We Are Not Ready for What Comes Next

• This is not an area we typically write on, but given the scale of capital being deployed into AI and the way it is now being framed in markets, it is difficult to ignore.
• The technology is real, and so are the investment flows behind it. What we think is being missed is not the importance of AI, but the shape of how returns will be distributed and the path required to justify the current level of investment.
• Markets are largely pricing a smooth transition: that infrastructure spend is rational, that enterprise adoption converts to monetisation within a reasonable timeframe, and that labour and regulatory effects remain secondary. Each of these assumptions is open to question.

A large investment cycle with a delayed revenue base

• The scale of spending is unprecedented. Hyperscalers are expected to commit over $600 billion in capital expenditure this year, with a significant share directed toward AI infrastructure.
• At the same time, revenue generated directly from AI services remains relatively modest. The gap between infrastructure investment and monetisation is therefore wide, and the current market structure relies on that gap closing over time.
• This introduces a timing problem. The companies leading this investment cycle have historically been valued on capital-light models with strong free cash flow generation. That is no longer the case. Capital intensity has increased materially, and in some cases, free cash flow is already under pressure.
• Enterprise data does not yet suggest a broad monetisation inflection. While productivity gains are being reported, revenue attribution remains limited, and full deployment of agentic AI systems is still at an early stage. Adoption and monetisation are being treated as equivalent in current pricing. They are not.

Value is moving away from the model layer

• The most visible parts of the AI stack are not necessarily the most durable from a return perspective.
• We see increasing evidence that value is shifting away from foundation models, where open-source competition is compressing differentiation, toward areas with stronger structural advantages. These include proprietary data, embedded enterprise applications, and vertically integrated software.
• More importantly, the physical layer of the system appears increasingly relevant. Power infrastructure, grid access, cooling systems, and data centre real estate represent constraints that are difficult to replicate quickly. These assets benefit from long lead times and, in many cases, contracted cash flows that are less sensitive to shifts at the model level.
• This suggests that the long-term return profile of the AI theme may be less dependent on model leadership than currently assumed, and more dependent on control over infrastructure and integration into real economic workflows.

Labour and regulation are likely to become binding constraints

• The distributional effects of AI are beginning to appear in early data. Certain categories of white-collar employment are already showing signs of displacement, even in the absence of a broader economic slowdown.
• At the same time, the regulatory environment is evolving. Legal frameworks are beginning to adapt to the presence of AI systems that are capable of autonomous decision-making. Questions around liability, accountability, and deployment are becoming more relevant, particularly in regulated sectors.
• These dynamics are not yet fully reflected in market pricing. However, they are likely to influence both the pace of adoption and the realised returns across the AI value chain.
• Some of the clearest ways to think about this come from outside traditional financial analysis. Bicentennial Man explores the institutional tension that emerges when a system becomes more than it was designed to be, while MANNA illustrates how efficiency-led systems can reshape labour outcomes over time. While not predictive, both highlight dynamics that are increasingly relevant.

The Cordoba View

• We are less focused on the most visible parts of the AI trade, where expectations are already elevated, and more focused on areas where demand is less conditional.
• In particular, we see stronger risk-adjusted opportunities in power infrastructure, grid technology, and data centre real estate, where supply constraints are physical, lead times are long, and cash flows are more stable.
• We also see a clearer path to value creation in enterprise software businesses with proprietary data and strong positions in regulated sectors such as healthcare, legal services, and financial compliance. Adoption in these areas is likely to be slower, but also more defensible once established.
• By contrast, we are more cautious on areas where capital deployment has moved ahead of demonstrated monetisation, and on segments of the model layer facing increasing competitive pressure from open-source alternatives.
• More broadly, we view AI less as a standard technology cycle and more as a structural shift in how economic value is created and distributed. These shifts rarely proceed in a straight line, and the points of friction are likely to be as important as the areas of growth.

For the full note, see the Research Library.