To create yourself

An accounting mechanism is inflating the entire artificial intelligence sector, while public money risks fuelling a mirage.

Imagine a system in which companies lend money to one another, record it as profit, and then use that apparent profit to attract yet more investors. In essence, this is what is happening among the major players in artificial intelligence.

Microsoft has invested more than $13 billion in OpenAI since 2019. A substantial portion of that investment was issued not in cash but in the form of cloud credits to be spent on Azure, meaning within Microsoft's own ecosystem. OpenAI uses those credits to purchase infrastructure; Microsoft records them as revenue; that revenue boosts its stock price; and a higher stock price attracts new capital. That new capital is then, once again, partly converted into credits. The circle closes.

This dynamic has been described as "vibe revenue": figures that create the appearance of growth without reflecting genuine underlying demand, but rather a network of reciprocal agreements among a small group of dominant players. The pattern repeats elsewhere.

Google has committed up to $40 billion to Anthropic — an initial $10 billion outright, with the remaining $30 billion contingent on performance milestones, partly structured around cloud compute. Amazon has committed up to $25 billion in the same company (with $5 billion paid immediately and up to $20 billion to follow), which in turn has pledged to spend $100 billion on AWS over time. Nvidia, alongside startups such as CoreWeave, helps finance the purchase of chips that are then used to resell computing power back to the very same tech giants that originally sustain the ecosystem.

In this context, demand for AI infrastructure is not emerging organically: it is being manufactured through financial engineering. Credit rating agencies, with Moody's among the most prominent, have already warned that operating costs are far outpacing real revenues. The comparison with 2008 is increasingly invoked: then, subprime mortgages were packaged and revalued until the system collapsed; today, AI valuations are built on revenues generated within the same closed loop that supports them. The system holds only as long as investor confidence remains intact.

The risk is that this artificial demand is now driving vast public investment. In the United States, the Stargate project envisions $500 billion in AI infrastructure. In Europe, the European Commission is preparing a €20 billion plan for up to five gigafactories equipped with around 100,000 chips each. The stated goal is technological sovereignty, but scepticism is growing among insiders. "No one can explain the business case," it is whispered in the European Parliament. Seventy-six proposals have been submitted to host one of the gigafactories across 16 countries, without clarity on their actual market need. Parts of the European innovation sector also argue that the future will not require more data — as suggested by major US platforms — but rather less data, used more efficiently and properly curated.

It is another financial bubble, certainly. But this time it may burst for a more tangible reason: physics. In 2024, US data centres consumed around 183 terawatt-hours of electricity — equivalent to the usage of more than 16 million households — and that figure is projected to more than double to around 426 terawatt-hours by 2030. A single 100-megawatt hyperscale data centre can consume hundreds of millions of gallons of water annually for cooling. In total, North American data centres used over one trillion litres of water in 2025. Land constraints are also becoming critical. In northern Virginia — a global hub for data centres — wait times for grid connections can reach seven years, while AI development cycles last just 12–18 months: a structural mismatch. In Arizona, a major project was rejected due to environmental concerns, while a coalition of 200 NGOs is calling for a moratorium on such infrastructure. Since 2023, projects worth $162 billion have been delayed or cancelled.