The author of this article gets to grips with how the firms running AI businesses can reconcile heavy capex with the returns that investors expect.

The author of this article, James Knoedler (pictured below), who is portfolio manager of the Evenlode Global Equity fund, and part of Evenlode Investment, discusses the trouble AI will encounter if these platforms keep deferring their bills. He examines the economics of AI and how the model of continual price cuts makes it difficult to work out if there is any underlying pricing power or not. He also touches on other aspects of the financial side of AI and how it affects business models. (The article was written in late 2025.)

The editors are pleased to share these views. The usual editorial disclaimers apply to opinions of guest authors. To comment, email tom.burroughes@wealthbriefing.com and amanda.cheesley@clearviewpublishing.com

James Knoedler

All bills become due eventually, and AI bills are mounting in two senses – first, the capital investment into AI capacity just keeps growing, and second, as a result the required payoff keeps getting more demanding. While capital markets can be notoriously generous at times, they also cannot function if they defer bills indefinitely.

The economics of AI
More than two years into the “AI era,” much is still not understood, and the contours of the landscape are getting sharper. The technology is very different from previous waves of digital disruption in that it has high variable costs, very high capital intensity, and therefore does not scale in the same attractive way as past superstar digital businesses as Microsoft, Alphabet, and Meta did.

For context, when Facebook had a similar number of monthly active users as OpenAI does now, it had a 25 per cent free cashflow (FCF) margin, whereas OpenAI expects to have a -54 per cent FCF margin in 2025, based on financials leaked to the press, despite massive subsidies from its partners and a reliance on non-cash expensing (xAI, meanwhile, expects a -2,640 per cent FCF margin). The big names at the heart of the AI category are burning cash at an historic rate.

This business model looks more like the Uber or WeWork generation of VC-funded “unicorns.” Investors in AI companies are betting that the economics will look like Uber in the long run, but the model of continual price cuts makes it very hard to work out if there is any underlying pricing power or not.

Providers of compute power to the AI names are also seeing gross margins erode, and capital intensity soar as the recent Oracle results revealed; the big tech earnings growth we cite above is significantly ahead of FCF growth.

The only big winner so far has clearly been Nvidia whose graphic processing units (GPUs) are the default backbone on which the industry has organised itself. It is significant that Nvidia’s marginal customer has shifted away from the cashed-up hyperscalers, to OpenAI’s new compute partners who are highly levered and much more exposed to capital markets cycles, such as Oracle, CoreWeave, and Soft Bank.

AI inflation
The other surprise relative to the expectations of 2023 has been the rapid penetration of consumer AI applications but the relatively disappointing enterprise uptake. As we learn more about the underlying technology, this is not that surprising. The technology tends to be approximate and unreliable in a way that consumers can accept but so far is prohibitive for businesses which carry strict liability and regulatory oversight.

While there are hopes that new developments within AI such as retrieval augmented generation (RAG), agentic models and reasoning models can address these issues, the very high cost of these innovations brings us back to the exorbitant capital demands of the industry.

Recent deals including Salesforce’s purchase of Informatica and Meta’s purchase of half of Scale AI, which we think are intended respectively to address the quality of retrieval data and post-training fine-tuning of models highlight the steady inflation of the AI bill while big revenue and profit payoffs seem frustratingly always just around the corner.

This takes us back to the everyday reality of our portfolio and what we are doing. We want to find companies where the competitive advantage is ideally improving at the margin, reflected in better pricing power and gross margins drifting gently upwards to enable more reinvestment in the competitive advantage.

Importance of data moats
As one well-respected information services company said to us, “AI is a UX (user experience) not a product in itself.” The natural language interface enabled by large language models can feel like an actual conversation, which makes it easier for non-technical users to surface content buried in large data sets. This is naturally useful to companies which control big proprietary bodies of information which are not available to the wider public and can be used to help businesses become more efficient. We are well invested in this sector through classic data companies such as RELX, Verisk, Wolters Kluwer, S&P Global, LSEG, and Experian.

There are many ways AI can improve the productivity of these companies’ clients. For instance, Verisk is already selling a tool which allows faster review and processing of insurance claims documents. RELX similarly is selling tools which speed up the bread-and-butter tasks of lawyers in big white-shoe law firms, allowing quicker “Shepardisation” of legal cases to see if they are valid precedents or not. Both have the common goal of improving the daily productivity of professionals who are already highly skilled.

We are also interested in companies which have large logs of transactional data from dealings which happen on their platforms, including Mastercard, CME, ICE, and Marsh McClennan, which again can use these tools to aggregate insights from the data and make them available to clients.

The common thread binding these companies is that they are not building the infrastructure for AI because they do not have to. The proprietary data they control cannot be reproduced with AI so they can afford to be price makers on making AI tiers of content available to clients; in other words, they price for the value it creates today. While this may be boringly conservative, we are not eager to take existential bets with client capital.

Weighing real-world impact
While there is a lot of lofty talk of superintelligence and the singularity looming around the corner, when and if we get the “wild abundance of intelligence and energy” recently promised in the CEO of OpenAI’s blog, everyone will be a winner.

We concentrate on the scenario where this does not happen and continue to focus on balance sheets, cashflow, and valuations. For now, we remain in a world where money and intelligence are relatively scarce and valuable. To be clear, we are not sceptics on the value of AI in its many guises. However, if it is to be a revolutionary technology, we assume that it will benefit a broad range of enterprises and consumers.

Its real impact so far has been muted outside a small pool of picks-and-shovels names. This is unsustainable – either it will produce substantial and widely shared benefits for the wider economy, or it will be deprioritised and moved back to more niche use cases such as coding copilots, language translation and natural language search tools.