Dr. Laurin Lilly-Laona Friedrich

For a long time, artificial intelligence was relatively easy to imagine as something essentially digital: models, algorithms and software operating somewhere behind a screen. Even as its capabilities expanded, the language surrounding it remained strangely weightless. We talked about data, compute and intelligence as though they existed somewhere separate from the physical economy.

That separation is becoming harder to sustain. AI needs data centres, semiconductors, electricity, cooling, grids, land and specialised supply chains. Building and operating that infrastructure requires increasingly large amounts of capital. AI may still appear to us through a screen, but its expansion now has physical, environmental and financial consequences far beyond it.

What makes this particularly interesting is that, at almost exactly the same moment, finance is beginning to look towards AI as a way of understanding those kinds of consequences.

Consider something like credit and ESG assessment. A lender trying to judge whether a company has a credible transition pathway may need to understand emissions and energy dependence alongside investment plans, sector dynamics, supply-chain exposure and financial resilience. The difficulty is increasingly not simply whether information exists. It is whether institutions can interpret enough of it, in combination and at sufficient speed, to reach a meaningful judgment.

This is one of the areas in which AI could become extraordinarily useful. More sophisticated analytical tools may allow institutions to identify relationships across large datasets, model scenarios and distinguish patterns that would otherwise be difficult to see. The promise is not simply more information, but potentially a greater ability to make sense of relationships between forms of information that have traditionally been assessed separately.

Those judgments already matter for where money flows. In its July 2026 bank lending survey, the Europäische Zentralbank found that climate considerations had an easing effect on credit standards for green firms and companies making progress in their transition, while high-emitting firms without credible transition plans experienced a tightening effect. The survey does not tell us that AI is making those decisions. It shows something more basic, and important for what comes next: the way institutions interpret environmental and transition information can already affect the conditions under which capital becomes available.

If AI becomes part of that interpretive process, the relationship is initially straightforward. Better analysis may lead to better differentiation of risk; that differentiation informs decisions about credit and investment; and those decisions help determine which activities receive capital, and on what terms.

But capital does not stop at the edge of the AI system. Some of it flows back into AI itself.

A striking example appeared this week. The Financial Times reported that Jane Street is in talks to refinance around $11bn of debt with a small group of investors including Pimco, potentially shifting substantial borrowing from public debt markets into private credit. Greater flexibility to make further investments in AI is part of the rationale. The transaction has other motivations, and it would be wrong to describe $11bn simply as “AI financing”. But the scale still tells us something: AI investment has become sufficiently material to feature in decisions about how very large amounts of capital are raised and structured.

Now follow that capital one step further. It finances models and computing capacity, but also the data centres, chips and energy infrastructure on which those models depend. The ECB has already described AI as driving a significant increase in capital expenditure among leading technology firms, particularly across data centres, semiconductors and energy systems. Its latest Economic Bulletin goes a step further in showing AI-related investment becoming visible in the wider economy: increased AI investment was one factor supporting resilience in digital services and manufacturing despite higher energy prices and uncertainty.

This is the point at which the distinction between AI as an analytical tool and AI as an economic activity begins to blur. The infrastructure needed to provide analytical capability has its own energy demand, environmental footprint, financing structures and operational dependencies. Greater reliance on a relatively small number of cloud providers, semiconductor producers or sources of compute can create concentrations of its own. At sufficient scale, AI does not simply help institutions interpret economic conditions; its development contributes to creating those conditions.

And those consequences do not disappear. They return to the financial system.

Energy demand becomes a question of costs, infrastructure and transition. New financing structures create exposures for lenders and investors. Concentrations in cloud and compute become questions of operational resilience. The environmental footprint of data centres becomes part of the wider sustainability picture. What began as infrastructure supporting an analytical tool returns to banks and investors as precisely the kind of credit, environmental, operational and financial information they need to understand.

And they may increasingly ask AI to help them understand it.

That is why I keep coming back to the ouroboros: the ancient image of the snake biting its own tail. Not as a symbol of destruction, but of self-reference. A system travels far enough outward that it eventually encounters itself.

The shape is remarkably close to what is emerging here. AI helps finance interpret increasingly complex economic and environmental information. Those interpretations contribute to decisions about where capital flows. Capital finances economic activity, including the continued expansion of AI. AI expansion creates physical, environmental and financial consequences of its own. Those consequences return to the financial system as new information requiring interpretation - and AI may once again be used to interpret them.

The interesting question, then, is not simply whether AI is good or bad for sustainability, or whether banks should use more AI in risk assessment. Both questions are too linear. The harder problem is what happens when the tool used to understand consequences becomes economically significant enough to produce consequences that subsequently return to it for analysis.

That produces some uncomfortable questions. If AI helps a lender assess the credibility of a company's transition pathway, how should the resource demands behind that analytical capability enter the wider assessment? If institutions depend increasingly on concentrated compute and cloud infrastructure in order to understand risk, when does that dependency become part of the risk picture itself? And if the expansion of AI begins to influence energy demand and financing structures, can we continue to treat the technology doing the analysis as conceptually separate from the economy being analysed?

None of this is an argument against using AI in finance. Quite the opposite. The ability to recognise relationships across quantities and types of information that no human team could realistically process may become one of its most valuable contributions. But the more powerful that capability becomes, the less convincing it is to imagine the machinery behind it as weightless.

AI has left the server room. Finance may increasingly use it to understand the consequences travelling through the economy, while simultaneously financing an AI infrastructure that creates new consequences of its own.

Eventually, those consequences come back around.

The snake has found its tail.