The stock market valuation of AI-related firms has increased tenfold over the past decade. As John Lanchester noted recently, all but one of the world’s ten largest companies are connected to the future value of artificial intelligence. All but one of those are American, and together their value is equal to well over half of the US economy. Over the past few years, anticipation of the AI ‘revolution’ has driven a surge in investment in these US tech companies. Promises of a radical breakthrough in post-human intelligence and miraculous productivity gains have captured the animal spirits of investors to the point where, as the FT’s Ruchir Sharma put it, ‘America is now one big bet on AI’. Fixed investment in the sector is so enormous that it was the primary driver of US growth in 2025. The training and operation of AI models requires a huge physical build-up of data centres, computing equipment, cooling systems, network hardware, grid connections and power provision. Tech firms are expected to spend a staggering $5 trillion on this costly infrastructure – which is still mostly concentrated in the US – to meet expected demand between now and 2030.
The problem is that the numbers do not add up. To meet its colossal financial needs the sector has shifted from a model dominated by cash-flow and equity financing to debt financing. In principle, this turn to debt could simply reflect growing profit opportunities and the anticipation of forthcoming prosperity. Increasingly exotic financial deals suggest otherwise. A large part of the hype is fuelled by financial loops in which suppliers invest in their clients and vice versa. OpenAI is a case in point. Its leading chip provider, Nvidia – the most valuable company in the world – is planning to invest $100 billion in OpenAI, effectively funding demand for its own products. OpenAI, meanwhile, is spending almost twice what it earns on Microsoft’s cloud platform Azure, which provides the computing power to run its services, thereby enriching its main backer while accumulating debt.
A lot of creative financing is underway. Take Meta’s plans to build a massive data centre in Louisiana. The $30 billion facility will be owned by Beignet Investor LLC, a joint venture between Meta and a private capital firm called Blue Owl. Neither Blue Owl clients nor Meta will provide the bulk of the financing, however, which comes from a vast pool of bondholders. Meta is primarily committing to a long-term lease to use the facility. As the FT’s Alphaville notes, ‘the cute structuring means that Beignet benefits from Meta’s creditworthiness, but Meta’s creditworthiness is magically not impacted by the financial liability that its long-term lease guarantee constitutes’.
Still, beneath the ingenious financial engineering, the bottom line is that Meta is willing to pay about 1% of its balance sheet to finance the construction of the data centre. And the reason is that, contrary to the claims parroted to bond investors, it seeks protection in the event that the promised future of superintelligence and superabundance fails to materialize. Meta’s data centre deal is symptomatic of the market conjuncture, which one financial analyst described as ‘the convergence of massive need for capital, issuers less willing to hold the residual risk . . . and dry powder’, i.e., available cash. In these circumstances, the job of investment bankers is to convince lenders to assume risks they do not really understand. ‘We’ve seen this story a million times’, the analyst warns, not least in the run-up to the 2008 financial crisis.
Looking narrowly at the strong balance sheets of the leading hyperscalers – Amazon, Meta, Microsoft, Alphabet – the AI boom may appear sustainable. But as cracks appear in weaker players like Oracle and in some corners of the AI development business, anxiety is building that there may not be enough profit to sustain the trend across the ecosystem as a whole. The AI rush comes after years of a booming US stock market and decades-long supercycle of fictitious capital, which entails its own fragilities. Hence the growing concern detectable beneath the bureaucratic language of the Bank of International Settlements: ‘If a decline in AI investment were to come with a significant stock market correction, negative spillovers could be larger than previous booms suggest. Investors have favoured US equities to gain exposure to AI firms and hidden leverage may lead to credit market spillovers.’
The limited evidence from field studies suggests that meaningful productivity gains take place in tasks such as writing, coding and assisting customers in call centres. There is an initial lag, as firms bear the cost of learning how to use the technology, but over time, adopters reap benefits. Since the technology is expected to become widely used and to drive continuous innovation and improvement, including in research and development processes, expectations of the economic benefits are high. If artificial intelligence increases productivity as promised, users will be willing to pay significantly more to access it. According to JP Morgan, given the size of the expected capital expenditure, AI providers ‘would require ~$650 billion of annual revenue in perpetuity’ to earn a 10% return – ‘an astonishingly large number’. That is equivalent to about $35 per month for each of the 1.5 billion active iPhone users, or 0.55% of global GDP. For the moment, prices are kept artificially low as AI firms hide the true economic costs to lock in customers. If the efficiency gains materialize, there will be no problem; flourishing businesses will have plenty of resources to pay the bill. Even if they are muted, AI investors could still emerge with bulging pockets. In a couple of years, when AI has infiltrated work processes to the point that exit costs are prohibitive, the customer base will be unable to escape and coerced into paying up. The world will be hooked on AI, and the tech firms will collect handsome profits.
No one should doubt that this is Big Tech’s strategy, and that even a cascade of failures in the AI business will not make them deviate from it. The history of capitalism is full of phases of crisis followed by dramatic moments of consolidation, and the leading tech companies could even benefit from industry upheaval. Moreover, given the tremendous political influence of Silicon Valley billionaires on the US government, one can expect them to fight tooth and nail to rally political support to achieve their goals. If required, they can always augment the promethean argument with a geopolitical one, presenting winning the AI race against China as an existential challenge for the country and buffing up juicy military contracts.
Still, strong headwinds are mounting. AI adoption went viral following ChatGPT’s release on 30 November 2022 and the value of companies has skyrocketed. But uptake in businesses has not been as high as anticipated. Notwithstanding the hype, the use of AI at work is not soaring and may even be slowing, and concerns only a small fraction of the workforce. Recent evidence indicates that there is no immediate productivity boost from using AI. In short, while some automation is underway, there is no evidence of an imminent AI disruption capable of generating the huge economic gains predicted.
As is well known to radical critics and forcefully argued by Daron Acemoglu and Simon Johnson, there is no such a thing as efficiency-driven capitalist development; increased technical efficiency is a macroeconomic outcome that depends on the institutional setting. Powerful technologies can prove unprofitable and fail to be deployed if the structure of the market prevents investors from reaping the rewards; and they can immiserate labour if they lead to massive layoffs. With AI, the most immediate danger seems to be an epidemic of workforce demoralization. Research suggests that intensive AI use is demotivating and deskilling, fuelling boredom and mediocrity. We could even see a reverse ‘productivity J-curve’: short-term productivity gains rapidly overwhelmed by a deterioration in labour quality.
Another problem is the waste that may result from the quasi-religious bet on AI by Big Tech, enabled by private leadership in the industry and mania-prone markets. The contrast between American and Chinese approaches to AI is instructive. Capitalist economies are beset by a deep coordination problem, as Michael Roberts has stressed: ‘in China there is a plan to meet key targets in technology that will boost the whole economy’ but ‘in the major capitalist economies, all the AI eggs are in a basket owned by the privately owned AI hyperscalers and the Magnificent Seven giant tech media companies – and for them, profitability is key, not technology outcomes’.
Further down the road, if financial strain on the sector intensifies, it is not clear that the material legacy of the boom will be comparable to that of previous bubbles. Indeed, construction and infrastructure account for only a minority of the expense of setting up data centre capacity; almost three-quarters of the investment consists of IT equipment – mostly advanced chips (Graphics Processing Units). Unlike the fibre-optic cables of the dot-com era or the railways of the nineteenth century, AI chips must be replaced frequently as their performance fades and technology improves. If, due to profitability concerns, investment suddenly seizes up, a shrinkage in AI availability relative to its current abundance is a material possibility. Theoretically, if the scale-back of capital expenditure outweighed cost reductions from improvements in AI processes, the legacy of the AI boom would not last long, and available computing power for ordinary AI queries could decline.
This problem of obsolescence has crucial financial implications. Indeed, data centre loans ‘are almost always non-amortizing loans: the payments do not go towards reducing the amount owed. Instead, they are perpetual financing for what is assumed to be a perpetual asset. The assumption is that at the end of the term of the loan – typically five to seven years – the whole balance will be refinanced’. But if the chips are almost worthless after five years, who will refinance an asset whose key component has fully depreciated?
This is to say nothing of the ecological stress caused by the surging demand for land, energy and water to run data centres, putting the whole AI rush on an unsustainable footage. In that context, the ideological function of Big Tech’s space-conquest narrative is to lend credibility to the fantasy of an all-digital future. As Google’s Project Suncatcher explains, ‘demand for AI compute – and energy – will continue to grow’ and ‘in the right orbit, a solar panel can be up to 8 times more productive than on earth, and produce power nearly continuously, reducing the need for batteries’, therefore ‘in the future, space may be the best place to scale AI compute.’
Down on earth, growing demand for cheap energy and rare earths materializes in old-fashioned imperialism. The new US security doctrine makes clear it wants ‘a hemisphere . . . that supports critical supply chains’. The Trump Administration’s seizure of Venezuelan oil and expansionary claims on Greenland for critical minerals coveted by tech billionaires show how serious it is. If AI continues to disappoint, imperialist adventures could well intensify – the digital pursuit of chimerical efficiency gains replaced by a predatory race to reduce costs in a new epoch of what David Harvey so accurately called ‘accumulation by dispossession’.
Read on: Cédric Durand, ‘Michel Aglietta’, NLR 156.