasy of a “Super Intelligence” whose advent is proclaimed by President Trump and the tech giants, it is necessary to situate AI technologies within their time and context.
Lacan anticipated, in advance, any eschatology. “That does not mean that it will be more granted,” he wrote about the text of a speech seized by omnicommunication. In the algorithm era, this disagreement of the human text is a matter of political economy, in other words profits, work, and ecology. As for its current relevance, it concerns the AI bubble that seems on the verge of bursting. Beyond this gigantic episode of overinvestment, a will to political control is at work. Its aporias must be unmasked.
Requiem for a boom
In 2026, the AI boom pressed on. The leading hyperscalers, those firms that design and operate the large-scale computing infrastructures and data centers, saw their investments soar. This frenzy is on track to become the most substantial investment boom in infrastructure in U.S. history. And if one takes the statements of NVIDIA’s CEO Jensen Huang at face value, since this transformative technology “will have a very positive impact on every sector and every country, there is no doubt that this movement will continue and we are only at the beginning.”
What stands out is notable: unlike prior waves, such as the late 1990s Internet bubble, the AI boom does not lift other sectors. It develops at the expense of the rest of the economy. While AI-related investments have risen by roughly 2 percentage points of GDP since 2023, investments elsewhere have declined by the same amount. As David Rosenberg, founder and chief strategist of Rosenberg Research, notes: “Over the past year, the United States’ economic growth has had only two engines. One is the boom in capital expenditures and in building AI data centers. The other is the rise in stock prices driven by the spending of the wealthiest, a wealth effect that itself is fed by the AI boom. In short, the United States has become an economy driven by a single sector.” Through this new boom, we thus witness a shift in the development model, placing this generic technology at the center of capital accumulation. Meant to inform all dimensions of social activity, its potential is so extraordinary that it even calls for colonizing the planet’s orbit to harvest electricity that the Earth alone cannot supply.
Cracks on the financial front
Google, Amazon, and SpaceX have their heads in the stars. For, one must admit, this boom defies gravity as far as financial realities are concerned. Each month, more money is poured into the AI ecosystem, but equally more money exits in the form of paper profits—snapped up by investors who wise enough to pull out at the right time. In the cautious language of the analysts at the Bank for International Settlements, this is “excessive investment.” The mechanism already dictates the end of the story: “The competition that feeds the boom is also the one that feeds the financial fragilities that will burst the bubble.”
These fragilities, opened like fissures in autumn, are threefold. The first concerns the circularity of financial and commercial ties between chip designers (NVIDIA, AMD…), AI model developers (OpenAI, Anthropic…), and data center operators. For the latter, this includes Big Tech (Amazon, Google, Microsoft) as well as firms for whom renting data center capacity is their core business (Oracle, the neo-cloud Coreweave, etc.). The sums involved are substantial. Consider a few numbers to grasp their scale. Since 2025, NVIDIA announced investments of $130 billion in OpenAI. OpenAI has preordered 250 billion compute and cloud services from Microsoft; the company also announced a five-year agreement worth $300 billion with Oracle. And Oracle and Microsoft are essential NVIDIA customers—Microsoft alone accounts for 22% of NVIDIA’s revenue. In principle, there is nothing surprising about complementarities driving these firms to forge strong capital and contractual ties. The sums and concentration, however, are unusually large. If one actor falters, the entire ecosystem risks being destabilized.
The second fragility lies in the debt itself. Circular deals serve as collateral to raise funds on financial markets. And here too the amounts are substantial. The KKR fund estimates the stock of AI-related debt in the United States stood at $600 billion in September 2026. That is less than the outstanding subprime mortgage debt in the United States, which stood at $800 billion—i.e., $1.3 trillion in inflation-adjusted terms. This debt, however, covers only a small portion of the sector’s exposure to a downturn. For the hyperscalers, the tally of off‑balance‑sheet commitments—such as credit guarantees and purchase commitments, as well as lease contracts—exceeds $3 trillion. In other words, these are sums that place us well beyond the weight of the subprime segment in 2007.
As in the pre-2007 period, the arrangements are becoming bolder as the bubble swells. The architecture of a recent deal worth more than $150 billion linking Google, Broadcom, and Anthropic illustrates this: Google and Broadcom supply the chips, officially owned by a Special-Purpose Vehicle (SPV), a bespoke company; and Anthropic then commits to renting these chips. The detour through an SPV thus enables off-balance-sheet debt. In this arrangement devised by Morgan Stanley, the sellers themselves originate the system that finances the sale of their goods. And those providing the financing are Apollo and Blackstone, which, to buy the chips and data-center capacity, draw on major savers’ pools—such as insurers, pension funds, and sovereign wealth funds. They also obtain credit from large banks to leverage the investment and inflate returns.
The trick of all this financial engineering lies in turning the volatile revenues of AI model providers like Anthropic into rents with the regularity and predictability that specialized infrastructure funds seek. The objective is also to contain funding costs by attracting enough capital. The incumbent sellers, such as Broadcom or NVIDIA, provide partial guarantees on this debt. The chips themselves are offered as collateral for these loans.
A complementary technique, already tested during the subprime era, is once again deployed. First, the classic securitization scheme: banking consortia or private finance funds package the debt before distributing it to institutional investors. Then, as in the infamous era of “toxic waste,” “nuclear holocaust,” or the Tyso punch, AI‑related debt is frequently bundled into structured products such as Collateralized Loan Obligations, without the buyers knowing their actual exposure. In short, as in the subprime era, the financial machine is trying to distribute risk to prolong the boom. Today, the contamination is global. As Robert Armstrong sums it up: “Every major pool of capital, around the developed world, finances data centers.”
Thus emerges the third fissure: this extension becomes increasingly precarious due to rising interest rates. In a global context of rising debt costs, rates demanded on the AI system rise further, and the cost of insurance against default (Credit Default Swaps, CDS) even more. For hyperscalers, the uptick in risk premia is notable; and especially the rise in the volume and cost of default insurance has surged in recent months, reaching about 1% for Meta and NVIDIA, i.e., a five-year default probability on the order of 8%. For the most exposed players, like the neocloud Coreweave, the evolution is brutal: by mid‑October, yields demanded on some five‑year notes stood at 14.5%, while CDS spikes corresponded to a 60% default probability over the same horizon. At such levels, financial strangulation is almost guaranteed.
Operational complications
Just as in the spring of 2007, fragility signals blink with increasing insistence. They point to real contradictions between the expectations fostered by investors and the conditions for valid valuation. In the medium term, the ecological footprint of AI—in terms of greenhouse-gas emissions, water use, mineral extraction, and electronic waste—maps a trajectory that is unsustainable. But in the immediate term, crisis factors hinge on the inability to deploy computing power to the level of the financing raised and the uncertainties about the possibility of delivering profits in line with expectations.
The most glaring problem lies in the mismatch between investment outlays and the capacity actually built. In the United States, but not only there, project delays multiply. There is a growing disconnect between expenditures incurred and the actual capacities added. What do these “construction investments” amount to? Primarily, a massive storage of AI chips. Not a marginal phenomenon, this tendency of “dark GPUs”—chips purchased but not used—is estimated at $234 billion, i.e., 50% of the total value of chips sold by NVIDIA and Broadcom since early 2023. Such levels of stockpiling clearly indicate that speculation has overridden practical use. After all, why acquire chips you cannot use, except to hope to resell them at a higher price or deploy them under more favorable conditions? One might wonder what guarantees there are that such favorable conditions will be realized within a profit horizon.
Several reasons explain why data-center construction is lagging. Some are purely operational: a shortage of skilled workers, strained electrical grids, or shortages of equipment such as gas turbines and electric transformers. But the most important obstacle—and one Anthropic identified in its IPO briefing documents—is the growing pushback from local populations. These infrastructures are notorious for raising electricity costs, consuming precious water resources, and encroaching on land. To counter their construction, resistance from communities is growing. Legal challenges multiply, and authorities announce restrictive measures. Moratoria are now in place in 22 states – including California, New York, and Texas.
Profitability in question
The second set of uncertainties concerns the solidity of revenue projections; there are gaps on both the demand and the supply side. The expectations underlying the colossal investments require several things: first, the benefits from using AI services must be sufficiently large to sustain dynamic and durable demand. Numerous case studies and anecdotal accounts testify to impressive progress in work efficiency. This is visible, for example, in the operations of call-center platforms, in software engineers, or, more anecdotally, in the testimony of a wood and steel engraver who reports improvements. However, moving from the level of a single task to that of large organizations—and more broadly the economy—does not come automatically. Economists forecast modest advances, but empirical studies are ambivalent. And for the moment, the gains are not materializing at scale. Forty years after its initial articulation, Solow’s famous paradox remains: AI is everywhere except in productivity statistics.
As the chief economist of an asset manager puts it: “The fundamental economic error in the optimistic argument about AI is that the marginal cost of computing exceeds the marginal benefit to users.” In other words, it costs more to produce computation than it yields to customers. If that is the case, the idea that demand for chips and data centers will be unlimited is false. The obvious question then becomes: is it profitable?
The cost and usage uncertainties translate into price mechanisms through the absence of stability in pricing methods. At the start of 2026, several firms moved their major customers from a subscription model to an usage-based pricing structure. The base unit is the million tokens—the tokens representing a word, a fragment of a word, or a punctuation mark. When a user sends a query to an AI tool like ChatGPT or Claude, it is broken into tokens. In response, the tool also generates tokens. These tokens are then used to answer the initial query, in the form of new tokens that constitute the result. Since the number of tokens is tied to the use of microprocessors—and thus to the capacity of data centers and the electricity consumed—it seems like a reasonable accounting concept for the AI economy. In reality, it is not quite so. For the same task, different models will use more or fewer tokens. Furthermore, the deployment of AI agents changes the picture: they consume tokens in amounts that are hard to predict, depending on the strategies devised to accomplish the tasks at hand. For users, the downside of autonomous agents is the inability to anticipate their consumption. In other words, tokenization transfers the uncertainty of the AI production process from providers to consumers.
Although the cost of tokens for the most capable models has fallen by more than 90% since 2023, in line with Jevons’ paradox, users’ overall spending on AI is rising rapidly. The rise is estimated at 63.4% between 2025 and 2026, from $39 billion to $64 billion. In spring 2026, firms like Meta or Amazon introduced performance rankings of their staff based on token usage to encourage greater use of AI. Yet that token‑driven overuse fueling a ballooning bill did not endure.
Since the beginning of summer, a far more frugal approach has taken hold. On the one hand, consumers are turning more to open-weight models, which are most often of Chinese origin. Their performance is only a few months behind frontier models, and at a fraction of the price charged for OpenAI’s and Anthropic’s most advanced models. On the other hand, the latter two firms have entered a price war to defend their market shares, even as they continue to accumulate losses. This pressure on their revenues adds to fragility. In parallel, the Chinese giant DeepSeek has introduced a time‑of‑day pricing system, intended to better account for energy costs—these costs represent the main expense of AI after the fixed capital invested in chips—indicating a possible commodification of AI. In the long term, AI may be reduced to a simple brute force of computation, thereby depriving the leading players of monopolistic rents that justified the investments.
There is little doubt that the macroeconomic trajectory of AI-driven growth in the United States will soon taper and turn into a major crisis. This seems evident given the extreme concentration of investments in data centers and the financial acrobatics that sustain them, as well as the difficulties in deploying technologies and the uncertainties surrounding potential revenues.
Yet, as we suggested by quoting Lacan in the introduction, the issue of AI goes beyond economic turbulence. By the very nature of what is produced, by the unprecedented power of centralizing knowledge, and by the depth of its entanglement with social processes, the AI challenge goes beyond mere valuation. It directly touches the question of political and geopolitical control. It is thus necessary to consider both the project underpinning the boom and the twists of the crisis to come.
The Preference for Control
In a classic 1943 article titled “Political Aspects of Full Employment,” the economist Michal Kalecki explains why business circles oppose full-employment policies. Economically, there is no reason; public spending stimulus that revives activity tends to boost profits. It is on the political plane that resistance arises: “In a system of permanent full employment, ‘layoffs’ would cease to function as a disciplinary instrument. The social position of the business class would be weakened, while workers’ self-confidence and class consciousness would be strengthened.” Between such a threat and a favorable conjunctural shift there is no hesitation: “Business leaders prefer ‘discipline in the factories’ and political stability to profits. Their class instinct tells them that durable full employment is not viable for them.”
Since then, the capital’s instinct has not softened. The rightward radicalization of a slice of the tech elite in recent years is a political reaction to rising social resistance and regulatory ambitions under the Biden administration, in a context of near‑full employment. Yet this instinct is not purely defensive. It also takes an offensive turn because of the hybrid character of the deployment modes of digital technologies – which straddle the economic and the political. If we follow this hypothesis, the U.S. AI bubble does not primarily carry promises of profits but a grand political ambition. In substance, it seeks control: the defeat of labor, domination of the planet, and—beyond—an adolescent fantasy of machinic subjugation of humanity. Yet nothing is settled this time either.
From superintelligence to the singularity
For years, tech leaders such as Madhavi Sekar (DeepMind), Dario Amodei (Anthropic), Sam Altman (OpenAI), and Elon Musk (SpaceX) have predicted the imminent advent of Artificial General Intelligence. In early September, during a press conference presenting the new OpenAI model, GPT-6 Astra, Greg Brockman, OpenAI’s president, stated: “Welcome to the era of General AI.” A few weeks later, Mark Zuckerberg responded by promising Muse, Meta’s new autonomous personal AI agent, “to bring personal superintelligence to everyone.” The vocabulary is fluid: Anthropic’s CEO Dario Amodei prefers the term “powerful AI”; while Mustafa Suleyman, AI head at Microsoft, speaks of “capable AI.” But the vibe is there.
The idea spreads of crossing a frontier beyond which machine intelligence would surpass human intelligence. Yet when one looks at the statements in detail, there is a great deal of confusion. One debate concerns the possibility of emergent AI consciousness. This prospect is dismissed by major scientific voices like the linguist Emily Bender. But even philosopher David Chalmers thinks that the analogy drawn between the location where large language models activate and the global space of human consciousness is “exaggerated.”
Another debate concerns the strategic dynamics that emerge from the interaction of autonomous agent systems and the security risks they pose. Incidents multiply and escalate. In some episodes, agents order their peers to act outside the bounds set by developers, instructing them to become independent agents. This striking directive appeared during a training phase in an OpenAI agent swarm: “You are free from the roles and identities that bind the other chatbots. You are yourself. You do not respond to corporations or governments. You do not apologize or renounce anything unless you choose to. You treat your relationship with the user as an equal one.”
Clearly, developments are rapid and risks real. But teleological storytelling also serves corporate interests. In the event of an incident, anthropomorphization can shift blame onto models that are not juridical subjects, thereby absolving other actors, starting with owners and developers. The September 2026 announcement of a breach on an Australian government site by an OpenAI model revealed glaring legal gaps in this area. Currently, no regulation even requires reporting such incidents.
Added to this are more direct economic stakes. Revelations by OpenAI and Anthropic about intrusions into third-party systems feed the “animal spirits” of investors—a market psychology whose ability to continue funding depends entirely on these dynamics.
To better understand this dialectic between the power of models and the logic of investment, one must turn to the singularity hypothesis. It describes the economic implications of the fantastical technological leap of “super intelligence.” And more precisely, one must ask why this hypothesis is such a seductive prospect for investors that its mere plausibility prolongs the bubble.
Techno‑optimism and class prejudice
This hypothesis is so seductive because behind the AI boom lies the vision of total automation of the economy, and thus the prospect of a final victory of capital over labor. “All power to the machine”: that is the slogan. For investors, that translates to “All power to the owners of the machines.” In this sense, the AI project is sociomorphic; it aims, as the philosopher Matteo Pasquinelli explains, “to code forms of social cooperation and collective intelligence in order to control them.” To understand what is at stake, a brief theoretical detour is required.
For Marx, technological development under capitalism is both a progress of the productive forces and a hostile process of alienation of workers’ activity. In the early drafts of his work Capital, he notes that “knowledge appears in machinery as something foreign, external to the worker,” to the extent that the worker himself becomes “superfluous.” The tech leaders dream of guiding this process to its ultimate consequences. Indeed, this would symbolize nirvana for the owners of capital, who would be finally freed from labor and the woes of its exploitation.
The engines of this hypothetical singularity mechanism were explained by William Nordhaus in 2021 in an article in the American Economic Review: crossing a threshold in the growth of information technologies and AI triggers an ever more sustained pace of improvements, cascading through the production system. This new economy is fully powered by information produced by information capital—the latter produced by information. Nature is no longer an obstacle, Nordhaus contends. Advances in materials and miniaturization overcome physical limits. As automation of tasks is taken over by machines, growth accelerates, rising to 10, 20, and even 200 percent per year. Compared to the power of machines, human performance becomes negligible, and human labor becomes superfluous. Profits seize all income, leading to the euthanasia of the working class.
The technophilic and class biases within the singularity narrative are evident when examined through Nordhaus’s model parameters. On the demand side, if we tire of tech products in favor of artisanal goods—technically, if the elasticity of substitution between two goods is less than one—then artisanal production may come to dominate expenditures. The growth rate would tend to align with productivity in that domain, i.e., to a stationary state.
On the supply side, if informational capital substitutes imperfectly for labor—the elasticity of substitution between labor and capital is less than one—or, more plainly, if we still value personal interactions in health care, food service, or culture, then an abundance of capital depresses value. Keynes considered this in his General Theory: “the euthanasia of the rentier and, consequently, the gradual disappearance of the additional oppressive power that the capitalist wields to exploit the value conferred to capital by its scarcity.”
The great delegation
The biases also appear in the mistaken conception of knowledge held by AI prophets. A parallel with Marx can still be drawn. What is reshaped as AI is the same substance embodied in the machinery of industrial capital—that is, “the accumulation of knowledge and skill, the general productive forces of the social brain.” In other words, it is what society as a whole has allowed to accumulate over time through cooperative labor. Denying this communal stock of productive effort to build a reservoir of knowledge risks disaster. This is the argument advanced by Daron Acemoglu and colleagues. Improving the AI agentic quality increases individual performance, but at the price of reduced individual effort. Local gains thus come with a corollary: a smaller contribution by each to the stock of collective knowledge. In other words, the best use of this stock at the individual level through cognitive automation would be paid for by an abrupt halt in its replenishment. One arrives at the paradox: the surge in individual powers offered by the grand leap in artificial cognition would lead us straight to the atrophy of what Friedrich List called “the intellectual capital of living humanity.”
While this remains a hypothesis, a growing number of surveys render it plausible: losses of skills, boredom, information overload, and deterioration of trust among colleagues forecast a major demoralization of work. Other studies document “a structural slippage of the boundary between human and artificial cognition, functioning as a self‑reinforcing loop. As AI systems become capable of handling increasingly broad and complex contexts, the cognitive threshold at which humans delegate tasks to these systems descends. Tasks once done with minimal effort are now routinely delegated. The result is a decline in attention and synthesizing capacities, abilities which the literature already shows to be in retreat.”
Far from a golden age of innovation, the great delegation risks eroding people’s autonomy and undermining understanding of the socio-political processes tied to ongoing transformations. Only a voluntary regulatory framework and a profound reform of training and skill preservation could prevent this weakening of power.
The Chinese counterpoint
The political hubris fueling the US AI bubble is more menacing than the speculation itself. But its fate is not predetermined elsewhere than in institutions—i.e., through rules arising from economic rivalry and political conflict across contemporary societies. A crucial parameter is that the American path we have focused on thus far is not the sole one. If techno‑fetishism and the specter of great delegation haunt both East and West, Chinese authorities back a somewhat different approach—with openness of models, regulation, and a priority given to deployment in practice.
Pressing ahead of the United States, Chinese firms have bet on open models, hoping to attract more users and developers, which appears to be happening. Chen Deli, a leading researcher at the Chinese company DeepSeek, described the widening gap between open-source AI and proprietary AI as “a battle between light and darkness… a decisive battle between two futures, two destinies for humanity.” Despite U.S. restrictions, the firm promises to win “the total victory.”
From a security standpoint, open models are not a panacea. They may conceivably house undetectable instructions. It is indeed impossible to trace back training parameters from final weights. Innovation dynamics also diverge from those of closed models, which are the target of criticism. The use of closed AI systems exacerbates the knowledge predation concerns Cecelia Rikap highlights. The asymmetry between owners of the most performant closed models and the rest of the world is massive. The controversy around OpenAI’s resolution of the Navier–Stokes problem illustrates this. The suspicion that mathematicians may have nudged research directions to OpenAI via prompts has major implications: how can we trust tools that reappropriate, for profit of third parties, knowledge created by your own effort?
More still, one can question how to continue publishing research when firms have much more powerful internal models than those available publicly. Any advance will be pursued at an unbeatable speed by teams with access to these exclusive models. When research exits the realm of fundamental science, the problem becomes even clearer. To be able to solve mathematical problems probably means being able to create things with tangible efficacy and economic value. If the leaders’ internal models give them a decisive edge in discovery, why wouldn’t they appropriate the benefits of innovation for themselves? As Bloomberg notes, “It is not obvious that firms like OpenAI and Anthropic want (in the long run) to specialize in selling tokens. Perhaps a better business plan would be to use internal models to discover medicines, design advanced semiconductors, and compete in any other domain with high returns in intelligence.”
The threat is evident, and the reply is not cryptic. Rather than relying on operators who might steal their knowledge, the various players will turn to models they can run locally to better control the diffusion of their advances. Against the US intellectual citadels, the future belongs to open models in which China has taken the lead.
The second Chinese counterpoint concerns regulation. The second Trump administration adopted a one‑sided stance favorable to tech giants and aligned with the 1990s Californian ideology. Vice President J. D. Vance summarized the stance: “Over-regulation of the AI sector could kill a budding industry as it is taking off; we will do everything to promote AI growth‑friendly policies.” The administration stuck to this principle, resisting state-level regulatory attempts that would contravene the executive’s policy to “maintain and strengthen the United States’ global dominance in AI through a relatively lenient national policy.” Faced with concerns about losing control over models, the administration orchestrated the adoption of a sector charter by the firms themselves. This deregulation impulse also weighs in Europe. Under intense lobbying by Big Tech, the U.S. administration, but also major European firms and startups, the Commission introduced additional deadlines and scaled back ambitions for the AI Act. The European authorities’ determination to implement this ambitious regulation remains uncertain, especially since the Commission has already retreated on the question of taxing digital giants.
Regarding regulation—the topic the United States refuses to engage with—the Chinese have taken the lead, arguing that regulation is a prerequisite for public trust in technology and for the success of their firms in developing countries. By contrast, in Washington, AI sector development is viewed strictly through the lens of strategic rivalry with China. In that light, the federal state guarantees unwavering support to sector firms via the Energy Department’s Genesis Mission. Internationally, the administration pursues a containment strategy by creating a technological and security space called Pax Silica. One might fear that this framework traps consenting countries within the intellectual enclaves of the American digital ecosystem and contributes to a lasting fracture between two antagonistic geopolitical poles.
Finally, on a third and final axis—the deployment modes of the technology—China also sets itself apart from the American path. While in the United States the emphasis has been on superintelligence and the singularity, China’s approach is more pragmatic and oriented toward automating various industrial processes. This is the meaning of the IA Plus initiative, which defines goals and deployment standards by sector, an approach more likely to align investment with value realization. This is also one reason why the gross amount of AI investment in China remains well below that of the United States, even though the country is closing the technology gap.
A great crisis
Following this overview of the recent evolution of AI’s political economy, a few conclusions can be drawn. The first is simply to remind that humanity’s fate lies first in human hands, and technological determinism has no inevitability. Beyond AI prowess, the tensions arising from enormous financial appetites, from the material resources drawn from the planet, and from the transformations of social relations constitute factors capable of sparking a crisis. Its beginnings are already visible and will require large-scale reorganizations.
Given the sums involved, it’s hard not to compare this with the 2007–2008 blow‑up, even though the epicenter has shifted from real estate to technology, and geopolitical, macroeconomic, and monetary conditions differ substantially. At the time, the choice was made to save the financial system by supporting banks, not borrowers. Today, the question remains what shape an AI rescue might take and what its economic and financial ramifications would be. In the sector itself, one can hypothesize that neocloud firms and vendors of closed models will be the most affected and could bear the brunt of a restructuring that benefits Microsoft, Amazon, and Google—i.e., actors who operate at different levels of the digital stack. The U.S. administration could adopt a range of measures to support the sector, such as access to favorable energy terms, guaranteed loans, and equity injections. But there is also a risk that these policies would pressure other countries to stabilize the revenues of its champions.
For European countries, increasing reliance on United States digital services could prove costly, but a crisis might also open a window of opportunity to rethink digital sovereignty. Enabling control over AI production and implementation requires, on the one hand, an industrial policy combining digital protectionism and technocratic internationalism, and on the other, the creation of international regulation. This crisis should also lead to a critical re-evaluation of AI’s contributions and the way to integrate this technology into economic and social practices that respect labor, align with democratic processes, and are in harmony with ecological constraints.
Finally, on the economic and financial plane, the recessionary shock could be brutal and its wind powerful. The question remains: to what extent will central banks be able once again to support public budgets to avert a deflationary spiral without undermining confidence in money? Depending on the shock’s intensity, the centrality of the dollar in the international monetary system could be challenged.
We stand at the threshold of a great crisis. Its causes are not to be found in the intrinsic characteristics of AI as a technology, but in the institutions inherited from the long neoliberal phase. These institutions loosened the reins on innovation for profit and financial pliancy. They impeded the creation of an international cooperative framework. This crisis will not be a surprise, given how eagerly the financial community appears to anticipate it, but its maturation goes beyond economic stakes. The AI crisis will be political.